BaseState Player Value Model
Context-neutral player and team value in runs, 2015–2026, and the 2027 forecast: every element, how it was derived, and how it was tested.
1.Summary
The BaseState Player Value Model values every major-league player in runs above or below a league-average player, per plate appearance for hitters and per batter faced for pitchers, and adds players up to teams. It is built from the complete MLB Stats API play-by-play for every regular-season game from 2014 through 2026 (30,052 games; 2014 is used only for the 2015 park factors), rebuilt one plate appearance at a time, plus FanGraphs fielding, positional, WAR and projection inputs and, for the forecast's pitcher prior and the pitch-quality views, the Stats API's pitch tracking. The dashboard shows the model for 2015–2026, with FanGraphs prospect grades (its 2015 and 2016 article lists, then its prospect board from 2017), and a forecast for 2027; each player card also lists his earlier MLB seasons with FanGraphs' stat lines and WAR.
The value is context-neutral: an event is worth what that event is worth on average across all base-out situations, so a hitter is not credited for when his hits happened, a pitcher is not charged for his fielders, and a runner is credited for what he did on the bases relative to an average runner on the same kind of ball in the same situation. Every component adds to zero across the league (the eight component totals in each season's ledger are zero to within 6e-04 runs; the exception is 2020 fielding at −1.00 runs: the 4 fielders with innings but no plate appearance carry +1.00 runs of centered fielding that no hitter record holds), and the players on a team add up to a team total that tracks actual run differential: r = 0.88–0.98 across the 30 clubs in every season 2015–2026.
The first tile compares the model's batting runs with FanGraphs' Batting runs for the 267 hitters with 300+ PA in 2026 (95% interval 0.982–0.989); the second is across the 30 clubs (0.84–0.96). The third is not a miss of the 2027 forecast, which cannot be scored yet, but of the same method predicting 2026 (below).
The 2027 forecast combines three seasons of each player's run values, pulls every component toward average by an amount fitted to how much it repeats (pitchers' strikeouts, walks and contact toward what their pitch tracking predicts instead), ages batting, baserunning and fielding (a straight line each), and pitchers' strikeouts and walks (a quadratic curve) and pitchers' contact quality (a straight line), with shapes fitted on 2015–2026 and chosen by back-test (the other components are not aged), takes playing time and teams from FanGraphs Depth Charts, and turns team totals into projected wins. In a back-test that predicts 2026 from 2024–2025 with settings fitted on 2025, the typical miss (the root mean squared error of total runs per 600, weighted by playing time in 2026, over the 248 hitters with 300+ PA in 2026 and an MLB season in 2025) fell from 27.6 runs per 600 PA with last year's rate unregressed to 16.9, and for the 250 pitchers with 250+ batters faced from 17.0 to 13.0 per 600 batters faced. A paired bootstrap of the players (paper_checks_forecast.json) puts those gains at +10.68 [+5.67, +18.01] and +3.97 [+2.73, +5.23] runs per 600, better in every resample. The test is not fully out of sample: each component's aging shape was picked by how well it predicted 2026 itself (forecast.py, pick_aging). Which shape is picked hardly matters: against a straight line fitted on 2015–2025 for every component, the per-component choice lowers the miss by 0.000 and 0.004 runs per 600. Aging at all matters more: without it the misses are 17.32 and 13.21, a gain of +0.41 [+0.05, +0.76] for hitters (better in 98.8% of resamples) and +0.16 [−0.09, +0.41] for pitchers (89.9%). With the shapes picked on 2025 instead, out of sample, the gain over no aging is +0.28 [+0.09, +0.46] for hitters and +0.12 [−0.04, +0.29] for pitchers. Pitch quality as the pitchers' prior gains +0.26 [+0.08, +0.43] (Section 12).
Minor leaguers enter the forecast through a translation of their Triple-A and Double-A seasons (Section 14): the same model built from minor-league play-by-play, carried over to MLB piece by piece with weights fitted on the players who made the jump. With the translation weights fitted on earlier seasons (2022–2024 for the 2025 test, 2022–2025 for 2026) and tested on 2025–26, it cut the typical miss for the 99 hitters with no MLB playing time in the three seasons before the target season (the forecast's history window) from 21.6 (the forecast's earlier treatment: the newcomer level measured on the seasons before each test) to 20.3 runs per 600 PA. A paired bootstrap of those 99 players (4,000 resamples) puts the gain at +1.25 [−0.53, +3.20], larger than zero in 92% of resamples; the 95% interval includes zero, so the gain is suggestive, not established. The MLB side of both tests uses the forecast's final k, the average of the settings fitted on 2025 and on 2026, so a little 2026 information enters the 2026 test. Both versions still project these players below what they did (mean error −6.1 and −7.6 runs per 600).
FanGraphs WAR is shown beside the model's runs throughout the dashboard for reference; Section 11 explains how the two scales relate (a full-time average player is about 2 WAR) and how team WAR is built. Section 16 sets out what the model borrows from other methods and what it does differently.
Known defects. Where the code does something other than what the method intends, the section that describes the method says so and measures it; nothing in the model was changed to make these numbers, except three defects fixed after they were measured, whose numbers before and after are both given. Until 2026-11-16 a pinch runner's first move on a plate-appearance-ending play was recorded as the replaced runner staying put (5,075 times in 2015–2026, 1,498 runs of absolute error; Section 3 has the defect, the fix and its effects, Section 6 its effect on the runner weights), the automatic extra-inning runner's stays before his first move were never written (3,793 rows in 2020–2026, −104.3 runs; Section 3), and fielder's choices that retire nobody credited their play value to the batter's situational outs (4,315 plays in 2015–2026, +4800 runs; Section 6). The largest of those that remain: park factors are keyed by the home club, not the ballpark (Section 6); the positional adjustment does not cancel in team sums (Sections 2 and 7); bunt hits leak their result into the expected value of contact (Section 8). The checks are in paper_checks_core.py, which writes output/paper_checks_core.json.
The dashboard's Method view is the short version. This paper documents every element, how it was derived, the tests that chose between alternatives, the earlier versions that were wrong, and the known limitations.
2.Design principles
One currency on both sides of every plate appearance
Each plate appearance is valued once from the offense's side (batter and runners) and once from the defense's side (pitcher, catcher and fielders). Both sides use the same run values, so the league's batting runs add to zero, the league's pitching runs add to zero, and a team's offensive runs plus its runs prevented (both positive when better than average) is a run differential relative to the league. Values are always relative to the league average for the season, never to a replacement level, so a full-time average player is worth zero.
Context-neutral
- Timing is excluded. A single is worth the average value of a single over all 24 base-out states, whether it came with the bases empty or loaded. Timing (sequencing, clutch) is measured separately and shown on team views as part of the % difference.
- Park is removed from batting with a two-season park factor (keyed by home club; Section 6).
- Opportunities are normalized. Situational outs and baserunning are rated per opportunity and applied at the league-average rate of opportunities, so a hitter is not rewarded for batting behind teammates who reach base often.
- Credit goes to the player responsible. Runner advancement belongs to the runner; double plays on grounders belong to the batter; what fielders do with a batted ball belongs to the fielders, not the pitcher; catcher framing belongs to the catcher, not the pitcher. The rule has documented exceptions where the play-by-play cannot tell who is responsible: catcher's interference is charged to the pitcher (83 in 2026, +0.337 runs each; Section 8); a fielder's choice that retires nobody goes to no one, into the % difference, because the fielders who chose cannot be identified from the feed (408 in 2026, +475.3 runs; until 2026-11-16 it went to the batter's situational outs, Section 6); a batter who reaches on a dropped third strike gets baserunning credit (52 in 2026, +32.1 runs); a steal or caught stealing on the plate appearance's final pitch is valued as the runner's advancement on that plate appearance, not in the running game, so it is not charged to the pitcher or the catcher (350 runners in 2026, +21.2 runs; Section 6).
Additive by construction
Components add to a player's runs; players add to a team. Everything not credited to a player is collected in one difference between actual run differential and the players runs total, shown as the % difference, so the books always balance against actual run differential. It holds:
- timing (the base-out state each event came in);
- home runs beyond what the contact predicted, and batted-ball conversion that the fielding metric does not assign to a player, including FanGraphs' team shift and positioning runs, which no player row carries (+579 runs league-wide in 2026 in FanGraphs' team table; Section 7);
- the play residual on every plate appearance that is not a ball-in-play out (hits, strikeouts, walks, hit batters, reached on error, interference) beyond what the runners are credited with: outfield arms, hit location, errors on the play;
- runner events credited to nobody: defensive indifference and the feed's other advances, outs and errors between pitches (in 2026: defensive indiff +22.9, other out −17.1, forced balk +4.4, error +1.6);
- the pitching side of the park, which batting's park factor does not touch;
- the half-inning terms of Section 4, on both sides of a club's games: the run expectancy each half-inning opens with (the bases empty, or from 2020 the automatic runner on second in extra innings, whose runs no plate appearance is credited with) and the final state of half-innings that end without a third out (walk-offs). On defense they are almost all of the gap between a club's run-prevention layers and its runs allowed (r = −0.9998 across the clubs in 2026; Section 5);
- innings played: clubs bat and defend different numbers of half-innings (a visiting club does not pitch the bottom of the ninth when the home club leads), which enters through the same terms;
- players with no plate appearance: the model keeps no hitter record for them, so their fielding is dropped (Section 7) and a pinch runner who never batted keeps no baserunning.
Per unit and per season
Hitters are rated per plate appearance, pitchers per batter faced, and both are reported as season totals and per 600 units, so part-time players compare with regulars and season totals still add up.
Splits
Every hitter is shown against left- and right-handed pitching, every pitcher against left- and right-handed batters, every team lineup against each pitcher hand, and every team staff against each batter hand. Parts of a player's value that do not depend on the matchup are spread by playing-time share (baserunning by the share of the runner's opportunities), so the two splits always add up to the total.
Positional adjustment kept apart
The positional adjustment is shown for comparing a shortstop with a first baseman, but it is left out of team sums. The reason first given here, that every club fields one of each position so the adjustments cancel, is wrong. Summed by each player's share of time with the club, FanGraphs' positional runs come to −20.1 runs per club in 2026 (SD 1.9 across clubs); in the seasons with a universal designated hitter (2020, 2022–2026) the club sums are nearly the same for every club but far from zero. In 2015–2019, 2021 they spread with an SD of 24–26 runs, because pitchers who batted carry FanGraphs' pitcher positional adjustment (+557 to +627 runs a season league-wide), so National League clubs sum far above American League ones. Adding the adjustment would shift every club by a near-constant since 2022 and bias the earlier seasons by league; it is kept out for that reason (Section 7).
3.Data and play-by-play reconstruction
Sources
| Source | What is used | Seasons |
|---|---|---|
MLB Stats API game/{pk}/playByPlay | Every play: event type, base runners and their movements (start, end, out, reason, play index), outs, score, batter and pitcher with handedness, exit velocity, launch angle and trajectory of the batted ball, and the play's events (pitch flags, substitutions). Requested with a 51-name field filter. | 2014–2026 (2014 only for the 2015 park factors) |
| MLB Stats API schedule, standings, teams, people | Final regular-season games, wins and losses, team names, player names, birth dates, bats and throws | 2014–2026 |
| FanGraphs leaders API (players) | Fielding: DRS and its parts, Statcast fielding parts, catcher framing, blocking and throwing. Batting and pitching: positional adjustment; WAR, batting and baserunning runs for comparison; stat lines for the player cards | fielding 2015–2026; batting and pitching 1990–2026 (career rows) |
| FanGraphs leaders API (teams) | Team totals used in the Brewers study and the fielding-metric test | 2025–2026 |
| MLB Stats API game feed (pitch level) | Pitch tracking for pitch quality (Sections 8 and 12) | 2015–2026 |
| MLB Stats API play-by-play, minor leagues | Triple-A, Double-A, High-A and Single-A play-by-play for the translations (Section 14) | AAA 2021–2026; AA 2021–2026; High-A 2021–2026; A 2021–2026 |
| MLB Stats API daily rosters | Active days for rookie status (the Top 100 rookies) | 2015–2026 |
| FanGraphs Depth Charts projections | 2027 playing time (PA, batters faced, games, starts, innings), team and projected WAR | 2027 |
| FanGraphs RosterResource depth-chart pages | Check that the projection feed's teams match current rosters | 2027 |
| FanGraphs prospect lists and board | Prospect grades shown on the dashboard (2015–2016 article lists, the board from 2017) | 2015–2026 |
All requests are plain HTTPS. FanGraphs answers curl but refuses Python's default client (HTTP 403), so the FanGraphs scripts shell out to curl; FanGraphs also rate-limits fast runs (HTTP 429), so the career-row fetcher pauses 20 seconds between files and backs off up to 10 minutes. The play-by-play is requested with a field filter that keeps about one fifth of the full feed (11.8–13.5 MB compressed per full season). fetch_pbp.py takes the schedule with gameType=R and keeps games whose coded state is final (F); a suspended game listed on two dates is kept once (by game id, dated by its official date). It downloads with 16 threads, a 60-second timeout and up to five tries with a back-off of 2, 4, 6 and 8 seconds. The MLB files drop the park id the schedule carries (the minor-league files keep it), which is why MLB park factors are keyed by home club (Section 6).
Plate appearances and runner events
Each play in the feed is either a plate appearance (one of 26 event types: batter_interference, catcher_interf, double, double_play, fan_interference, field_error, field_out, fielders_choice, fielders_choice_out, force_out, grounded_into_double_play, grounded_into_triple_play, hit_by_pitch, home_run, intent_walk, sac_bunt, sac_bunt_double_play, sac_fly, sac_fly_double_play, single, strikeout, strikeout_double_play, strikeout_triple_play, triple, triple_play, walk) or a play that ends without one. In 2026 there were 479 plays without a plate appearance (caught stealing 2b 282, pickoff 1b 73, pickoff caught stealing 2b 56, other out 17, caught stealing 3b 16, caught stealing home 13, pickoff 2b 8, wild pitch 4, pickoff caught stealing home 3, pickoff 3b 2, pickoff caught stealing 3b 2, game advisory 1, stolen base 2b 1, stolen base 3b 1). Within a plate appearance, runner movements are split by the play index on which they happened, relative to the index of the plate appearance's last pitch:
mid-PA = movements with playIndex < index of the last pitch event (isPitch) and a start base final = every other movement of the play (the batter's own movements are always final) non-PA play: every movement is a runner event
- Mid-PA runner events: grouped by play index, applied as one batch each, and labeled with the first runner's event type. They become runner/battery events and are never credited to the batter. In 2026 there were 5,708 such groups: stolen base 2b 2,536, wild pitch 1,336, stolen base 3b 346, caught stealing 2b 344, defensive indiff 300, passed ball 208, balk 132, pickoff 1b 112, pickoff error 1b 88, pickoff caught stealing 2b 87, caught stealing 3b 60, pickoff error 2b 51, forced balk 22, pickoff 2b 18, other out 16, pickoff caught stealing 3b 13, pickoff error 3b 10, stolen base home 8, pickoff 3b 7, error 5, pickoff caught stealing home 5, caught stealing home 4. The 479 plays without a plate appearance add 484 more rows (caught stealing 2b 282, pickoff 1b 73, pickoff caught stealing 2b 56, other out 17, caught stealing 3b 16, caught stealing home 13, pickoff 2b 8, stolen base 2b 5, wild pitch 5, pickoff caught stealing home 3, pickoff caught stealing 3b 2, stolen base 3b 2, pickoff 3b 2), so the runner-event table has 6,192 rows (grouping replicated from
pa_table.pyon the feed). - The plate appearance itself: the final pitch and everything that happened on it, including a steal or caught stealing on that pitch (Section 6).
Tracking the bases
The base state is tracked by runner id through each half-inning. Three details matter:
- Batch moves. One runner can have several movement entries on one play (first to second on the hit, second to third on the throw), and two runners can pass through the same base. Applying entries one at a time let a later runner overwrite an earlier one. Each play's movements are therefore applied as one batch: take each runner's final destination, lift every moving runner off the bases, then set them down.
- Pinch runners. Substitutions are not tracked, so after a pinch runner enters, the tracker still holds the player he replaced. When a play is applied, a moving runner clears whoever was recorded on his starting base, which keeps the base state right but not the runner table, because a plate appearance's runner rows are written from the tracker before its play is applied. Until 2026-11-16, on the play where a pinch runner first moved, if that play ended a plate appearance, the replaced player was written as staying on his base and the pinch runner's own movement was lost. Measured before the fix (the check in
paper_checks_core.py, its pre-fix output frozen inpinch_fix_history.json), it happened 5,075 times in 2015–2026 (376–482 a full season, 463 in 2026); every one of them followed an offensive substitution of that runner earlier in the same half-inning (5,075 of 5,075), which identifies them as pinch runners (the substitution can sit in the same play, before the pitch). Valued with the season's cells and weights (Section 6), the recorded "stays" carried −745.2 runs and the real moves +54.9, a total absolute error of 1,498 runs; 291 were on home runs, on which the model values no runner at all, so they carried no error; 1,381 were runners forced on ground-ball or bunt outs, whose deviation goes to the batter (−162.5 runs net); the other 3,403 were charged to the replaced runner's baserunning instead of credited to the pinch runner. The check looked only for lost moves. The same cause had a second, value-neutral form it could not see: a pinch runner who entered and stayed put on the plate appearance that ended was written under the replaced player's id, with the right deviation but on the wrong player's record.
The fix (2026-11-16,pa_table.py) corrects the tracker's copy of who was on each base before the rows are written, in two passes, and corrects the tracker itself after the play:moved for each runner m moving from base b on the final pitch: if the tracker holds on b someone who does not move on the play, and m is not recorded on another base, m is b's runner stayed for each base b whose recorded runner r does not move: if the feed's postOn id q for b differs from r, does not move, is not recorded on another base and is not the batter, q replaced r (he entered and stayed put) synced after a plate appearance that does not end the half-inning and whose occupancy matches the feed, each occupied base takes the feed's runner idThe automatic runner's placeholder was excluded from all three, which left that defect (next item) as it was until it was fixed the same day. In 2015–2026 the first pass fired 3,945 times and the second 2,484; the sync fired 2,484 times, on the same plate appearances as the second pass (the two counts are equal in every season), consistent with a pinch runner who has not yet moved being the only source of a disagreement in identity between the tracker and the feed. The first pass fires less often than the check found lost moves (3,945 against 5,075) because a pinch runner who stays put on one plate appearance has been put on his base by the sync before he moves on a later one. An instrumented copy ofpa_table.py(inpinch_fix_history.py) marks each id the sync places, drops the mark when that runner moves or the half-inning ends, and counts the rows written for marked runners; on 2026 it gives 359 first-pass moves and 104 moves after a sync, 463 in all, and 244 second-pass stays and 132 later stays, 376 in all (its runner table is identical to the production one); the file comparison below finds 463 and 376. Rerun after the fix, the check finds no lost move in any season. Since the automatic-runner fix the automatic runner is seated by id and goes through the same rules, so the counts now include pinch runners for him: 4,167, 2,719 and 2,719. The 2026 tables, the only season whose pre-fix files were kept, compared file to file: the plate-appearance and runner-event tables are identical, and the runner table (295,457 rows, the same count) differs in 839 rows, 463 now recording the pinch runner's move and 376 the pinch runner staying put. In every season the three tables have the same row counts as before the fix (a corrected row replaces a row; none is added or dropped). The three reconciliation checks below still find no mismatch in any season.
Effects. Among the player components, only hitters' baserunning and situational outs changed: over the 9,835 hitter-seasons of 2015–2026 (the same players before and after), the absolute changes sum to 1,423 runs of baserunning and 204 of situational outs; batting and fielding did not move, and no pitcher component moved. Both components are centered, so each season's net change is zero (largest 7e−14). 125 hitter-seasons moved by more than a run of baserunning; the largest is Victor Martinez in 2018 (−8.2 to −1.7 in 508 PA), whose pinch runners' moves had been charged to him as stays. The runner weights of Section 6 moved too. With the fix the runs per extra base rose in every season (+0.0071 to +0.0098, 3.1%–4.5%), the direction measurement error in the regressor predicts: a stay recorded where a move happened pairs the move's play residual with a deviation that is too low by the bases gained, and such points flatten the fitted slope. The runs per out on the bases moved toward zero in every season (+0.0013 to +0.0067). That is the opposite of what attenuation alone predicts for this coefficient (a lost out recorded as a stay would have pulled it toward zero before the fix, not away from it); with two correlated regressors in one regression, an error in the extra-base deviations also moves the out coefficient, and that spillover is not decomposed here. Team totals barely moved: the players-against-run-differential correlation changed by at most 0.0027 in any season. The error was persistent per player: a hitter's change in baserunning from the fix in one season correlates with his change in the next at r = 0.374 [0.333, 0.412] over the 1,819 player pairs of the year-to-year table in Section 6 (hitters with 300+ PA in both seasons of a pair, the 9 season pairs stacked; 0.31 to 0.50 season pair by season pair; the Fisher interval treats the stacked pairs as independent though a player can appear in several, so it is too narrow, but every season pair is positive on its own), as it would be if the same hitters are replaced on base season after season. That persistence had been counted as skill. With the fix the pooled year-to-year r of hitters' baserunning fell from 0.541 [0.508, 0.573] to 0.509 [0.474, 0.542], and the forecast's regression constant for baserunning, fitted by back-test, rose from 3,500 to 4,000 PA (more pull toward average). The 2026 back-test barely moved: hitters 16.920 to 16.917 runs per 600, pitchers unchanged at 13.047. Through the minor-league tables, rebuilt with the same code (their pre-fix ledgers were not kept, so the change is measured only downstream), the translation's 2025–26 test for hitters moved from 17.267 to 17.236 runs per 600 PA over all 767 hitters and from 20.307 to 20.360 for the 99 with no MLB time; the pitchers’ fitted translation and test did not change. The rebuild also moved the minor-league ledgers' stored ages to Opening Day:opening_day.pychanged every age from July 1 to Opening Day on 2026-11-15, after those ledgers were last built (the translation they fed is stamped 20:50 that day), so the mean ages quoted in Section 14 were ages on July 1 until this rebuild. The translation computes ages from birth dates and the dashboard recomputes them, so neither was affected. That the pitchers' translation came out identical is the check that nothing else changed on the pitchers' side, which the runner rows do not reach; the hitters' side has no such check. In the 2027 forecast, hitters' batting and fielding moved by at most 2e−05; their baserunning moved by 0.19 runs per 600 on average (47 of 761 hitters by more than half a run; largest +1.90 for Giancarlo Stanton and −0.69 for Jihwan Bae) and their situational outs by at most 0.16; pitchers did not move, and the largest change in a team's projected baserunning is −1.11 runs (MIL). - The automatic extra-inning runner (2020 on) is detected from the data, not assumed from the inning number: a half-inning starts with a runner on second if a runner's first movement on its first play starts at second, or someone is on second after that play without having moved there. An inning-number rule fails for the 2020–21 seven-inning doubleheaders, whose extra innings begin in the 8th; it caused 40 base-state mismatches in 2020–21 and the data-based rule removes all of them. Until 2026-11-16 the tracker then held him as a placeholder until he first moved, and the runner table skipped the placeholder: 3,793 runner rows were never written in 2020–2026 (3,715 in the 10th inning or later, 78 earlier, the seven-inning doubleheaders; measured before the fix by the occupied-base check in
paper_checks_core.py, frozen inauto_fix_history.json). Written as the stays they were, they would have carried −104.3 runs of baserunning, so the automatic runners' baserunning was too high and their stints on base were undercounted.
The feed names him. The first play of every one of the 3,790 automatic-runner half-innings in MLB 2020–2026 and the 11,016 in the four minor-league levels 2021–2026 carries exactly onerunner_placedevent, and elsewhere it appears on 3 later plays, all in the minor leagues (Single-A 2022, Double-A 2025, Triple-A 2023), where the fix does not read it (auto_fix_history.py, from the play-by-play alone). Its player is the runner on second on that play (the one who moves from second on it, or is on second after it) in 13,334 of them and a pinch runner substituted on that same play in the other 1,472. The fix seats the placed player on second by id when the half-inning opens; when he is replaced, the pinch-runner rules above take over, and the placeholder remains only for a half-inning without the event (0 in MLB 2020–2026; 3,790 were named). Rerun after the fix, the check finds no occupied base without a runner row in any season. Compared file to file against the pre-fix snapshots, every plate-appearance and runner-event table (MLB 2020–2026, the four minor-league levels 2021–2026) is identical, and the runner tables gained exactly the rows the check had found missing, season by season (3,793 in MLB; 10,181 in the minor leagues, where the check does not run), every one a stay on second, with no row removed or changed.
Effects. The 3,793 MLB stays belong to 1,407 hitter-seasons. Their baserunning fell by 49.8 runs in all (4.2 to 9.0 a season), and everyone else's rose by the same total, because baserunning is centered on the league. That is 48% of the −104.3 runs the rows carry as raw credit (valued with the pre-fix weights).ledger.pyrecenters baserunning per stint on base (each hitter's total less the league's baserunning per stint times his stints), so the charge the new rows add to the league is spread back over every hitter's stints, the automatic runners' own included, and the rows add stints too, to automatic runners who never moved, which had none for that half-inning. The advancement part is also rated per opportunity and scaled to the league's opportunities per stint (Section 6), and the rows add opportunities. No player-season moved by more than 0.67 runs of baserunning (the largest, Ryan McKenna in 2022, +2.83 to +2.16); situational outs moved by at most 0.35 (on ball-in-play outsledger.pygives the batter's situational value whatever part of the play residual the runners' credits do not take, so a stay the table lacked was charged to the batter and is now the runner's), batting and fielding did not move, and no pitcher component moved. 2 of the runners who gained rows had no plate appearance that season, so the model keeps no hitter record for them and their stays fall into the difference. The runner weights of Section 6 moved by at most 0.0003 and the players-against-run-differential correlation by at most 0.0001. In the minor leagues the 10,181 stays lowered the automatic runners' baserunning by 159.3 runs in all and added 1,445 stints on base; the pitchers' records came out identical in every level and season. In the minor leagues that charge sat in the batters' situational outs, and with it gone the translation carries the piece over much more, as it would if the charge had been noise in the predictor: the fitted slope of MLB situational outs on minor-league ones went from 0.35 to 0.64 at Triple-A and from 0.12 to 0.31 at Double-A, with the reliability constant from 300 to 1,000 PA (it had sat at the bottom of its grid, 300, 1,000, 3,000, 10,000); the baserunning piece barely moved (0.787 to 0.790 at Triple-A). The translation's 2025–26 test for hitters moved from 17.236 to 17.239 runs per 600 PA over all 767 hitters and from 20.360 to 20.391 for the 99 with no MLB time; the pitchers' fitted translation did not change. In the 2027 forecast hitters' baserunning moved by 0.02 runs per 600 on average (largest −0.24 for Ali Sánchez), situational outs by at most 0.31; pitchers did not move, the largest change in a team's projected baserunning is −0.21 runs (SF), and the 2026 back-test went from 16.917 to 16.909 for hitters.
One data quirk is handled explicitly: once (2023-10-01, game 716358) the feed records a run as a movement ending at 4B instead of score; it is treated as a run.
Reconciliation
Every game is checked three ways, and the counts are printed by pa_table.py (and saved by the checks script):
- after each plate appearance that does not end the inning, the tracked base occupancy (which bases are occupied, not by whom) must match the feed's
postOnFirst,postOnSecondandpostOnThird; on a mismatch the count goes up and the tracker is reset to the feed's runners; - after each plate appearance the outs must match the feed's count (the tracker then takes the feed's count); after a play without a plate appearance the tracker takes the feed's outs without a check;
- the runs from plate appearances and runner events must add to the final score on the game's last play.
All 30,052 regular-season games from 2014 to 2026 pass all three checks with zero mismatches. Because the base check compares occupancy only, it cannot see a wrong runner id, which is how the pinch-runner defect above passed it for every season until 2026-11-16. Since the fix a differing id with matching occupancy is taken from the feed and counted (the sync above), not treated as a mismatch, so the check is still blind to identity: the sync corrects the tracker, it does not test it.
| Season | Final games | Clubs with other than a full schedule | PA rows | Runner rows | Runner-event rows | Plays without a PA | Feed, MB | Mismatches (outs / bases / score) |
|---|---|---|---|---|---|---|---|---|
| 2014 | 2,430 | – | 183,929 | 293,367 | 6,014 | 424 | 11.8 | 0 / 0 / 0 |
| 2015 | 2,429 | CLE 161, DET 161 | 183,628 | 291,909 | 5,923 | 491 | 12.7 | 0 / 0 / 0 |
| 2016 | 2,428 | ATL 161, CLE 161, DET 161, MIA 161 | 184,580 | 294,503 | 5,972 | 434 | 12.8 | 0 / 0 / 0 |
| 2017 | 2,430 | – | 185,293 | 296,028 | 5,804 | 403 | 12.9 | 0 / 0 / 0 |
| 2018 | 2,431 | CHC 163, COL 163, LAD 163, MIA 161, MIL 163, PIT 161 | 185,139 | 295,142 | 5,814 | 440 | 12.9 | 0 / 0 / 0 |
| 2019 | 2,429 | CWS 161, DET 161 | 186,517 | 296,535 | 5,402 | 363 | 12.9 | 0 / 0 / 0 |
| 2020 | 898 | DET 58, STL 58 | 66,506 | 107,312 | 2,055 | 141 | 4.7 | 0 / 0 / 0 |
| 2021 | 2,429 | ATL 161, COL 161 | 181,818 | 291,472 | 5,312 | 344 | 12.8 | 0 / 0 / 0 |
| 2022 | 2,430 | – | 182,052 | 290,155 | 5,246 | 416 | 12.6 | 0 / 0 / 0 |
| 2023 | 2,430 | – | 184,104 | 295,651 | 6,259 | 402 | 13.3 | 0 / 0 / 0 |
| 2024 | 2,429 | CLE 161, HOU 161 | 182,449 | 290,424 | 6,426 | 454 | 13.2 | 0 / 0 / 0 |
| 2025 | 2,430 | – | 182,926 | 292,303 | 6,316 | 461 | 13.4 | 0 / 0 / 0 |
| 2026 | 2,429 | BAL 161, NYY 161 | 183,849 | 296,001 | 6,192 | 479 | 13.5 | 0 / 0 / 0 |
From a rerun of pa_table.py on each season's feed with its file writers disabled (paper_checks_core.py); the rerun reproduces the row counts of the tables on disk in every season.
Seasons short of 2,430 games had games never made up (a club at 161); 2018 include tiebreaker games (clubs at 163), which MLB counted as regular-season games through 2021.
The rebuild produces three tables per season: one row per plate appearance (181,818–186,517 in a full season), one row per runner per plate appearance with start base, destination and out reason (290,155–296,535), and one row per runner event (5,246–6,426).
The FanGraphs rows come from the leaders API, not the CSV exports: they match the exports to within about a run on every row (the API carries later updates), and switching to the API moved 2026 hitter runs by 0.14 on average (2.0 at most) when it was made.
4.Run expectancy and event values
Run expectancy
For each season, the expected runs from each of the 24 base-out states to the end of the half-inning are the average actual runs scored from that state, using innings 1–8 (later innings are cut short by walk-offs and skipped bottom halves):
RE(s, o) = mean over plate appearances starting in state (s, o), innings 1-8, of runs_to_end
runs_to_end = runs scored from the start of that PA (after its mid-PA runner events) to the end
of the half-inning, runs on the PA itself and on later runner events included
The estimate is the season's own, in sample, and each plate appearance that starts in a state is one observation (a half-inning contributes as many observations as it has plate appearances). Two edge cases sit inside the innings 1–8 sample: rain-shortened games, and the 2020–21 seven-inning doubleheaders, whose 7th inning can end in a walk-off and whose 8th is an extra inning that opens with the automatic runner. The 2026 table:
| Runners | 0 outs | 1 out | 2 outs |
|---|---|---|---|
| empty | 0.503 | 0.269 | 0.105 |
| 1B | 0.887 | 0.537 | 0.238 |
| 2B | 1.181 | 0.663 | 0.316 |
| 1B 2B | 1.539 | 0.917 | 0.455 |
| 3B | 1.440 | 0.908 | 0.369 |
| 1B 3B | 1.767 | 1.181 | 0.526 |
| 2B 3B | 2.053 | 1.357 | 0.508 |
| loaded | 2.511 | 1.642 | 0.744 |
The run value of a plate appearance is its change in run expectancy plus the runs it scored:
RE24(PA) = RE(end state) - RE(start state) + runs scored on the PA (RE = 0 after the third out)
The start state is the state after any mid-PA runner events, so the batter is never charged or credited for a steal or wild pitch during his plate appearance. The league mean RE24 per PA, μ, is slightly negative (−0.0035 in 2026); every value below is expressed relative to μ, so the league-average PA is worth exactly zero. The first explanation given for the sign, that runner events between pitches carry some of the run expectancy, is only part of it. Over a half-inning the plate appearances and the runner events telescope:
sum over PAs of RE24 + sum over runner events of (RE after - RE before + runs) = sum over half-innings of ( runs scored - RE(opening state) + RE(final state) ) RE(final state) = 0 when the half-inning ends on the third out
The identity holds to 6e-16 runs in every half-inning (checked against the feed's own final states). In 2026 the plate appearances sum to −651 runs; the runner events carry +460, so the half-inning terms are −191. From the 9th inning on they are −118: −264 from runs minus opening states (extra innings from 2020 open at second and none out: 570 half-innings) and +147 from the final states of the 203 half-innings that ended with fewer than three outs, walk-offs (101 in the 9th and 102 in the 10th or later). In innings 1–8 they are −73 (1 half-inning ended short, worth +0.50), which the estimator does not force to zero because RE is a mean over plate-appearance observations of each state, not over half-innings. Every season:
| Season | PAs | mu per PA | mu x PAs | Runner events, change in RE | Half-innings 1-8 | Half-innings 9+: runs minus opening state | Half-innings 9+: final states | Half-innings ending short |
|---|---|---|---|---|---|---|---|---|
| 2015 | 183,628 | −0.0036 | −663 | +468 | +23 | −359 | +140 | 207 |
| 2016 | 184,580 | −0.0045 | −836 | +527 | −54 | −377 | +122 | 191 |
| 2017 | 185,293 | −0.0046 | −844 | +605 | −44 | −324 | +129 | 204 |
| 2018 | 185,139 | −0.0043 | −794 | +529 | −93 | −306 | +134 | 214 |
| 2019 | 186,517 | −0.0037 | −694 | +489 | −20 | −327 | +142 | 203 |
| 2020 | 66,506 | −0.0041 | −272 | +173 | −1 | −147 | +49 | 69 |
| 2021 | 181,818 | −0.0043 | −783 | +602 | −36 | −298 | +154 | 211 |
| 2022 | 182,052 | −0.0036 | −656 | +554 | −43 | −213 | +153 | 209 |
| 2023 | 184,104 | −0.0049 | −900 | +565 | −124 | −344 | +133 | 192 |
| 2024 | 182,449 | −0.0041 | −742 | +612 | −41 | −240 | +152 | 208 |
| 2025 | 182,926 | −0.0042 | −771 | +510 | −106 | −329 | +174 | 222 |
| 2026 | 183,849 | −0.0035 | −651 | +460 | −73 | −264 | +147 | 203 |
Runs. mu x PAs = (half-innings 1–8) + (half-innings 9+, both columns) − (runner events). Half-innings ending short: 9th inning or later, fewer than three outs at the end (walk-offs). Half-innings that end short in innings 1–8 (rain-shortened games, and the 7th-inning walk-offs of the 2020–21 seven-inning doubleheaders) stay in the 1–8 column: 2015 2, 2016 4, 2017 2, 2018 4, 2019 4, 2020 12, 2021 18, 2022 4, 2024 1, 2025 3, 2026 1.
Three ways to value an event
- LW0, wOBA-style. Single, double, triple, home run, walk, hit by pitch, and one value for every out (reached-on-error, sacrifice fly, double play, fielder's choice and strikeout all count as an out). Intentional walks, sacrifice bunts and interference are left out, as in FanGraphs' wOBA: they are valued at μ, i.e. zero. The 2026 out value is −0.262 runs. Because the left-out plate appearances are set to zero rather than to their own mean, LW0 does not sum exactly to zero over the league; the detail layer below absorbs the difference.
- LW1, context-neutral detail. The same, but outs split by the recorded batted-ball trajectory (strikeout, ground ball, fly ball, line drive, popup, bunt, and "no trajectory" for the few outs the feed leaves blank), and reached-on-error, intentional walks and interference valued as their own events. Each class is one value, the mean RE24 of its plate appearances over all base-out states. This is the batting value used in the model.
- LW2, in context. The LW1 class valued in the specific base-out state where it happened: the mean RE24 of that class in that state.
The exact mapping, read from gap_decomp.classes:
| Event types in the feed | LW0 class | LW1 class |
|---|---|---|
single | 1B | 1B |
double | 2B | 2B |
triple | 3B | 3B |
walk | BB | BB |
sac_bunt, sac_bunt_double_play | EXCL | BUNT |
hit_by_pitch | HBP | HBP |
home_run | HR | HR |
intent_walk | EXCL | IBB |
batter_interference, catcher_interf, fan_interference | EXCL | INTERF |
strikeout, strikeout_double_play, strikeout_triple_play | OUT | K |
field_error | OUT | ROE |
double_play, field_out, fielders_choice, fielders_choice_out, force_out, grounded_into_double_play, grounded_into_triple_play, sac_fly, sac_fly_double_play, triple_play | OUT | by trajectory: BIP?, BUNT, FB, GB, LD, PU |
OUT = one wOBA out value; EXCL = left out of wOBA (valued at zero). A trajectory starting with bunt makes any out a bunt out; hits and reached-on-error keep their class whatever the trajectory.
Two consequences of the mapping matter later. A fielder's choice in which nobody is retired is an out in both LW0 and LW1 (its class is its trajectory), so its play residual is large and positive; until 2026-11-16 the model gave it to the batter, and it now goes to the % difference (Section 6). And in practice the interference class is catcher's interference only: the feed's catcher_interf is the only type seen as a plate-appearance result in 2015–2026, worth 0.308–0.433 runs a season.
| Event (LW1) | Runs vs average PA, 2026 | Count |
|---|---|---|
| Home run | +1.398 | 5,575 |
| Triple | +1.071 | 671 |
| Double | +0.762 | 7,493 |
| Single | +0.474 | 26,110 |
| Reached on error | +0.449 | 1,041 |
| Hit by pitch | +0.357 | 2,118 |
| Interference | +0.337 | 83 |
| Walk | +0.335 | 15,792 |
| Intentional walk | +0.234 | 545 |
| Bunt out | −0.191 | 1,136 |
| Fly-ball out | −0.249 | 24,860 |
| Line-drive out | −0.262 | 11,006 |
| Strikeout | −0.272 | 40,700 |
| Ground-ball out | −0.275 | 37,880 |
| Popup out | −0.279 | 8,839 |
Valuing outs by type matters little in total (a strikeout costs about the same as a ground-ball out once double plays and runner advancement are averaged in), but it keeps the batting value honest about what kind of out was made. It does not make the decomposition below exact: that identity is a telescoping sum and holds for any choice of classes (an earlier version of this paper said otherwise). What the classes decide is which layer a piece of value lands in.
Reached on error is the batter's. LW1 credits a batter who reaches on an error with nearly a single (+0.449 runs in 2026), where wOBA counts an out. The evidence that it is a batter skill is thin: a hitter's rate of reaching on error repeats at r = 0.13 [0.08, 0.18] pooled over nine season pairs (300+ PA both seasons; pairs range −0.06 to 0.32), and a club's detail value from reaching on error at 0.04 [−0.10, 0.19]. It is kept because ground-ball hitters and fast runners do reach more often, but it is the least defended choice in the batting value.
Check against FanGraphs. The model's wOBA-style team runs (LW0) track FanGraphs' team wRAA at r = 0.9997 [0.9994, 0.9999] in 2025 and 0.9995 [0.9988, 0.9997] in 2026 (30 clubs), so the run values themselves agree with the standard ones before any of the model's changes.
The decomposition identity
For every plate appearance the gap between its actual value and its wOBA-style value splits exactly into three layers:
RE24 - LW0 = (LW1 - LW0) detail: events wOBA lumps together (context-neutral)
+ (LW2 - LW1) timing: the base-out state the event came in (context)
+ (RE24 - LW2) play: what happened beyond the event class in that state
(runner advancement, double plays, errors on the play)
This identity is the backbone of the model: the detail layer stays with the batter, the timing layer is left out, and the play layer is credited to whoever was responsible (runners, the batter on ball-in-play outs, or nobody). By construction the play layer averages zero within every class-state cell, and the timing layer averages zero within every class.
5.The Brewers study: context versus repeatable skill
The model began with a question: did the 2026 Brewers (103–59, +214 run differential, World Series champions, projected before the season for about 81 wins, a figure not kept with the model's data) do the little things right in ways projection systems miss? Their record matched their Pythagorean expectation (Pythagenpat, exponent ((RS + RA) / G)0.287 = 1.876: 103.0 wins from 832 runs scored and 618 allowed), so the gap was in runs, not in turning runs into wins. FanGraphs WAR accounted for about 96 of the 103 wins: the club's 48.7 WAR in FanGraphs' team table plus a replacement-level club's 47.7 wins (162 / 2 − 1,000 / 30). Built as in Section 11, from each player's WAR times his share of time with the club, it is 49.9 WAR and 97.6 wins.
A trap in the obvious measurement
The first measurement, FanGraphs' team RE24 minus wRAA, put Milwaukee well above the league in situational value and showed the gap persisting from year to year across teams. The figures first quoted (+60 runs, r about 0.6) were read from FanGraphs on the study date and cannot be reproduced from the team files saved since; from those files the gap is +47.3 for Milwaukee in 2026 and +44.3 in 2025, and it repeats at r = 0.72 [0.48, 0.86] from 2025 to 2026. That persistence turned out to be a ballpark artifact: FanGraphs' RE24 is park-adjusted and its wRAA is not. The residual of FanGraphs' RE24 against the model's unadjusted RE24 is −81 runs for Colorado in 2026 (−95 in 2025) and correlates r = 0.88 [0.75, 0.94] (2026) and 0.89 [0.77, 0.94] (2025) with FanGraphs' own park and league adjustment (Batting minus wRAA). Subtracting the two FanGraphs figures mixes a park factor into the gap. The model therefore computes RE24 itself and never subtracts FanGraphs' two numbers.
What the clean decomposition showed
The gap is decomposed with the identity of Section 4 for every club and season. The year-to-year correlations were first read from one pair of seasons (2025 to 2026, 30 clubs, so a 95% interval about ±0.3 wide); they are now also pooled over every pair of consecutive seasons in 2015–2019, 2021–2026, 2020 left out (9 pairs, 270 club-pairs), with an interval from a bootstrap that resamples clubs (2,000 draws), since a club appears in several pairs.
| Brewers offense | Brewers 2026 | Brewers 2025 | r, 2025->2026 [95%] | r, all pairs pooled [95%] | In the model? |
|---|---|---|---|---|---|
| Gap, RE24 minus wOBA-style runs | +49.3 (1st) | +43.5 (1st) | 0.20 [−0.17, 0.52] | 0.21 [0.10, 0.33] | |
| Detail (outs by type, reached on error) | −2.0 (27th) | +7.2 (8th) | 0.10 [−0.27, 0.44] | 0.11 [−0.06, 0.30] | Yes, to batters |
| Timing | +14.3 (5th) | +8.3 (11th) | −0.08 [−0.42, 0.29] | 0.13 [0.00, 0.26] | No |
| Play, all plate appearances | +37.0 (1st) | +27.9 (1st) | 0.40 [0.05, 0.67] | 0.45 [0.35, 0.56] | |
| Play, on hits (runner advancement) | +10.2 (2nd) | +9.7 (3rd) | 0.02 [−0.34, 0.38] | 0.32 [0.18, 0.45] | Yes, to runners (rest to the % difference) |
| Play, on ball-in-play outs | +24.7 (1st) | +20.5 (1st) | 0.41 [0.06, 0.67] | 0.30 [0.18, 0.44] | Yes, to batters (situational) and runners |
| Play, on strikeouts | +1.3 (4th) | −2.0 (28th) | −0.09 [−0.43, 0.28] | 0.09 [−0.03, 0.21] | To runners (rest to the % difference) |
| Double-play avoidance, at league opportunity rates | +13.5 (1st) | +5.1 (7th) | 0.40 [0.04, 0.66] | 0.29 [0.17, 0.43] | Yes, to batters (situational) |
| Runner from third on outs, at league opportunity rates | +6.1 (4th) | +11.6 (2nd) | 0.04 [−0.32, 0.40] | 0.17 [0.04, 0.31] | Yes, to batters and runners |
Runs summed over the club's plate appearances, each measured against the league-average plate appearance, and rank among 30 clubs. The layers are not centered across clubs (2026 club means: gap +2.9, detail +2.9, timing 0.0, play 0.0; the two opportunity-rate rows average about zero). The play rows do not add to the play total: the play residual on walks and hit batters (+1.0 in 2026), reached on error (−0.2) and home runs (0.0) is not shown.
In the one-pair reading the persistent piece was value on ball-in-play outs (r = 0.41 [0.06, 0.67] against 0.02 [−0.34, 0.38] for play on hits). Pooled over nine pairs that is no longer so: play on hits (0.32 [0.18, 0.45]) persists as much as play on ball-in-play outs (0.30 [0.18, 0.44]); the difference in pooled Fisher z is +0.014 [−0.202, +0.225] (club bootstrap), and hits lead in 6 of 9 pairs. The ball-in-play-out part has two pieces. The situations and the double-play rate come from the study's script (brewers_detail.py); the scaling to league opportunity rates is not in that script: the study's figures for it (+13.5 and +5.1, +6.1 and +11.6) were recorded in the README without code, and paper_checks_core.py reproduces them exactly with this definition:
double-play avoidance = [ play residual on ground-ball OUTS with a runner on first, none on third, < 2 outs ]
/ [ the club's PAs with a runner on first, none on third, < 2 outs ]
x [ league share of PAs in that situation ] x [ the club's PAs ]
runner from third = the same over ball-in-play outs (GB, FB, LD, PU, bunt) and PAs with a runner on third, < 2 outs
double-play rate = double plays (2+ outs on the play) / ground-ball outs in the double-play situation
Milwaukee's hitters grounded into double plays on 34.5% of their 238 double-play-situation ground-ball outs in 2026, against 45.1% for the rest of MLB (39.6% vs 46.0% in 2025). (On defense the Brewers turned 47.6% of their double-play chances, ground-ball outs with a runner on first and fewer than two outs, against 43.9% for the league.) The denominator counts outs only: grounders that became hits or errors are not in it, and states with a runner on third are left out. Counting every ground-ball out with a runner on first, the rate is 33.2% against 44.3%. Avoidance was worth +13.5 runs at league opportunity rates in 2026 (1st) and +5.1 in 2025; scoring the runner from third +6.1 and +11.6. From 2025 to 2026 double-play avoidance repeated at r = 0.40 [0.04, 0.66] and runner-from-third at 0.04 [−0.32, 0.40]; the difference between the two is not significant (z = 1.40, p = 0.162, treating them as independent). Over all nine pairs they pool to 0.29 [0.17, 0.43] and 0.17 [0.04, 0.31]: both are distinguishable from zero (runner-from-third only just), but which persists more is not settled. The pooled difference in Fisher z is +0.120 [−0.053, +0.316] (club bootstrap, larger for double plays in 92% of draws), and double-play avoidance leads in only 5 of 9 pairs. The study read timing as not repeating at all (−0.08 [−0.42, 0.29] from 2025 to 2026); over nine pairs it pools to 0.13 [0.00, 0.26], only just distinguishable from zero. Reaching on error was not a factor: Milwaukee's detail value from it was +20.6 in 2026 against a club mean of +24.7 (27th).
Run prevention
The same plate appearances seen from the defense (prevention_decomp.py; Section 8 defines the layers), read from prevention_2025.json and prevention_2026.json:
| Brewers run prevention | Brewers 2026 | Brewers 2025 | r, 2025->2026 [95%] | r, all pairs pooled [95%] | In the model? |
|---|---|---|---|---|---|
| Strikeouts and walks | +73.0 (3rd) | +19.3 (10th) | 0.59 [0.29, 0.78] | 0.71 [0.63, 0.78] | Yes, to pitchers |
| Contact quality allowed | +36.1 (6th) | +40.7 (2nd) | 0.45 [0.10, 0.69] | 0.48 [0.34, 0.60] | Yes, to pitchers |
| Fielders' batted-ball conversion | +32.1 (5th) | +18.6 (9th) | 0.71 [0.47, 0.85] | 0.48 [0.33, 0.61] | Via DRS, to fielders |
| Running game | +5.3 (5th) | +2.6 (12th) | 0.58 [0.28, 0.78] | 0.24 [0.13, 0.37] | Yes, to pitchers and catchers |
| On the play | +7.5 (8th) | +2.8 (16th) | 0.04 [−0.33, 0.39] | 0.05 [−0.07, 0.16] | No (part reaches fielders through DRS) |
| Home runs beyond expected | −28.6 (28th) | −0.1 (14th) | 0.38 [0.03, 0.65] | 0.54 [0.39, 0.66] | No |
| Timing | −9.9 (20th) | +5.2 (11th) | −0.03 [−0.39, 0.33] | 0.03 [−0.13, 0.16] | No |
| Total of the layers | +115.4 (3rd) | +89.1 (3rd) | 0.60 [0.30, 0.79] | 0.61 [0.50, 0.72] |
Runs prevented relative to the league average (positive = fewer allowed) and rank. Actual runs allowed relative to the league average: +107.6 in 2026, +86.5 in 2025. The running-game layer here counts every runner event that is not a wild pitch, passed ball or balk as a steal or pickoff (defensive indifference included); the model's running game uses a narrower set (Section 6).
The layers add to the club's runs allowed: across the 30 clubs the total tracks actual runs allowed relative to the league average at r = 0.998 in 2026, within 4.8 runs per club on average. The remainder is the half-inning term of Section 4 seen from the defense. Per club, the sum over its defensive half-innings of RE24 plus runner-event changes minus runs (that is, RE of the final state minus RE of the opening state) tracks the remainder at r = −0.9998 in 2026: the more run expectancy a club's defensive half-innings open with, the more its layers overstate its runs prevented. Extra innings alone (the automatic runner's opening state at second, and walk-off endings) give r = −0.85. The number of plate appearances a club defended explains almost none of it (r = −0.26; removing a linear term in plate appearances moves the mean absolute remainder only from 4.79 to 4.63 runs).
Two conclusions shaped the model. First, a model of player value should keep what can be attributed to a player and does not depend on the base-out state; persistence is evidence, not the rule by itself. Timing is left out because it is context by definition, and it barely repeats (0.13 [0.00, 0.26] pooled). Play on hits, first read as 0.02 [−0.34, 0.38] from one pair of seasons, the lowest of the nine, pools to 0.32 [0.18, 0.45], about as much as play on ball-in-play outs (0.30 [0.18, 0.44]); both are kept and credited to the runners and batters responsible. Detail pools to 0.11 [−0.06, 0.30], not distinguishable from zero, and is kept because it is the batter's own event. Home runs beyond expected repeat strongly at the club level (0.54 [0.39, 0.66]), because a club's home park repeats; for individual pitchers they repeat at 0.13 [0.08, 0.18] per batter faced (250+ both seasons), against 0.29 [0.23, 0.36] for contact quality, and they are left out (Section 8). Second, contact quality matters: Milwaukee's staff ranked 6th and 2nd at suppressing contact in 2026 and 2025 although only 60% of its 2026 batters faced came from pitchers who had also pitched for it in 2025, which FIP-based measures cannot see. That is why pitchers are credited with the expected value of the contact they allow, not just strikeouts, walks and home runs.
What the finished model says about the Brewers
The question the study began with, answered by the model it led to (team totals of the model components, players split by time with the club):
| Component | Brewers 2026 ledger | Brewers 2025 ledger |
|---|---|---|
| Batting | +59.6 (3rd) | +60.9 (8th) |
| Situational outs | +15.8 (1st) | +16.2 (1st) |
| Baserunning | +13.8 (3rd) | +10.9 (2nd) |
| Fielding and catcher defense | +44.8 (4th) | +14.7 (9th) |
| Strikeouts and walks | +82.7 (3rd) | +23.7 (7th) |
| Contact quality | +32.9 (5th) | +38.2 (2nd) |
| Running game and pitchers' fielding | −6.8 (26th) | +1.8 (13th) |
| Players runs total | +242.8 (1st) | +166.5 (5th) |
| Actual run differential | +214 | +172 |
Runs and rank among 30 clubs. % difference in runs (actual run differential minus players runs total): −28.8 in 2026, +5.5 in 2025.
Situational outs, the measured version of "the little things", put Milwaukee 1st in 2026 and 1st in 2025, but at +15.8 runs it is a small part of a players runs total of +243, which itself exceeds the actual run differential of +214.
6.Hitter components
A hitter's runs are the sum of four components (five for catchers): batting, situational outs, baserunning, fielding, and catcher defense. All are runs relative to an average player and all add to zero across the league.
Batting
Batting is the LW1 value of each plate appearance minus a park adjustment:
batting(PA) = LW1(event class) - mu - (PF(park) - 1) x league runs per PA
PF(team) = 1 + 0.5 x ( [runs per PA in that team's home games, both clubs, two seasons]
/ [runs per PA in that team's road games, both clubs, two seasons] - 1 )
league runs per PA = all runs in the two seasons (PA runs + runner-event runs) / all PAs
batting is then shifted by its league mean per PA, so it sums to exactly zero
The park factor pools the season and the one before (so 2015 uses 2014) and is regressed halfway to neutral because two seasons of home/road splits are noisy. Runs include runner-event runs. The league rate in the adjustment is the two-season pooled one (0.1183 runs per PA for 2026). The adjustment applies to every plate appearance in the park, home and visiting hitters alike. The 2026 factors at the ends of the range:
| Park | PF 2026 |
|---|---|
| COL | 1.105 |
| ATH | 1.089 |
| PHI | 1.072 |
| LAD | 1.034 |
| WSH | 1.029 |
| … | |
| SD | 0.957 |
| MIA | 0.948 |
| STL | 0.948 |
| TEX | 0.936 |
| SEA | 0.935 |
Three limits of the park factor, none of them fixed:
- It is keyed by the home club, not the ballpark. The MLB play-by-play files drop the park id (Section 3), so every home game of a club counts as its park: neutral-site games (international series and special-event games) count as the home club's park, and a club that changed parks mixes two parks in its two-season pool, as for the Blue Jays in 2020–21, the Athletics' move to Sacramento in 2025 and the Rays' 2025 season at Steinbrenner Field. The minor-league ledgers key parks by venue.
- The halfway regression is fixed, the same for two full seasons as for a pool that includes 2020's 60 games. It was assumed, not fitted. The minor-league ledgers regress by n / (n + 10,000) PA instead, which is about halfway for two seasons.
- No road-schedule correction. The road denominator is the club's own road games, at whatever mix of parks its schedule held; the factors are not iterated.
Pitchers' values are not park-adjusted, because their main inputs (strikeouts, walks, and the expected value of contact from exit velocity and launch angle) depend much less on the park than runs do. That was first an assumption; it is now tested. For each club, the per-600 difference between its home games and its road games (both clubs' plate appearances) is computed for the pitcher parts and for the batting value, and its spread across the 30 parks is set beside the spread sampling alone would give, for the last two seasons pooled (the window of a park factor) and for all 2015–2026:
| Value (runs per 600 PA) | Home minus road, SD across parks | Expected from sampling | Excess over noise | Extremes | Coors Field (COL) |
|---|---|---|---|---|---|
| Pitchers' strikeouts and walks, 2025–2026 | 2.6 | 1.3 | 2.2 | ATH −4.0 / HOU +6.3 | −2.0 |
| Pitchers' strikeouts and walks, 2015–2026 | 2.2 | 0.5 | 2.1 | COL −4.5 / SEA +4.8 | −4.5 |
| Pitchers' contact quality, 2025–2026 | 1.8 | 2.0 | 0.0 | KC −3.7 / DET +4.3 | −1.1 |
| Pitchers' contact quality, 2015–2026 | 1.8 | 0.8 | 1.6 | KC −3.6 / HOU +4.3 | −1.1 |
| Batting value (LW1), 2025–2026 | 7.0 | 3.3 | 6.2 | SEA −15.1 / ATH +19.1 | +15.8 |
| Batting value (LW1), 2015–2026 | 5.6 | 1.4 | 5.4 | SEA −10.7 / COL +20.7 | +20.7 |
Excess over noise = square root of (observed variance minus the mean squared standard error). Pitcher parts are runs prevented; batting value is from the hitters' side.
The pitcher parts do have a park component, smaller than the batting value's: over 2015–2026 an excess spread of 2.1 runs per 600 for strikeouts and walks and 1.6 for contact, against 5.4 for the batting value (Coors Field: −4.5 strikeouts and walks, +20.7 batting value). In two seasons, the contact spread (1.8) is within sampling noise (2.0). The park's effect on pitchers lands in the % difference.
Situational outs
This component is where the ball-in-play-out part of the Brewers finding lives (double plays avoided, runners moved and scored on outs). At the club level it persists (0.30 [0.18, 0.44] pooled, Section 5); at the player level it is the least repeatable hitter component (below). On a hitter's ball-in-play outs (LW1 classes BIP?, BUNT, FB, GB, LD, PU: sacrifice flies, sacrifice bunts, double plays and fielder's choices included; strikeouts, including strikeout double plays, and reached-on-error are not), the play layer (RE24 minus LW2) is the value of what happened beyond the average for that kind of out in that base-out state: double plays avoided or hit into, forced runners, runners moved up. After the credits to runners who were not forced are taken out (below), the remainder belongs to the batter: his own deviation as batter-runner and the deviations of runners forced on his ground-ball and bunt outs (a double play is the batter's doing, not the runner's) stay with him.
The opportunities are plate appearances with runners on and fewer than two outs. To avoid rewarding a hitter for batting behind high-OBP teammates, the value is rated per opportunity and applied at the league-average rate of those opportunities, separately against each pitcher hand h:
raw(h) = sum over his ball-in-play outs vs hand h of [ RE24 - LW2 - credits to runners not forced ]
situational(h) = raw(h) / [his PAs vs h with runners on and < 2 outs] x 0.257 x [his PAs vs h]
= 0 if he had no such PA against h
then: situational(h) -= league situational per PA x [his PAs vs h] (sums to zero)
25.7% of league PAs were situational in 2026. The numerator includes every ball-in-play out, also those with the bases empty or two out (whose play residual is the batter-runner's small deviation), while the denominator counts only situational plate appearances of any kind; a hitter with no situational PA against a hand loses that hand's numerator.
Fixed: fielder's choices that retire nobody. The feed has two fielder's choices: fielders_choice_out retires a runner, and fielders_choice retires nobody (all 4,315 of them in 2015–2026 record no out). The second is a ball-in-play out in LW1 (its trajectory class), so its play residual is large and positive (+1.22 runs each in 2026). Until 2026-11-16 the model gave the batter what the runners' own credits left of it: his own batter-runner deviation, the forced runners' deviations on grounders, and the remainder. The lead runner who was not retired is usually forced, so that was most of it (96% in 2026): measured before the fix (2026: 408 plays, +496.0 runs of play residual, +475.3 of it to batters before the opportunity scaling; 4,315 plays and 4,800 runs over 2015–2026, frozen in fc_fix_history.json). That is the fielders' choice, not the batter's skill. Since the fix those parts are credited to no one and go to the % difference: ledger.py skips the batter's credits on such a play, keeps the residual out of his situational value, and records what it sends there (meta.fc_no_out; milb_ledger.py does the same). Non-forced runners keep their own credits, and the batter keeps his batting value, an out of that trajectory. The replica in paper_checks_core.py reproduces the model's record exactly in every season. Effects. Only situational outs moved: their absolute changes sum to 4,841 runs over 2015–2026, the largest a player-season 5.80; batting, baserunning and fielding did not move, and the pitchers and the runner weights are identical (the weights are fitted on every runner row, whoever is credited). The 2,898 hitter-seasons with such plays lost 2,176.3 runs of situational value in all and the rest gained the same, because situational outs are recentered per plate appearance; per play the loss averaged 0.35 to 0.76 runs in a season, well under the play's raw credit, because the recentering spreads the removed credit back over every hitter's plate appearances, those with such plays included. The largest change is Alexei Ramirez in 2015 (5 such plays, +2.44 to −3.36). The players-against-run-differential correlation moved by −0.0040 to +0.0046 in a season, not consistently one way, and the play residual no one is credited with (Section 10) rose by the same runs, from +2.3–+18.3 to +179.5–+503.6 a season. The credits behaved like noise: the pooled year-to-year r of hitters' situational outs rose from 0.165 [0.119, 0.209] to 0.259 [0.215, 0.301] (the intervals do not overlap), and the forecast, which fits its pull toward average by back-test, now pulls situational outs less (below). In the minor leagues 8,520 such plays sent +9,756 runs to the difference; the pitchers' records came out identical. The translation's situational piece now has slopes 0.53 (Triple-A) and 0.59 (Double-A), from 0.64 and 0.31, with k 300 (from 1,000), a third set in one day (the automatic-runner fix had moved them from 0.35 and 0.12 with k 300), so the piece is weakly identified: small changes to minor-league situational outs move it a long way. Its 2025–26 test for hitters moved from 17.239 to 17.259 runs per 600 PA (all 767) and from 20.391 to 20.326 (the 99 with no MLB time). In the 2027 forecast hitters' situational outs moved by 0.32 runs per 600 on average (largest −1.73 for Ali Sánchez), the fitted k for situational outs is 7,000 (was 10,000), a team's projected situational outs moved by at most 2.27 runs, and the 2026 back-test went from 16.909 to 16.913 for hitters.
Line-drive double plays are not forced-runner plays: the runner doubled off is charged on his baserunning. In 2026, 268 line-drive double plays carried −123.8 runs of play residual, −97.7 to the runners and −26.1 to the batters.
How much it repeats. Situational value is the least repeatable hitter component. It was first described from one pair of seasons (r about 0.20, 2025 to 2026, before the fielder's-choice fix above raised it); over all nine pairs, per PA, hitters with 300+ PA in both seasons:
| Component (per PA) | r, 2025->2026 | Range over 9 pairs | Pooled [95%] |
|---|---|---|---|
| Batting | 0.41 [0.29, 0.52] | 0.41–0.60 | 0.51 [0.45, 0.57] |
| Situational outs | 0.33 [0.20, 0.45] | 0.20–0.33 | 0.26 [0.21, 0.31] |
| Baserunning | 0.44 [0.32, 0.55] | 0.43–0.64 | 0.51 [0.47, 0.56] |
| Fielding | 0.43 [0.31, 0.54] | 0.38–0.54 | 0.47 [0.41, 0.52] |
| Runs (total) | 0.40 [0.27, 0.51] | 0.40–0.61 | 0.49 [0.44, 0.54] |
| Value with position | 0.43 [0.31, 0.54] | 0.42–0.61 | 0.50 [0.46, 0.55] |
Pairs of consecutive seasons 2015–2026 without 2020. Pooled: Fisher z weighted by n − 3; interval from 1,000 bootstrap draws resampling players, which keeps a player's repeated pairs together.
The forecast's back-test pulls it hard toward average: its k is 7,000 weighted PA, the largest value on the search grid (forecast.K_GRID stops at 10,000), so the fit may want more. With the hitters' 5/4/3 season weights, a regular with three 600-PA seasons (U = 7,200 weighted PA) keeps U / (U + k) = 51% of his weighted rate, and a hitter with one 600-PA season (3,000) keeps 30%. "Almost entirely to average", as this was first put, overstates it for regulars.
Baserunning
Baserunning combines advancement on every plate appearance except home runs and the running game between pitches.
Advancement. For every runner on every plate appearance except home runs (including the batter-runner), the destination is compared with what an average runner achieved in the same cell:
cell = (LW1 class, start base state, outs, runner's start base, exit-velocity band)
bands: under 85, 85-95, 95+ mph, and a fourth band for no exit velocity (every strikeout, walk
and hit batter, and a batted ball the feed did not measure)
the band is dropped (cell = class, state, outs, base) when the banded cell has fewer than 30 runners
a_safe = 0 if out on the play, else max(destination - start base, 0) destination: 1, 2, 3, home = 4;
unknown = stayed
d_adv = a_safe - cell mean of a_safe
d_out = 1[out on the play] - cell out rate
forced = runner's start base > 0 and every base below it occupied when the pitch was thrown
Cells are means over the season's own runner rows (each runner is in his own cell's mean). Missing exit velocity was 8.8%–13.1% of batted balls in 2015–2019 and under 1% from 2020 (Section 8), so in the early seasons the fourth band is not small. 97.3% of 2026's runner rows used the banded cell.
The run value of an extra base and of an out on the bases are not assumed: they are fitted each season by an ordinary least-squares regression, without intercept, of every non-home-run plate appearance's play-layer residual on the sum of its runners' deviations, the batter-runner's included:
RE24 - LW2 = w_adv x sum(d_adv) + w_out x sum(d_out) + e one row per PA
The weights are stable across seasons and precisely estimated (the replica in paper_checks_core.py reproduces every ledger’s weights exactly):
| Season | Runs per extra base (SE) | Runs per out on the bases (SE) | R2 | Fine cells used | PAs |
|---|---|---|---|---|---|
| 2015 | +0.228 (0.0006 / 0.0020) | −0.342 (0.0014 / 0.0047) | 0.684 | 96.8% | 178,719 |
| 2016 | +0.240 (0.0006 / 0.0019) | −0.319 (0.0014 / 0.0045) | 0.697 | 96.7% | 178,970 |
| 2017 | +0.234 (0.0006 / 0.0019) | −0.358 (0.0014 / 0.0046) | 0.712 | 96.7% | 179,188 |
| 2018 | +0.240 (0.0006 / 0.0019) | −0.331 (0.0014 / 0.0045) | 0.703 | 96.8% | 179,554 |
| 2019 | +0.237 (0.0006 / 0.0018) | −0.376 (0.0014 / 0.0047) | 0.730 | 96.8% | 179,741 |
| 2020 | +0.222 (0.0011 / 0.0033) | −0.399 (0.0023 / 0.0082) | 0.711 | 90.8% | 64,202 |
| 2021 | +0.230 (0.0006 / 0.0020) | −0.356 (0.0014 / 0.0050) | 0.704 | 97.5% | 175,874 |
| 2022 | +0.235 (0.0006 / 0.0019) | −0.335 (0.0014 / 0.0048) | 0.704 | 97.2% | 176,837 |
| 2023 | +0.243 (0.0006 / 0.0018) | −0.345 (0.0014 / 0.0048) | 0.716 | 97.4% | 178,236 |
| 2024 | +0.240 (0.0006 / 0.0020) | −0.346 (0.0014 / 0.0050) | 0.701 | 97.4% | 176,996 |
| 2025 | +0.222 (0.0006 / 0.0020) | −0.383 (0.0015 / 0.0053) | 0.690 | 97.5% | 177,276 |
| 2026 | +0.234 (0.0006 / 0.0021) | −0.362 (0.0014 / 0.0053) | 0.704 | 97.3% | 178,274 |
Standard errors: classical / heteroskedasticity-robust (HC0). The play residual is far from homoskedastic (a double play and a routine out share a cell), so the robust ones are the ones to read; both are small against the season-to-season movement of the weights. R2 is uncentered (no intercept).
Each runner's credit is w_adv x d_adv + w_out x d_out. Credits go to the runner, with two exceptions: the batter-runner's own deviation goes to the batter (to his baserunning on every plate appearance that is not a ball-in-play out, to his situational value on ball-in-play outs), and runners forced on ground-ball and bunt outs go to the batter's situational value. Whatever the runner credits leave of the play residual on plate appearances that are not ball-in-play outs (hits, strikeouts, walks and hit batters, reached on error, interference) goes to the % difference (much of it on hits is outfield arms and hit location). Two edge cases are in the code as it stands, and a third records two that were fixed:
- A batter who reaches on a dropped third strike gets baserunning credit for it (52 in 2026, +32.1 runs, about 0.62 each), although a wild pitch or passed ball usually made it.
- A steal or caught stealing on the plate appearance's final pitch (a runner going on a full-count pitch that ends in a walk or strikeout) is part of the plate appearance, so it is valued in the advancement cells, not as a steal: it is not centered on the league steal rate and not charged to the pitcher or the catcher's throwing (350 runners in 2026, +21.2 runs of advancement credit). The feed gives no wild-pitch or passed-ball reason on any final-pitch movement (0 in 2015–2026), so advances on a wild pitch that ends a plate appearance, if any, cannot be told from ordinary advancement.
- The pinch-runner and automatic-runner rows of Section 3, both fixed on 2026-11-16. Before the fixes 5,075 replaced runners were written as staying put (over 2015–2026 the rows as written totaled −745 runs against +55 for the moves that happened, an absolute error of 1,498 runs), and 3,793 automatic-runner stays were never written (−104.3 runs the automatic runners were spared). The weights in the table above are the refitted ones. Before the pinch-runner fix the runs per extra base were 0.0071 to 0.0098 lower in each season and the runs per out on the bases 0.0013 to 0.0067 further from zero; the automatic-runner fix moved either weight by at most 0.0003.
Running game. Runner events between pitches are valued by their change in run expectancy, runs included. Those whose type starts with stolen_base, caught_stealing or pickoff (pickoff throwing errors included: +37.0 runs to runners in 2026) go to the runners, split equally among the runners in the event (a double steal), and the opposite to the pitcher. Wild pitches, passed balls and balks are the battery's, not the runner's. Everything else is credited to nobody. In 2026, over every row of the runner-event table (mid-PA groups and plays without a plate appearance):
| Event | Events | Change in RE, runs | Credited to |
|---|---|---|---|
| stolen base 2b | 2,541 | +367.9 | the runner(s); the pitcher, less the catchers' throwing |
| wild pitch | 1,341 | +334.7 | the pitcher (battery), less the catchers' blocking |
| caught stealing 2b | 626 | −248.6 | the runner(s); the pitcher, less the catchers' throwing |
| stolen base 3b | 348 | +86.6 | the runner(s); the pitcher, less the catchers' throwing |
| defensive indiff | 300 | +22.9 | nobody (% difference) |
| passed ball | 208 | +49.6 | the pitcher (battery), less the catchers' blocking |
| pickoff 1b | 185 | −78.1 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff caught stealing 2b | 143 | −55.4 | the runner(s); the pitcher, less the catchers' throwing |
| balk | 132 | +33.0 | the pitcher (battery), less the catchers' blocking |
| pickoff error 1b | 88 | +18.4 | the runner(s); the pitcher, less the catchers' throwing |
| caught stealing 3b | 76 | −40.0 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff error 2b | 51 | +13.4 | the runner(s); the pitcher, less the catchers' throwing |
| other out | 33 | −17.1 | nobody (% difference) |
| pickoff 2b | 26 | −15.5 | the runner(s); the pitcher, less the catchers' throwing |
| forced balk | 22 | +4.4 | nobody (% difference) |
| caught stealing home | 17 | −6.7 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff caught stealing 3b | 15 | −8.5 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff error 3b | 10 | +5.1 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff 3b | 9 | −7.1 | the runner(s); the pitcher, less the catchers' throwing |
| stolen base home | 8 | +6.0 | the runner(s); the pitcher, less the catchers' throwing |
| pickoff caught stealing home | 8 | −6.7 | the runner(s); the pitcher, less the catchers' throwing |
| error | 5 | +1.6 | nobody (% difference) |
Change in run expectancy summed over the season's events, runs scored included, before centering. The catchers' throwing and blocking (Statcast) are taken back out of the pitchers (Section 7).
Opportunity normalization. A runner's chances depend on how often teammates put the ball in play behind him. Advancement is rated per opportunity and applied at the league rate of opportunities per stint on base; steals are centered on the league steal value per stint:
advancement = (credits / opportunities) x (league opportunities / league stints) x stints
(= credits unscaled if he had no opportunity)
steals = steal-type runs - league steal-type runs per stint x stints
opportunity = one runner row credited to him as a runner: on base when a PA ended, not a home run,
not forced on a ground-ball or bunt out (those rows are the batter's)
stint = a distinct (game, half-inning) in which he was on base: in a runner row as a runner,
or in a steal-type event
A stint is a half-inning, not a stay: a runner who reaches twice in one half-inning has one stint. The batter-runner's own credits (stretching a hit, taking a base on a throw) are added to his advancement credits and scaled by the same factor, although they are not counted as opportunities. The baserunning total is recentered to zero per stint and split between pitcher hands by the share of the runner's opportunities against each hand (by plate-appearance share if he had none); the hand of the pitcher on a steal is not used.
7.Fielding, catcher defense and the positional adjustment
Why DRS
Fielding is the one major component the play-by-play cannot assign to individual fielders without spray angles and fielder positioning, so it comes from published metrics. To choose between Defensive Runs Saved and Statcast's fielding run value, each club's non-catcher fielding under each metric was compared with the model's own team measure of batted-ball conversion (actual results minus expected results from exit velocity and launch angle, the in-park part of Section 8, plus the play layer on defense). The test was first run once and its numbers typed into this paper; it is now saved (paper_checks_core.py), and rerunning it changed the conclusion's footing:
| Fielding measure (club total) | r with conversion + play, 2025 | r with conversion + play, 2026 |
|---|---|---|
| DRS minus catcher parts (rSZ, rCERA, rSB), FanGraphs team table: the test as first run | 0.86 [0.73, 0.93] | 0.82 [0.65, 0.91] |
| The same without team shift runs (rTS) | 0.79 [0.61, 0.90] | 0.79 [0.59, 0.89] |
| DRS minus catcher parts, the model's player rows summed by time share: the input the model uses | 0.76 [0.55, 0.88] | 0.81 [0.64, 0.91] |
| Statcast FRV (range + arm + infield DP + 1B receiving), FanGraphs team table | 0.57 [0.27, 0.77] | 0.73 [0.50, 0.86] |
| Statcast FRV, the model's player rows | 0.58 [0.28, 0.78] | 0.71 [0.47, 0.85] |
| 50/50 blend, team table | 0.82 [0.66, 0.91] | 0.85 [0.70, 0.93] |
| 50/50 blend, player rows | 0.73 [0.51, 0.86] | 0.83 [0.66, 0.91] |
| Team shift runs (rTS) alone | 0.46 [0.13, 0.71] | 0.21 [−0.16, 0.53] |
30 clubs; 95% Fisher intervals. Player rows: each player's fielding times the share of his own plate appearances that came with the club, as in the team ledger.
The first version of this table read 0.86 / 0.82 for DRS, 0.47 / 0.76 for Statcast and 0.78 / 0.86 for the blend. The DRS row reproduces exactly from FanGraphs' team fielding table, but that table's DRS includes the team shift and positioning runs (rTS: +497 runs league-wide in 2025, +579 in 2026) that no player row carries, so the test validated a number the model does not use. The Statcast and blend rows could not be reproduced from any saved file with the definitions stated; the reproducible values are above. Three findings:
- On the model's own input (player rows), DRS and the blend are statistically indistinguishable. DRS against the blend: 2025 t = 0.91, p = 0.370; bootstrap difference −0.03 [−0.10, +0.03]; 2026 t = −0.33, p = 0.745; bootstrap difference +0.01 [−0.05, +0.08] (Williams' test for correlations sharing the club outcome, df = 27, positive t when DRS correlates more; bootstrap of the 30 clubs, 4,000 draws, blend minus DRS). DRS against Statcast alone: 2025 t = 2.13, p = 0.043; bootstrap difference −0.18 [−0.36, −0.01]; 2026 t = 1.32, p = 0.198; bootstrap difference −0.11 [−0.25, +0.04]. Even on the team table with rTS, the blend beat DRS in 2026, as the first table already showed.
- Over every season with Statcast fielding (2016–2026, player rows), DRS pools to r = 0.72 [0.66, 0.77], Statcast to 0.63 [0.56, 0.69] and the blend to 0.74 [0.68, 0.78] (Fisher z weighted, seasons treated as independent). The blend's pooled advantage in Fisher z is +0.038 [−0.017, +0.098] (bootstrap of clubs, 2,000 draws), positive in 92% of draws.
- Against conversion alone (the earlier, README test), FanGraphs' team DRS correlates 0.79 [0.60, 0.89] (2025) and 0.76 [0.56, 0.88] (2026), OAA 0.53 [0.21, 0.75] and 0.63 [0.35, 0.81], FRP 0.44 [0.09, 0.69] and 0.68 [0.42, 0.83].
| Season | DRS r | Statcast r | Blend r | Williams test, DRS vs blend |
|---|---|---|---|---|
| 2016 | 0.51 [0.19, 0.74] | 0.40 [0.05, 0.67] | 0.55 [0.23, 0.76] | −0.52 (p 0.605) |
| 2017 | 0.66 [0.40, 0.83] | 0.53 [0.21, 0.75] | 0.69 [0.44, 0.84] | −0.43 (p 0.668) |
| 2018 | 0.75 [0.54, 0.88] | 0.71 [0.47, 0.85] | 0.79 [0.59, 0.89] | −1.09 (p 0.286) |
| 2019 | 0.75 [0.54, 0.88] | 0.68 [0.42, 0.83] | 0.77 [0.56, 0.88] | −0.57 (p 0.575) |
| 2020 | 0.69 [0.44, 0.84] | 0.68 [0.42, 0.83] | 0.75 [0.53, 0.87] | −1.45 (p 0.158) |
| 2021 | 0.68 [0.42, 0.83] | 0.65 [0.38, 0.82] | 0.72 [0.49, 0.86] | −1.17 (p 0.253) |
| 2022 | 0.78 [0.58, 0.89] | 0.73 [0.50, 0.86] | 0.79 [0.60, 0.90] | −0.45 (p 0.659) |
| 2023 | 0.73 [0.50, 0.86] | 0.54 [0.23, 0.76] | 0.73 [0.50, 0.86] | −0.04 (p 0.972) |
| 2024 | 0.69 [0.43, 0.84] | 0.61 [0.32, 0.79] | 0.68 [0.43, 0.84] | 0.08 (p 0.933) |
| 2025 | 0.76 [0.55, 0.88] | 0.58 [0.28, 0.78] | 0.73 [0.51, 0.86] | 0.91 (p 0.370) |
| 2026 | 0.81 [0.64, 0.91] | 0.71 [0.47, 0.85] | 0.83 [0.66, 0.91] | −0.33 (p 0.745) |
Player rows summed to clubs by time share against conversion plus on-the-play, 30 clubs. 2015 has no Statcast fielding.
On this evidence DRS is not significantly worse than the blend and is better than Statcast alone, not the best of the three. The blend's correlation is higher than DRS's in 9 of 11 seasons and pooled (0.74 against 0.72), and wins 92% of the bootstrap draws, but its pooled advantage (+0.038 [−0.017, +0.098] in Fisher z) is not distinguishable from zero. DRS beats Statcast in 11 of 11 seasons, by +0.165 [+0.044, +0.302] in pooled Fisher z (100% of draws). DRS stays the model's fielding source because the choice predates the saved test: the README records it as chosen because it "tracks team conversion best", on the first table above, which the reproducible test no longer supports against the blend. Part of DRS's agreement with conversion is by construction (DRS credits plays made by zone, closer to an outcome measure than Statcast's catch-probability model). Switching to the blend would cost nothing on this evidence and may help a forecast; it has not been made. The dashboard shows the Statcast figure in every player's breakdown as a second opinion.
The fielding calculation
fielding(player) = sum over his non-pitcher positions of [DRS - rSZ - rCERA - rSB]
- league non-pitcher rate per inning x his innings at non-pitcher positions
pitcher fielding = [DRS - rSZ - rCERA - rSB] at P - league pitcher rate per inning x his innings at P
league rate = league sum of [DRS - rSZ - rCERA - rSB] / league innings, positions and pitchers apart
The catcher parts are subtracted on every row, and they are not zero for pitchers: in 2026, 254 pitchers carry a nonzero rSB (stolen-base runs, 314 absolute runs in all) and 868 a nonzero rSZ (199). Removing them from pitchers is right, since the pitcher's running game is charged through run expectancy (Section 8) and his strikeouts and walks through the plate appearances, but the formula given before omitted it. The centering is per defensive inning, separately for position players and pitchers. The reason first given for it, that DRS averaged about +19 runs per club in 2026, came from FanGraphs' team table (rTS again). The player rows the model centers sum to −34 to +151 runs a season league-wide (−1.1 to +5.0 per club), so the centering moves a club by a few runs at most:
| Season | Non-P DRS less catcher parts | Per club | Pitcher DRS less parts | Pitchers with rSB | Team-endpoint DRS | Team shift runs (rTS) |
|---|---|---|---|---|---|---|
| 2015 | +151 | +5.0 | −38 | 201 (255) | – | – |
| 2016 | +105 | +3.5 | −16 | 202 (260) | – | – |
| 2017 | +126 | +4.2 | −28 | 225 (268) | – | – |
| 2018 | +80 | +2.7 | −29 | 215 (260) | – | – |
| 2019 | +123 | +4.1 | −32 | 201 (236) | – | – |
| 2020 | +35 | +1.2 | −3 | 78 (85) | – | – |
| 2021 | −25 | −0.8 | −26 | 225 (267) | – | – |
| 2022 | −1 | 0.0 | −21 | 220 (245) | – | – |
| 2023 | −2 | −0.1 | −26 | 283 (363) | – | – |
| 2024 | +1 | 0.0 | −42 | 288 (353) | – | – |
| 2025 | −34 | −1.1 | −25 | 274 (355) | +427 | +497 |
| 2026 | +10 | +0.3 | −21 | 254 (314) | +569 | +579 |
League totals in runs from fg_players_fld_<season>.json (team columns from fg_team_fld_<season>.json, on disk for 2025 and 2026 only). Pitchers with rSB: count and absolute runs.
FanGraphs is queried with team=0, so a traded player has one combined row per position (team - - -; 224 such rows in 2026) and no innings by club. Fielding is therefore hand-neutral, spread across a hitter's splits by plate-appearance share, and allocated to his clubs by plate-appearance share, not by where he fielded. A fielder with no plate appearance has no hitter record and his fielding is dropped (2 in 2026). The positional label on a hitter's record is the position with the most non-pitcher innings, or DH if none, so pitchers who batted are labeled DH (2112 of the 2128 pitcher-batter seasons).
Catcher defense
Catchers are credited with Statcast framing, blocking and throwing (FanGraphs API fields fFRP, bFRP, tFRP), each centered per catcher inning. Those runs show up in the pitching numbers too (framing changes strikeouts and walks; throwing and blocking change the running game), so the same runs, after centering, are taken back out of the team's pitchers, in proportion to batters faced:
team framing(t) = sum over catchers of centered framing x PA(catcher, t) / PA(catcher) pitcher strikeouts & walks -= team framing(t) x (pitcher BF for t / team BF) pitcher running game -= team (throwing + blocking)(t) x (pitcher BF for t / team BF) (the strikeouts-and-walks offset is spread over his batter-hand splits by batters-faced share)
A traded catcher's framing is allocated to clubs by his plate-appearance share, like his fielding. In 2015 FanGraphs' Statcast catcher fields are zero on every row (framing, blocking and throwing absolute totals 0, 0, 0), so 2015 catcher defense is zero, 2015 pitchers keep their catchers' framing in strikeouts and walks and their throwing in the running game, and 2015 catchers lose DRS's rCERA and rSB (subtracted above) with nothing in their place.
Field-name trap. In FanGraphs' fielding API, fFRP is framing and cFRP is first-base receiving, not catcher framing. The mapping was verified against the CSV exports (within about a run on every row).
Positional adjustment
The positional adjustment is FanGraphs' Positional runs from its batting table. It is shown as a separate column and added only in "Value with position" for comparing players across positions; it is never part of team totals. It does not cancel in a team total (Section 2):
| Season | League sum | Club mean | Club SD | Lowest club | Highest club | Pitchers who batted |
|---|---|---|---|---|---|---|
| 2015 | +267 | +8.9 | 24.3 | TB −17.4 | CHC +38.7 | 328 (+557) |
| 2016 | +286 | +9.5 | 25.0 | HOU −18.5 | CHC +40.9 | 336 (+572) |
| 2017 | +341 | +11.4 | 26.3 | ATH −16.8 | WSH +42.3 | 333 (+627) |
| 2018 | +326 | +10.9 | 26.1 | DET −18.4 | WSH +40.7 | 362 (+614) |
| 2019 | +340 | +11.3 | 26.3 | DET −16.5 | LAD +43.3 | 355 (+623) |
| 2020 | −218 | −7.3 | 0.5 | CIN −8.6 | SD −6.1 | 4 (+1) |
| 2021 | +286 | +9.5 | 24.4 | ATH −16.4 | LAD +41.4 | 390 (+566) |
| 2022 | −580 | −19.3 | 1.3 | SF −23.6 | SD −16.4 | 5 (+1) |
| 2023 | −590 | −19.7 | 1.4 | SEA −23.2 | LAD −16.9 | 4 (+1) |
| 2024 | −584 | −19.5 | 1.6 | AZ −22.9 | SF −16.4 | 2 (0) |
| 2025 | −581 | −19.4 | 1.3 | TEX −22.8 | LAA −17.2 | 6 (+1) |
| 2026 | −604 | −20.1 | 1.9 | DET −24.9 | MIA −14.8 | 3 (0) |
Runs. Club sums weight each player by the share of his own plate appearances that came with the club. Pitchers who batted: hitters who also faced more batters than they batted, their number and summed positional runs.
Before the universal designated hitter the pitchers who batted carried +557 to +627 positional runs a season league-wide (Madison Bumgarner alone +10.5 in 97 PA in 2016), nearly all on National League clubs. Leaving the adjustment out of team totals is correct; the reason is that it would add a near-constant (2022 on) or a league bias (before), not that it cancels.
8.Pitcher components
A pitcher's runs prevented are the sum of strikeouts and walks, contact quality, running game, and his own fielding. Each plate appearance's value from the defensive side splits exactly into layers; the pitcher owns the first two.
for each PA, v = RE24 - mu, sign flipped so positive = runs prevented:
strikeouts & walks : classes BB, HBP, IBB, INTERF, K at their LW1 value;
every batted ball at the league average batted-ball value b
contact quality : xv - b, the expected value of the batted ball minus the league average
conversion : what the fielders made of it (fielders, park, luck)
HR vs expected : home runs beyond what the contact predicted (park, luck)
timing : LW2 - LW1 (context)
on the play : RE24 - LW2 (fielders and runners)
the six add to v exactly; b = mean LW1 value of all batted balls (+0.039 in 2026)
The strikeouts-and-walks class list includes interference, which in practice is catcher's interference (Section 4): the pitcher is charged +0.337 runs for each (83 in 2026), although the catcher committed it. A batted ball here is everything else: balls in play, home runs, reached on error and bunts.
Expected value of contact
Every batted ball's expected value comes from the season's league average of balls hit with the same exit velocity and launch angle. The value averaged is the LW1 value of the result (context-neutral, reached-on-error included as a result), not the ball's RE24, and the cells are fitted in sample on the season's own batted balls, each ball inside its own cell. The cell rule (prevention_decomp.expected_bip):
keys, tried in order: fine = (bunt flag, floor(EV / 3), floor(LA / 4)) 3 mph x 4 degrees
coarse = (bunt flag, floor(EV / 6), floor(LA / 8)) 6 mph x 8 degrees
trajectory = (recorded trajectory, bunt flag)
all batted balls
a key is used if it holds at least 25 balls and at least one that stayed in the park
no EV or no LA: start at the trajectory key
bunt flag = the ball's LW1 class is a bunt out
The replica of this rule in the checks script reproduces the model's own expected values exactly in every season. How the batted balls were valued, season by season:
| Season | Batted balls | No EV or LA | Fine cell | Coarse cell | Trajectory | All batted balls | Bunt hits and ROE |
|---|---|---|---|---|---|---|---|
| 2015 | 130,475 | 13.11% | 83.8% | 2.3% | 13.9% | 0.03% | 542 |
| 2016 | 128,821 | 12.17% | 84.1% | 2.8% | 13.0% | 0.02% | 498 |
| 2017 | 127,554 | 10.63% | 85.9% | 2.7% | 11.4% | 0.02% | 429 |
| 2018 | 126,284 | 8.77% | 87.6% | 2.6% | 9.7% | 0.02% | 459 |
| 2019 | 125,755 | 8.82% | 87.6% | 2.5% | 9.8% | 0.02% | 393 |
| 2020 | 43,972 | 0.70% | 87.8% | 8.0% | 4.2% | 0.02% | 140 |
| 2021 | 121,706 | 0.36% | 94.8% | 3.2% | 1.9% | 0.02% | 320 |
| 2022 | 124,267 | 0.48% | 95.5% | 2.7% | 1.8% | 0.01% | 384 |
| 2023 | 124,234 | 0.30% | 95.5% | 2.9% | 1.6% | 0.02% | 352 |
| 2024 | 124,203 | 0.33% | 95.3% | 2.9% | 1.8% | 0.01% | 377 |
| 2025 | 124,888 | 0.36% | 95.3% | 2.9% | 1.9% | 0.01% | 350 |
| 2026 | 124,611 | 0.45% | 94.8% | 3.2% | 1.9% | 0.02% | 507 |
Shares of the season's batted balls by the cell level that valued them. Bunt hits and ROE: bunt-trajectory batted balls that were not outs.
The paper first said that about 0.4% of batted balls lack exit velocity or launch angle. That is true from 2021 (0.30%–0.48%), not before: 2015–2019, with the earlier tracking, 8.8%–13.1% of batted balls had no measurement and 9.7%–13.9% were valued at their trajectory's average. Unmeasured balls are not a random sample of contact: in 2015, 25.4% of the unmeasured batted balls were popups and 51.5% ground balls, against 3.7% and 44.3% of the measured ones. Contact quality in those seasons is rougher, and the fallback shares differ by park and pitcher.
Defect: bunts are kept apart from swings only when they are outs. The bunt flag is the LW1 class, and only outs have the bunt class; a bunt single or a bunt that reached on an error keeps its hit class. In 2026, 1,136 bunts were outs and 507 were hits or errors. The bunt-out cells hold only outs, so a bunt out's expected value is the average bunt out. The bunt hits fall into the swing cells for their speed and angle (13 in fine and 210 in coarse cells, expected value 0.049 on average in the coarse ones, against 0.475 for the result), or, when those cells are too sparse (265 of the 273 at the next level have exit velocity and launch angle), into a trajectory cell that holds only bunt hits and errors (273, expected value 0.474 against a result of 0.474), and 11 fall through to the all-batted-balls value (0.039 against 0.474). In every case the split is by result, so the pitcher's contact quality for bunts is partly the bunt's outcome. Flagging bunts by trajectory instead would move 198 absolute runs of 2026 contact quality between pitchers, 1.74 at most (largest: Jose Cabrera +1.74 in 129 BF; Will Warren +1.67 in 677 BF; Yoshinobu Yamamoto +1.49 in 708 BF), a spread of 0.71 runs per 600 among the 277 pitchers with 250+ batters faced. It is small for pitchers; it would matter more in a predictor, which is why the minor-league translation keys bunts by trajectory.
The residual must be split carefully. Splitting it by the actual result makes every in-park ball look like a fielding win and every home run a loss (in 2026, +110 runs of "conversion" and −111 of "home runs" for the average club, from +24 to +195 and −174 to −84; a trap found in the first version of prevention_decomp.py). Instead each cell gives a home-run probability p and the mean value m of the balls that stayed in:
xv = p x v(HR) + (1 - p) x m HR part = (1[HR] - p) x (v(HR) - m) averages zero within a cell in-park part = v1 - m (non-home-runs only) averages zero within a cell v1 - xv = HR part + in-park part exactly, for every ball
Because a ball is valued at the first key that qualifies, not every ball's cell is the same set of balls, so the expected values do not sum exactly to the actual LW1 values; the model therefore recenters contact quality (and the pitchers' running game and fielding) to zero per batter faced after the fact. Before recentering, league contact quality was +28 runs in 2026 (+0.09 per 600 batters faced) but −615 in 2015 (−2.01 per 600), when 13.9% of batted balls fell back to their trajectory.
Running game and offsets
Steals, caught stealing and pickoffs (pickoff errors included) and wild pitches, passed balls and balks while a pitcher is on the mound are valued by their change in run expectancy, against two separate league rates per batter faced:
running(pitcher) = - [ (steal-type RE change - league steal-type RE per BF x his BF)
+ (wild pitch, passed ball, balk RE change - league rate per BF x his BF) ]
- his share of his catchers' throwing + blocking (Section 7)
The pitcher on an event is the pitcher of the plate appearance during which it happened (for a play without a plate appearance, that play's pitcher). Defensive indifference, the feed's other advances, outs and errors between pitches, and balks the feed labels as forced are not charged to anyone (Section 6). The pitcher's own fielding is his DRS at pitcher less its catcher parts (rSZ, rCERA, rSB), centered per inning at pitcher (Section 7). Strikeouts and walks include his share of his catchers' framing, removed.
What pitchers are not charged for
Batted-ball conversion (fielders), home runs beyond what the contact predicted, timing and the on-the-play layer are left out of pitcher value. In the Brewers decomposition these layers carried −29 (home runs beyond expected), −10 (timing), +32 (conversion) and +7.5 (on the play) runs in 2026, +1.0 together: the model does not charge the staff for the home runs or the timing, and does not credit it with the fielding. Calling the home-run layer luck was too strong: across clubs it repeats (0.38 [0.03, 0.65] from 2025 to 2026), largely because a club's home park repeats. What justifies leaving these layers out of a pitcher's value is how little they repeat for the pitcher himself, per batter faced, 250+ batters faced in both seasons:
| Pitcher layer (per BF) | r, 2025->2026 | Range over 9 pairs | Pooled [95%] | Charged to the pitcher? |
|---|---|---|---|---|
| Strikeouts and walks | 0.60 [0.49, 0.68] | 0.50–0.67 | 0.62 [0.58, 0.66] | Yes |
| Contact quality | 0.33 [0.18, 0.45] | 0.17–0.37 | 0.29 [0.23, 0.36] | Yes |
| Running game (after catcher offsets) | 0.30 [0.15, 0.43] | 0.18–0.46 | 0.32 [0.25, 0.39] | Yes |
| Home runs beyond expected | 0.12 [−0.03, 0.27] | 0.03–0.24 | 0.13 [0.08, 0.18] | No |
| Conversion (in-park part) | 0.14 [−0.01, 0.28] | −0.01 to 0.15 | 0.08 [0.02, 0.13] | No (fielders) |
| Timing | 0.08 [−0.07, 0.23] | −0.13 to 0.15 | 0.05 [0.00, 0.10] | No |
| On the play | 0.14 [−0.01, 0.29] | −0.15 to 0.20 | 0.06 [0.01, 0.12] | No |
Pooled over the nine pairs of consecutive seasons without 2020; interval from 1,000 bootstrap draws resampling pitchers.
Strikeouts and walks repeat far more than anything else; contact quality repeats much less but more than the layers left out, which is the line the model draws.
Pitch quality: what the stuff is worth
Pitch quality (PQ) is the stuff model of Section 12 turned into a measure of its own. Every pitch is valued from its velocity, movement, spin, extension and release, by a model that never sees where the pitch went or what happened to it. The values are summed per pitcher into the same two parts the model charges him for, strikeouts and walks and contact quality (above), and put per 600 batters faced against that season's average pitcher. Each season is centered by batters faced over the model's pitchers, so 0 is an average arm that year, as it is for the runs. The gap between a pitcher's pitch quality and his actual runs in those two parts is everything the stuff does not explain: command, sequencing, framing and luck.
The dashboard shows it wherever pitchers are: a "Pitch quality" section on each pitcher's season card (his stuff against his actual runs, then the same broken down by pitch type: usage, velocity, spin, movement, each pitch's value per 100 pitches against the season's average pitch, and its share of his pitch quality, the shares adding up to his total; pitch_types.py), two columns in each card's season-by-season table, two sortable columns in the Pitchers view (in runs or per 600, for both hands or either side), rows in the cards' By batter hand tables, the Teams view's staff columns (both hands and the staff splits), Top 100 rookies, a Best stuff tab (the top 50 pitchers of each season and of 2027 by pitch quality, in either part or both, with each one's best pitch, fastball velocity and the gap to his actual or projected runs), and each team's popup (the top pitchers' pitch quality and a staff table ranked across the clubs). By batter hand, each side counts only the pitches thrown to it but keeps the same baseline as the total (the season's average pitcher overall), so the platoon edge shows and the two sides, weighted by batters faced, give the total back; a right-handed sinkerballer like Clay Holmes in 2023 rates −5.2 in strikeouts and walks against left-handed batters and +12.7 against right-handed ones. For 2027 the hand split uses MLB pitches only, because Triple-A pitch quality is not split by hand. For 2027 the columns show each pitcher's MLB pitch quality over the three seasons before, weighted 3/2/1 by batters faced like the forecast's history (863 of 975 projected pitchers). That is not the forecast's own pull (beta-scaled and shrunk, Section 12). A pitcher with no MLB pitches shows his Triple-A pitch quality instead, weighted the same way by tracked batters faced, against each season's average MLB pitcher and marked AAA (78 of the 112 pitchers coming up from the minors, 10+ tracked batters faced); only the 34 with Double-A time alone, which is untracked, show a dash. The hand split uses MLB pitches only. Seasons run from 2015 to 2026; the ones before 2020 were measured by older systems (below).
How it tracks results, for pitchers with 250+ batters faced (100+ in 2020). The first two columns compare stuff and results in the same season. The last two ask what predicts next season's runs: this season's stuff, or this season's results.
| Season | Tracking | Pitchers | Same season: K&BB r | Same season: contact r | Next season K&BB: stuff / results r | Next season contact: stuff / results r |
|---|---|---|---|---|---|---|
| 2015 | PITCHf/x | 277 | 0.40 | 0.27 | 0.49 / 0.65 | 0.32 / 0.26 |
| 2016 | PITCHf/x | 280 | 0.51 | 0.28 | 0.39 / 0.67 | 0.22 / 0.31 |
| 2017 | TrackMan | 290 | 0.52 | 0.40 | 0.45 / 0.60 | 0.19 / 0.17 |
| 2018 | TrackMan | 274 | 0.52 | 0.39 | 0.56 / 0.66 | 0.21 / 0.24 |
| 2019 | TrackMan | 279 | 0.59 | 0.35 | 0.34 / 0.51 | 0.32 / 0.21 |
| 2020 | Hawk-Eye | 244 | 0.45 | 0.31 | 0.56 / 0.53 | 0.44 / 0.22 |
| 2021 | Hawk-Eye | 273 | 0.55 | 0.48 | 0.46 / 0.63 | 0.44 / 0.37 |
| 2022 | Hawk-Eye | 274 | 0.48 | 0.46 | 0.51 / 0.50 | 0.52 / 0.35 |
| 2023 | Hawk-Eye | 280 | 0.51 | 0.50 | 0.39 / 0.57 | 0.38 / 0.35 |
| 2024 | Hawk-Eye | 277 | 0.49 | 0.35 | 0.43 / 0.66 | 0.43 / 0.26 |
| 2025 | Hawk-Eye | 289 | 0.46 | 0.38 | 0.52 / 0.60 | 0.46 / 0.33 |
| 2026 | Hawk-Eye | 277 | 0.55 | 0.42 |
Correlations of pitch quality and actual runs per 600 batters faced. Next season: pitchers who reached the minimum in both seasons.
Stuff and results line up at about r = 0.5 for strikeouts and walks in the same season and lower for contact. Strikeout and walk results repeat better than stuff predicts them, so for that part the results lead and pitch quality only adds (Section 12). Contact is the other way round from 2019 on: next season's contact is predicted better by this season's stuff than by this season's contact in 7 of 7 pairs. Before that, with the older systems, it is about even or behind (2 of 4), so read those seasons' contact numbers as rougher.
The best stuff of 2015–2026, pitcher-seasons with 500+ batters faced, per 600, next to the pitcher's actual runs in the same part:
| Season | Pitcher | Team | BF | Pitch quality | Actual runs |
|---|---|---|---|---|---|
| 2022 | Spencer Strider | ATL | 528 | +26.8 | +28.7 |
| 2019 | Gerrit Cole | HOU | 817 | +26.2 | +39.7 |
| 2022 | Gerrit Cole | NYY | 793 | +22.6 | +20.0 |
| 2018 | Gerrit Cole | HOU | 799 | +21.7 | +22.8 |
| 2023 | Spencer Strider | ATL | 763 | +21.0 | +28.1 |
| 2021 | Gerrit Cole | NYY | 726 | +20.3 | +26.8 |
| 2017 | Marco Estrada | TOR | 806 | +20.3 | +0.3 |
| 2022 | Hunter Greene | CIN | 531 | +19.0 | +13.2 |
Strikeouts and walks.
| Season | Pitcher | Team | BF | Pitch quality | Actual runs |
|---|---|---|---|---|---|
| 2022 | Corbin Burnes | MIL | 797 | +12.9 | +2.1 |
| 2021 | Corbin Burnes | MIL | 657 | +11.7 | +18.9 |
| 2017 | Clayton Kershaw | LAD | 679 | +11.4 | +7.7 |
| 2016 | Clayton Kershaw | LAD | 544 | +11.2 | +15.6 |
| 2022 | Sandy Alcantara | MIA | 886 | +11.1 | +14.2 |
| 2021 | Sandy Alcantara | MIA | 837 | +10.9 | +8.4 |
| 2017 | Charlie Morton | HOU | 617 | +10.6 | +5.1 |
| 2023 | Graham Ashcraft | CIN | 624 | +10.4 | +2.6 |
Contact quality.
Spencer Strider and Gerrit Cole own the strikeout list, and their results matched the stuff. Marco Estrada is the model's clearest miss: it rated his 2017 stuff +20.3 against +0.3 in results, and +9 to +16 in his other seasons with 250+ batters faced, where the results were ordinary too.
Pitch quality before 2021: three tracking systems on one scale
Scope. The dashboard shows pitch quality for 2015–2020 too (cards, tables, leaderboards, team popups); the forecast and its back-tests never read those seasons (pitch_quality_early.py; the files sit in data/pitches/early/ precisely so that pitch_quality.py, which lists data/pitches/, cannot train or score the forecast on them). The stuff model is fitted only on Hawk-Eye pitches (2021–2026), so every earlier pitch has to be expressed in Hawk-Eye's measurements before the model can score it. The Stats API feed carries three systems:
- PITCHf/x, 2015, 2016 (camera tracking): start speed 50 feet from the plate, the 40-foot movement pfx (inches of break relative to a spinless trajectory over the last 40 feet), release coordinates at y = 50 ft, and a spin rate estimated from the trajectory alone (Magnus spin only, no gyro component: a 90 mph fastball shows 1,276 rpm, sliders near 1,000). No extension, no induced vertical break, no horizontal break as Statcast defines them.
- TrackMan, 2017, 2018, 2019 (Doppler radar): release speed, induced vertical and horizontal break, total spin, extension and release point, as Hawk-Eye reports them, but with its own calibration, and recalibrated between 2017 and 2018 (below).
- Hawk-Eye, from 2020 (optical): the model's own system. 2020 is scored as is apart from the park step.
Four steps, in this order, then the scoring: (1) for PITCHf/x, break from pfx; (2) each park's calibration evened out to its Hawk-Eye level; (3) the system bridges, chained toward Hawk-Eye; (4) each non-fastball's gaps from the pitcher's main fastball recomputed on the corrected values. TrackMan seasons are then scored by the full stuff model; PITCHf/x seasons by a reduced model (same design without spin and extension, fitted on the same Hawk-Eye pitches). On 2021, a season neither model saw in fitting, the reduced model ranks pitchers like the full one, r = 0.971 for strikeouts and walks and 0.968 for contact (440 pitchers with 150+ batters faced); and scored both ways, 2023's Hawk-Eye pitches get the same mean strikeouts-and-walks value in every pitcher-hand × batter-side × pitch-group cell within 0.045 runs per 100 pitches, so the reduced model is not where the older seasons differ.
Frames. Horizontal break and release side are mirrored for left-handers (arm side positive) everywhere in the model, because the pitch-value regressions pool both hands. A calibration error, though, lives in the catcher's frame: a lateral bias b in a system's raw horizontal measurement becomes +b for right-handers and −b for left-handers once mirrored. So for the two mirrored measurements every offset below is estimated and applied separately for each pitcher hand; for the others (velocity, vertical break, spin, extension, release height) the two hands share one offset. This was not the first version (see the history below).
Step 1, movement map. Fitted on the 2,428,164 2017–2020 pitches that carry both pfx and Statcast's breaks, least squares, mirrored frame:
ivb = 0.7283 + 1.5687 pfx_z - 0.0103 pfx_z (velo - 90) R2 0.905, RMSE 2.77 in hb = -0.5149 - 1.6188 pfx_x + 0.0064 pfx_x (velo - 90) + 0.0503 h R2 0.919, RMSE 2.69 in h = +1 right-handed, -1 left-handed (the catcher-frame intercept, which mirroring flips)
The slopes near 1.6 are the extra flight time: Statcast's breaks run from release (about 54 feet) where pfx covers 40 feet, and (54/40)² = 1.82 before air resistance; the velocity term carries the small dependence of that ratio on speed. PITCHf/x spin is then set to missing (not mapped: the trajectory-only estimate is a different quantity, not a biased version of total spin), as is extension.
Step 2, parks. For each season and measurement x, with cells k = pitch group × pitcher hand for horizontal break and release side and pitch group otherwise, alternate three times:
mu(p,g) = mean over p's group-g pitches of ( x - o(v,k) ) pitcher-group means o(v,k) = mean over the park's cell-k pitches of ( x - mu(p,g) ) park offsets (500+ pitches) x' = x - ( o_season(v,k) - o_HawkEye(v,k) ) o_HawkEye = mean of the 2021 and 2022 offsets
The alternation matters because a home staff throws about half its pitches in its own park: one pass would push half of a park's offset into its home pitchers' means. Subtracting the park's own Hawk-Eye offset instead of the whole offset keeps real park effects (thin air at Coors Field reduces break there in every system) and removes only the calibration gap; a park with no Hawk-Eye seasons (the Rangers' old park, Turner Field) loses its whole offset. Spread over parks of the fastball offsets, before correction:
| Season | mph | IVB | HB | Spin | Extension | Release height | Release side |
|---|---|---|---|---|---|---|---|
| Hawk-Eye reference (2021–2022) | 0.15 | 0.6 | 0.5 | 10 | 0.05 | 0.05 | 0.04 |
| 2015 (PITCHf/x) | 0.34 | 1.1 | 1.6 | – | – | 0.08 | 0.08 |
| 2016 (PITCHf/x) | 0.33 | 1.1 | 1.7 | – | – | 0.08 | 0.11 |
| 2017 (TrackMan) | 0.17 | 1.7 | 2.2 | 10 | 0.11 | 0.07 | 0.07 |
| 2018 (TrackMan) | 0.11 | 0.6 | 0.5 | 13 | 0.05 | 0.07 | 0.04 |
| 2019 (TrackMan) | 0.11 | 0.6 | 0.5 | 11 | 0.06 | 0.05 | 0.04 |
| 2020 (Hawk-Eye) | 0.18 | 0.6 | 0.5 | 11 | 0.08 | 0.04 | 0.04 |
Standard deviation over parks of the park offset o(v,k) for fastballs, in mph, inches, rpm and feet. For horizontal break and release side both hands' cells are pooled into one spread, so these two columns also contain each system's catcher-frame lateral bias, which differs in sign between the hands.
Step 3, bridges. For a measurement and cell, the same-pitcher change between seasons a and b is a weighted mean over pitcher-groups with 30+ pitches in both:
D_k(a->b) = sum_p w_p ( xbar_p,b - xbar_p,a ) / sum_p w_p , w_p = min( n_p,a , n_p,b ) bridge_k = D_k(across the boundary) - mean of D_k over the clean baseline pairs that measure x x'' = x' + sum of the bridges between the pitch's season and Hawk-Eye
Subtracting the baseline removes what a same-pitcher change contains apart from measurement: aging (velocity falls a few tenths of a mph a year for returning pitchers), league trends and survivor selection. The baselines are chosen to be clean (no boundary inside them): PITCHf/x → TrackMan 2017 uses 2015→16 and 2018→19; TrackMan 2017 → 2018–19 uses the same two; TrackMan → Hawk-Eye uses 2018→19 and 2021→22, skipping 2020→21 because the mid-2021 enforcement against grip substances changed spin and movement for real. A measurement a baseline system lacks (PITCHf/x spin and extension) is averaged over the baselines that have it, not counted as a zero change. Applied: 2018–19 take the TrackMan → Hawk-Eye bridge; 2017 that plus its own; 2015–16 all three, on the measurements the reduced model uses. Fastball bridges:
| Bridge | mph | IVB | HB, right-handed | HB, left-handed | Spin | Extension |
|---|---|---|---|---|---|---|
| PITCHf/x → TrackMan 2017 (2016→17) | +0.34 | +1.90 | +2.59 | −1.52 | – | – |
| TrackMan 2017 → TrackMan 2018–19 (2017→18) | −0.29 | −1.22 | −0.74 | +0.13 | −21 | −0.07 |
| TrackMan → Hawk-Eye (2019→20) | −0.22 | +0.31 | −0.11 | −0.09 | −11 | +0.20 |
Bridge offsets for fastballs (four-seam, sinker), added to the earlier system's values; inches, mph, rpm, feet. Spin and extension do not exist in the PITCHf/x feed.
History: two flaws found by the platoon analysis. The first version (2026-11-16, morning) pooled both hands in every offset and treated 2017–2019 as one TrackMan system, with 2017→18 as a baseline pair. Its season-to-season check was weak across 2016→17 and 2017→18 (table below), which was then read as pitcher-level noise in 2017 because park × month and per-game offsets did not help. That reading was wrong. Pitch quality by batter hand (below) showed left-handers' 2015–16 stuff to be worse against left-handed batters than against right-handed ones, by 3.6 and 4.4 runs per 600, where their results said the opposite. Two diagnostics then located the causes:
- Catcher frame. Across 2016→17 the same right-handers' fastball horizontal break moved +2.61 inches and the same left-handers' −1.40 (mirrored frame). Writing the two as s + b and s − b gives an arm-side (frame-free) change s = +0.61 and a catcher-frame bias b = 2.00 inches; the pooled bridge of version 1 (+1.83 inches, the boundary change less the baseline changes, set mostly by the more numerous right-handers) left right-handers 0.8 inches off and left-handers 3.2. Release side showed no such split (all changes within ±0.07 ft), and vertical break moved alike for both hands (+1.90 and +1.79), as a frame artifact predicts.
- TrackMan's 2017. From 2017 to 2018 the same pitchers lost 1.2 to 2.0 inches of induced vertical break in every pitch group and for both hands, and horizontal break moved −0.78 for right-handers and +0.30 for left-handers, where a normal year moves these by about 0.1 to 0.3. A change that size, that uniform across pitch types, and with a catcher-frame signature is a recalibration, not pitching. Using 2017→18 as a "normal" baseline had therefore also carried that recalibration into the PITCHf/x bridge.
Version 2 estimated the mirrored measurements per hand (offsets, park cells, and the h term in the movement map); version 3 (published) also gave 2017 its own bridge, with clean baselines. A third diagnostic, the platoon edge by pitch group, showed what version 2 still missed: right-handers' fastballs to left-handed batters were valued +0.301 and +0.211 runs per 100 pitches in 2016–17 (mean predicted strikeouts-and-walks value, uncentered, pitcher's view), above the same pitchers' fastballs to right-handed batters (+0.180 and +0.184), a reversal of the platoon direction that 2018 does not show (−0.095 to left-handed batters); with version 3 they are −0.154 and −0.088 (to right-handed batters −0.078 and +0.020). The checks, all three versions:
| Season pair | Version 1: pooled offsets, 2017–19 one system | Version 2: mirrored measurements per pitcher hand | Version 3 (published): + 2017 its own bridge | Pitchers |
|---|---|---|---|---|
| 2015–2016 | 0.899 / 0.863 | 0.910 / 0.884 | 0.893 / 0.885 | 290 |
| 2016–2017 (PITCHf/x → TrackMan 2017) | 0.790 / 0.607 | 0.862 / 0.791 | 0.848 / 0.798 | 300 |
| 2017–2018 (TrackMan 2017 → TrackMan) | 0.851 / 0.754 | 0.866 / 0.800 | 0.885 / 0.828 | 305 |
| 2018–2019 | 0.882 / 0.874 | 0.882 / 0.873 | 0.882 / 0.873 | 302 |
| 2019–2020 (TrackMan → Hawk-Eye) | 0.887 / 0.812 | 0.886 / 0.813 | 0.886 / 0.813 | 110 |
| 2020–2021 | 0.905 / 0.815 | 0.904 / 0.815 | 0.904 / 0.815 | 115 |
| 2021–2022 | 0.912 / 0.855 | 0.912 / 0.855 | 0.912 / 0.855 | 293 |
Correlation of pitchers' pitch quality per batter faced in consecutive seasons, strikeouts and walks / contact, pitchers with 150+ batters faced in both. 2021–2026 pairs are the cards' leave-one-season-out scores and do not depend on the calibration; 2019–20 and 2020–21 involve 2020's 60 games (fewer pitchers reach the minimum, and lower contact correlations).
| Season, or staff check | Version 1: pooled offsets, 2017–19 one system | Version 2: mirrored measurements per pitcher hand | Version 3 (published): + 2017 its own bridge |
|---|---|---|---|
| 2015 | +1.40 / −3.55 | −0.49 / +2.71 | +2.52 / +3.01 |
| 2016 | +1.32 / −4.42 | −0.51 / +1.98 | +2.46 / +2.40 |
| 2017 | +0.81 / +2.44 | +0.89 / +2.87 | +2.88 / +3.23 |
| 2018 | +3.35 / +3.09 | +3.35 / +3.02 | +3.35 / +3.02 |
| 2019 | +3.20 / +3.22 | +3.21 / +3.11 | +3.21 / +3.11 |
| Staff r, PITCHf/x club-seasons | 0.20 | 0.25 | 0.38 |
| Staff r, TrackMan club-seasons | 0.60 | 0.60 | 0.62 |
Upper rows: the league platoon edge in strikeouts-and-walks pitch quality, right-handers / left-handers, runs per 600 batters faced (definition below); the actual edge in those seasons is +6.2 to +7.3 for right-handers and +2.7 to +4.9 for left-handers. Lower rows: correlation of staff pitch quality with staff results over each system's club-seasons (below).
The two pairs that cross a system change gain most: 2016–17 from 0.790 / 0.607 to 0.848 / 0.798, and 2017–18 from 0.851 / 0.754 to 0.885 / 0.828. 2015–16, inside PITCHf/x, moves little (0.899 / 0.863 to 0.893 / 0.885): both seasons take the same bridges, and the strikeouts-and-walks figure can move at all only because the stuff model is quadratic, so the absolute level of a measurement changes how pitchers rank. Pairs from 2018 on are unchanged, as they must be. The 2021–2026 pairs run 0.886–0.912 / 0.855–0.869; the pairs that cross 2017 are still a little below that, so some 2017 noise remains, but most of what the first version blamed on it was the two calibration flaws.
Pitch quality by batter hand: the platoon edge
Definitions. Each pitch carries the side the batter stood on as he batted (switch hitters on the side they took). For pitcher i, component c and batter side s:
PQ_c(i,s) = ( sum over i's pitches to batters of side s of v_c / BF_i,s - m_c ) x 600 A_c(i,s) = actual runs_c(i,s) / BF_i,s x 600 (the model's runs against that side) edge_c(i) = X_c(i, side of his arm) - X_c(i, other side) X = PQ or A league edge_c = sum_i BF_i edge_c(i) / sum_i BF_i (pitchers with both sides faced)
The baseline mc is the same as for the total (the season's league predicted value per batter faced, all batters), not a per-side mean, for two reasons. First, additivity: BFi,L·PQ(i,L) + BFi,R·PQ(i,R) = BFi·PQ(i), so the sides give the total back (checked: within 0.002 runs a pitcher-season). Second, comparability: the model's own splits of actual runs (Section 9) are on the overall baseline too; league totals sum to zero overall, not per side. In 2023, for instance, right-handers' strikeouts-and-walks runs were +2.31 per 600 against right-handed batters and −2.89 against left-handed ones. Per-side centering would subtract the league platoon edge from both series and leave nothing to compare. The hand sums are taken with the same models and scoring as the cards, so a pitcher's two sides are the card's total split, not a separate estimate.
One modeling detail matters for reading this. The pitch-value regressions (Section 12) are fitted separately for same-hand and opposite-hand matchups (the model key is group × same/opposite × two strikes or not) and include count indicators (balls 1–3, one strike), so a pitch's predicted value depends on the matchup and on the count it was thrown in. A pitcher's edge in stuff therefore has three sources: how the model values his pitches' physical traits against each side, the pitch mix he chooses by side, and the counts he throws them in. Location is not one of them (the stuff model has no location).
The league edge, by season and hand (runs per 600 batters faced):
| Season | Tracking | K&BB stuff, RHP | K&BB stuff, LHP | K&BB stuff | K&BB actual | Contact stuff, RHP | Contact stuff, LHP | Contact stuff | Contact actual |
|---|---|---|---|---|---|---|---|---|---|
| 2015 | PITCHf/x | +2.52 | +3.01 | +2.65 | +6.62 | +3.95 | +3.12 | +3.73 | +3.84 |
| 2016 | PITCHf/x | +2.46 | +2.40 | +2.44 | +5.26 | +4.01 | +3.36 | +3.84 | +4.00 |
| 2017 | TrackMan | +2.88 | +3.23 | +2.97 | +6.71 | +3.69 | +3.13 | +3.55 | +4.93 |
| 2018 | TrackMan | +3.35 | +3.02 | +3.26 | +6.11 | +3.96 | +3.71 | +3.88 | +4.48 |
| 2019 | TrackMan | +3.21 | +3.11 | +3.19 | +6.06 | +3.92 | +2.74 | +3.60 | +4.32 |
| 2020 | Hawk-Eye | +3.89 | +3.29 | +3.73 | +6.79 | +3.76 | +3.09 | +3.58 | −0.40 |
| 2021 | Hawk-Eye | +3.51 | +2.31 | +3.15 | +6.34 | +2.89 | +3.37 | +3.03 | +5.82 |
| 2022 | Hawk-Eye | +4.47 | +2.58 | +3.96 | +4.30 | +3.77 | +4.69 | +4.02 | +4.01 |
| 2023 | Hawk-Eye | +4.34 | +3.07 | +4.00 | +4.24 | +4.73 | +5.70 | +4.99 | +5.13 |
| 2024 | Hawk-Eye | +3.68 | +2.64 | +3.40 | +5.37 | +5.22 | +6.12 | +5.46 | +4.92 |
| 2025 | Hawk-Eye | +3.75 | +2.61 | +3.44 | +5.37 | +4.96 | +6.30 | +5.32 | +5.39 |
| 2026 | Hawk-Eye | +3.38 | +3.32 | +3.36 | +4.96 | +4.61 | +6.03 | +5.03 | +6.24 |
BF-weighted mean over pitchers who faced both sides of (same-side minus other-side), stuff and actual. All seasons on the published calibration (version 3, above).
Over the seasons other than 2020, the edge in stuff averages +3.26 runs per 600 for strikeouts and walks against an actual +5.58 (58%), and +4.22 for contact against +4.83 (88%). So the physical traits and mix of a pitcher's pitches account for most of the contact platoon edge but only about three-fifths of the strikeouts-and-walks edge; the rest is what the stuff model cannot see, chiefly location (a pitcher works the same-side batter's edges differently) and sequencing. The strikeouts-and-walks stuff edge is steady across systems once the calibration is right (2.4 to 4.0 overall in every season; 2.3 to 4.5 by hand). The contact stuff edge rises over the Hawk-Eye years (+3.0 in 2021, +5.0 to +5.5 in 2024–26) where the actual contact edge has no such trend; what drives that rise is not tested here. The two hands differ in contact: in Hawk-Eye seasons left-handers' actual contact edge averages +9.1 against +3.8 for right-handers (left-handed batters' weak contact off left-handers), while their stuff edge is +5.4 against +4.4: the stuff gets the direction right but explains less of the left-handers' contact advantage. 2020's actual contact edge (−0.40) is a 60-game result (about 66,000 batters faced) and is left out of the averages above.
Pitchers. Pitcher-level agreement between his stuff edge and his actual edge in the same season (both sides with 100+ batters faced, 40+ in 2020):
| Season | Tracking | Pitchers | K&BB r | Contact r |
|---|---|---|---|---|
| 2015 | PITCHf/x | 276 | 0.18 | 0.31 |
| 2016 | PITCHf/x | 270 | 0.25 | 0.32 |
| 2017 | TrackMan | 290 | 0.22 | 0.38 |
| 2018 | TrackMan | 277 | 0.33 | 0.37 |
| 2019 | TrackMan | 283 | 0.35 | 0.29 |
| 2020 | Hawk-Eye | 253 | 0.01 | 0.16 |
| 2021 | Hawk-Eye | 267 | 0.32 | 0.37 |
| 2022 | Hawk-Eye | 253 | 0.27 | 0.38 |
| 2023 | Hawk-Eye | 273 | 0.32 | 0.32 |
| 2024 | Hawk-Eye | 298 | 0.28 | 0.31 |
| 2025 | Hawk-Eye | 297 | 0.25 | 0.30 |
| 2026 | Hawk-Eye | 307 | 0.31 | 0.34 |
Correlation over pitchers of edge_c(i) in stuff and in actual runs, same season.
The correlations are low (0.18–0.38 outside 2020, where the 60-game samples leave them near zero) because a pitcher's actual split rests on two small samples: a starter faces 250–450 batters from the less common side. The persistence of each edge from one season to the next (same thresholds in both seasons) shows which of the two is the signal:
| Seasons | Pitchers | K&BB: stuff repeats | K&BB: actual repeats | K&BB: stuff → next actual | K&BB: actual → next actual | Contact: stuff repeats | Contact: actual repeats | Contact: stuff → next actual | Contact: actual → next actual |
|---|---|---|---|---|---|---|---|---|---|
| 2015–2016 | 173 | 0.79 | 0.39 | 0.14 | 0.39 | 0.90 | 0.24 | 0.34 | 0.24 |
| 2016–2017 | 173 | 0.66 | 0.30 | 0.24 | 0.30 | 0.83 | 0.07 | 0.34 | 0.07 |
| 2017–2018 | 165 | 0.85 | 0.35 | 0.39 | 0.35 | 0.88 | 0.19 | 0.36 | 0.19 |
| 2018–2019 | 168 | 0.75 | 0.46 | 0.28 | 0.46 | 0.87 | 0.25 | 0.21 | 0.25 |
| 2019–2020 | 138 | 0.81 | 0.07 | −0.03 | 0.07 | 0.86 | 0.06 | 0.01 | 0.06 |
| 2020–2021 | 156 | 0.80 | 0.11 | 0.37 | 0.11 | 0.85 | 0.12 | 0.32 | 0.12 |
| 2021–2022 | 161 | 0.82 | 0.26 | 0.29 | 0.26 | 0.84 | 0.32 | 0.34 | 0.32 |
| 2022–2023 | 150 | 0.75 | 0.30 | 0.32 | 0.30 | 0.86 | 0.18 | 0.20 | 0.18 |
| 2023–2024 | 160 | 0.77 | 0.29 | 0.28 | 0.29 | 0.86 | 0.19 | 0.38 | 0.19 |
| 2024–2025 | 175 | 0.81 | 0.05 | 0.26 | 0.05 | 0.82 | 0.12 | 0.27 | 0.12 |
| 2025–2026 | 178 | 0.84 | 0.33 | 0.33 | 0.33 | 0.78 | 0.08 | 0.34 | 0.08 |
Correlations over pitchers in both seasons. “Repeats”: an edge with the same edge next season. “→ next actual”: this season's edge with next season's actual edge.
A pitcher's stuff edge repeats at about 0.79 (strikeouts and walks) and 0.85 (contact) on average over the pairs; his actual edge at 0.26 and 0.16. As a forecast of next season's actual split, this season's stuff edge does better than this season's actual edge in 6 of 11 pairs for strikeouts and walks and 9 of 11 for contact (mean r 0.26 against 0.26, and 0.28 against 0.16). Observed platoon splits are mostly sampling noise at these sample sizes, which is why the forecast's own splits (Section 12) regress a pitcher's gap toward the typical gap for his hand with kgap = 900 on the harmonic mean of his weighted batters faced from the two sides: a starter with 300 and 400 keeps n/(n + 900) = 171/1,071 = 16% of his observed gap. A pitcher's stuff edge is a far more stable estimate of the true gap and is a natural prior for that regression, in the place of the typical gap for his hand; that is a candidate for the forecast, not tested or used yet.
The pitch mix. Pooling the Hawk-Eye seasons other than 2020 (pitcher-seasons with 250+ batters faced), each edge is regressed on the share of his pitches that are offspeed (changeup, splitter, forkball, screwball) and on the share that are sliders or sweepers, per 10 percentage points:
| Pitchers and share | Pitcher-seasons | Stuff K&BB edge | Stuff contact edge | Stuff edge, both | Actual edge, both |
|---|---|---|---|---|---|
| Right-handers: offspeed share | 1,218 | +1.39 (r 0.25) | −1.45 (r −0.22) | −0.05 (r −0.01) | −2.28 (r −0.11) |
| Right-handers: slider and sweeper share | 1,218 | −0.38 (r −0.09) | +0.87 (r 0.17) | +0.49 (r 0.08) | +1.59 (r 0.10) |
| Left-handers: offspeed share | 452 | +1.38 (r 0.25) | −1.38 (r −0.18) | 0.00 (r 0.00) | −0.73 (r −0.03) |
| Left-handers: slider and sweeper share | 452 | −0.19 (r −0.05) | +0.58 (r 0.11) | +0.39 (r 0.06) | +1.48 (r 0.09) |
OLS slope per 10 percentage points of the share, runs per 600 batters faced, with the correlation. One regressor at a time; usage is the pitcher's overall share, not by side.
More offspeed moves a pitcher's stuff edge out of contact and into strikeouts and walks but leaves the total unchanged (slope about 0, for both hands); more sliders and sweepers do the reverse, slightly. In actual results, more offspeed shrinks a right-hander's total edge (the changeup's role against opposite-side batters), which the stuff edge does not reproduce: that part of the changeup's platoon value would sit in location and sequencing. The single regressions are descriptive (usage correlates with arm slot, velocity and role) and are not used anywhere in the model.
Staff pitch quality
Definition. A club's staff pitch quality is its pitchers' pitch quality, each over his batters faced for the club; the staff's actual runs are built the same way, so both are per 600 batters faced and comparable:
BF(i,club) = BF_i x share(i,club) share = the pitcher's fraction of his season with the club (the model's split) PQ_c(club) = sum_i BF(i,club) PQ_c(i) / sum_i BF(i,club) A_c(club) = sum_i BF(i,club) A_c(i) / sum_i BF(i,club)
A traded pitcher's whole-season pitch quality is spread over his clubs by batters faced (pitch quality is not split by club). The team popups and the Teams view show the same quantity (in 2027 from the weighted 2024–26 MLB pitch quality, with Triple-A for pitchers who have no MLB pitches then).
By season, correlation over the clubs of staff pitch quality and staff actual runs, per component and for both together, with a Fisher-z 95% interval for the latter (z = atanh r, standard error 1/√(n − 3)):
| Season | Tracking | Clubs | K&BB r | Contact r | Both r | 95% interval, both | SD staff PQ | SD staff actual |
|---|---|---|---|---|---|---|---|---|
| 2015 | PITCHf/x | 30 | 0.40 | 0.22 | 0.33 | [−0.03, 0.62] | 2.85 | 6.44 |
| 2016 | PITCHf/x | 30 | 0.33 | 0.35 | 0.47 | [0.13, 0.71] | 2.10 | 5.44 |
| 2017 | TrackMan | 30 | 0.36 | 0.34 | 0.43 | [0.09, 0.69] | 2.31 | 6.79 |
| 2018 | TrackMan | 30 | 0.55 | 0.46 | 0.57 | [0.26, 0.77] | 3.24 | 6.93 |
| 2019 | TrackMan | 30 | 0.68 | 0.66 | 0.77 | [0.56, 0.88] | 3.45 | 8.02 |
| 2020 | Hawk-Eye | 30 | 0.61 | 0.59 | 0.66 | [0.39, 0.82] | 3.77 | 8.80 |
| 2021 | Hawk-Eye | 30 | 0.77 | 0.62 | 0.77 | [0.56, 0.88] | 3.66 | 7.40 |
| 2022 | Hawk-Eye | 30 | 0.60 | 0.50 | 0.65 | [0.38, 0.82] | 3.81 | 6.91 |
| 2023 | Hawk-Eye | 30 | 0.59 | 0.64 | 0.68 | [0.42, 0.83] | 3.61 | 6.17 |
| 2024 | Hawk-Eye | 30 | 0.38 | 0.23 | 0.42 | [0.07, 0.68] | 2.77 | 5.09 |
| 2025 | Hawk-Eye | 30 | 0.62 | 0.46 | 0.57 | [0.27, 0.77] | 2.45 | 5.55 |
| 2026 | Hawk-Eye | 30 | 0.54 | 0.49 | 0.68 | [0.42, 0.83] | 3.21 | 6.63 |
SD: standard deviation over the clubs of the two parts together, runs per 600 batters faced.
Thirty clubs make each season's correlation imprecise (interval half-widths of about 0.15 to 0.35, wider the lower r is), so the systems are compared on their pooled club-seasons:
| Club-seasons | n | K&BB r | Contact r | Both r | 95% interval, both |
|---|---|---|---|---|---|
| PITCHf/x | 60 | 0.37 | 0.29 | 0.38 | [0.14, 0.58] |
| TrackMan | 90 | 0.53 | 0.52 | 0.62 | [0.47, 0.73] |
| Hawk-Eye | 210 | 0.59 | 0.52 | 0.65 | [0.56, 0.72] |
| Hawk-Eye without 2020 | 180 | 0.59 | 0.50 | 0.65 | [0.55, 0.72] |
Each club-season once; pitch quality and runs are already centered within each season.
Against Hawk-Eye's r = 0.65, the pooled older seasons give r = 0.54 [0.42, 0.65]; the difference is z = 1.51, two-sided p = 0.13 (Fisher test for independent correlations). Before the calibration fixes above, the PITCHf/x club-seasons gave r = 0.20 and the pooled test p = 0.0067: most of the apparent finding that staff stuff "means less" under the older systems was the calibration. Separately: TrackMan against Hawk-Eye, z = 0.42, p = 0.68, no difference; PITCHf/x against Hawk-Eye, z = 2.46, p = 0.014, still lower. The reduced model is not the reason (it matches the full model on Hawk-Eye pitches, above), so the PITCHf/x measurements themselves are: break from a 40-foot fit with an RMSE of 2.7 inches, speed and release measured at 50 feet, and no usable spin.
Why staffs track better than pitchers (under Hawk-Eye, per component, 0.59 and 0.52 for staffs against 0.46–0.55 and 0.35–0.50 for pitchers with 250+ batters faced in the by-season table that opens this subsection's pitch-quality section): a correlation between a measure and an outcome is capped by the reliabilities of both, robserved = rtrue √(relx rely), and a staff's actual runs rest on about 6,000 batters faced, a pitcher's on a few hundred. Aggregation raises the reliability of the outcome much more than that of pitch quality, which is already stable for single pitchers (season-to-season r 0.86–0.91 under Hawk-Eye). The staff figure is also not a simple average of pitcher correlations: it weights pitchers by batters faced, so relievers' small samples count for little.
Persistence. Consecutive seasons, pooled (a pair belongs to a system when both seasons do):
| Pairs | Club pairs | Stuff repeats | Actual repeats | Stuff → next actual | Actual → next actual |
|---|---|---|---|---|---|
| PITCHf/x | 30 | 0.65 | 0.66 | 0.19 | 0.66 |
| across a system change | 60 | 0.58 | 0.55 | 0.46 | 0.55 |
| TrackMan | 60 | 0.72 | 0.69 | 0.51 | 0.69 |
| Hawk-Eye | 180 | 0.79 | 0.56 | 0.55 | 0.56 |
Correlations over clubs present in both seasons, both parts together. Rosters turn over, so these mix staff stability with the persistence of the measures.
Under Hawk-Eye a staff's pitch quality repeats at 0.79 where its results repeat at 0.56, and this season's staff pitch quality predicts next season's staff results about as well as this season's results do (0.55 against 0.56). The single PITCHf/x pair (2015→16, 30 clubs) is the exception, where stuff predicts next season's results poorly; one pair of 30 is too few to separate a system effect from chance.
Does the park recalibration remove staff signal? A home staff throws half its pitches in its park, so the park offset and the staff's true stuff are partly collinear, and the alternation in step 2 could in principle absorb real staff-level stuff into a park offset. Test: rescore 2015–2019 with step 2 left out (everything else, including the version-3 bridges, unchanged) and compare.
| Season | Staff r, with | Staff r, without | Pitcher r, with | Pitcher r, without | SD of the shift | Shift vs staff actual r |
|---|---|---|---|---|---|---|
| 2015 | 0.33 | 0.25 | 0.40 / 0.27 | 0.39 / 0.26 | 1.38 | 0.14 |
| 2016 | 0.47 | 0.42 | 0.51 / 0.28 | 0.53 / 0.28 | 1.17 | −0.04 |
| 2017 | 0.43 | 0.37 | 0.52 / 0.41 | 0.50 / 0.41 | 2.62 | −0.07 |
| 2018 | 0.57 | 0.59 | 0.52 / 0.39 | 0.52 / 0.40 | 0.45 | −0.09 |
| 2019 | 0.77 | 0.73 | 0.59 / 0.35 | 0.59 / 0.35 | 0.46 | 0.21 |
Staff r: both parts together. Pitcher r: strikeouts and walks / contact, pitchers with 250+ batters faced. Shift: how much the recalibration moved each staff's pitch quality (with minus without), its SD over the clubs and its correlation with the staff's actual runs.
Leaving the recalibration out lowers the staff correlation where the recalibration is large (2015–17) and changes almost nothing where it is small (2018–19). The shift it applies to a staff is uncorrelated with that staff's results (|r| ≤ 0.22 over 30 clubs, inside the 95% band for no correlation, ±1.96/√27 = ±0.38), which is what a calibration correction should look like and not what absorbing real stuff would look like (that would correlate negatively). So the recalibration removes calibration, not signal.
Most and least beyond their stuff. Leaving out 2020's 60-game staffs, the ones that got most beyond their stuff were 2019 LAD (+4.4 stuff, +17.3 actual), 2015 NYM (+0.8 stuff, +13.6 actual) and 2015 LAD (+2.2 stuff, +14.0 actual). The ones that got least out of it were 2019 BAL (+0.5 stuff, −12.6 actual), 2021 BAL (+0.4 stuff, −12.2 actual) and 2017 BAL (+1.6 stuff, −9.5 actual), both parts together, runs per 600.
9.Splits by hand
Every player's season, and every club's, is also split by the hand of the opponent: hitters by the throwing hand of the pitcher in each plate appearance, pitchers by the side the batter hit from in that plate appearance (so a switch hitter counts on the side he batted). The split values are descriptive and unregressed; the forecast regresses each player's gap toward the typical gap for his hand (Section 12). This section gives the exact rule for each component, the typical gaps the forecast uses, and the team-level splits.
Players: what is split and how
Hitters, vs LHP and vs RHP. Only batting is summed directly by the pitcher's hand. Situational outs are rated per opportunity against each hand; baserunning is computed as one season total and then divided by opportunity shares; fielding and catcher defense are divided by plate-appearance shares (ledger.py):
batting_s = sum over his PAs against hand s of bat(PA) (exact)
bat(PA) = LW1 value - (park factor - 1) x league runs per PA - league shift
situational_s = (situational runs_s / sitPA_s) x q x PA_s - c_sit x PA_s (0 if sitPA_s = 0)
sitPA_s = his PAs against hand s with runners on and fewer than 2 outs
q = league share of such PAs (0.2569 in 2026)
baserunning = adv / opp x (league opp / league stints) x stints + (steal runs - league steal runs per stint x stints)
baserunning_s = baserunning x opp_s / opp (PA_s / PA when he had no opportunity) - c_br x stints x baserunning_s / baserunning
fielding_s = (fielding + catcher defense) x PA_s / PA
Here opp counts his advancement opportunities as a runner already on base (a forced runner on a grounder is scored under the batter's situational outs, not here), stints his times on base, and c_sit, c_br the league constants that re-center each component to zero. Steal, caught-stealing and pickoff runs are recorded by pitcher hand, but the split does not use them separately: they are part of the season total that is divided by the opportunity shares. An earlier version of this section said situational outs and baserunning were "summed directly by pitcher hand"; only batting is. The positional adjustment is not split in the ledgers; the dashboard spreads it by the share of plate appearances against each hand.
Pitchers, vs LHB and vs RHB. Strikeouts and walks and contact quality are summed by the batter's side; everything that does not depend on the batter (the running game, the catcher throwing and blocking offset, the pitcher's own fielding, the framing offset) is divided by batters-faced share:
K&BB_s = - sum over PAs against side s of charge(PA) + framing offset x BF_s / BF
charge = LW1 value for BB, HBP, IBB, INTERF, K; league-average batted-ball value bbar for any ball in play
contact_s = - sum over balls in play against side s of (xv - bbar) - c_con x BF_s
other_s = (running game + catcher throwing/blocking offset + own fielding) x BF_s / BF - c_oth x BF_s
The two sides add exactly to the season total in every component, for hitters and pitchers.
Typical platoon gaps
The forecast's starting point for a player's split is the typical gap for his hand: the pooled rate against right-handed opponents minus the pooled rate against left-handed ones, over every player of that hand in the four seasons 2023–2026, with no playing-time minimum (so it is weighted by plate appearances or batters faced):
g(h) = sum runs_R / sum PA_R - sum runs_L / sum PA_L over all player-seasons of hand h, 2023-2026
hitters: batting; pitchers: strikeouts & walks + contact, by throwing hand
share_L(h) = sum PA_L / sum PA (pitchers: batters faced who hit left-handed)
s(h) = clamp( gap in strikeouts & walks(h) / g(h), 0, 1 ) pitchers only
s(h) = 0.5 if |g(h)| x 600 < 1 or fewer than 1,000 BF on either side, 2023-2026
| Handedness group | Player-seasons | Pooled 2023–2026 | 2023 | 2024 | 2025 | 2026 | SD of the four | Share vs left-handed opponents |
|---|---|---|---|---|---|---|---|---|
| Left-handed hitters | 908 | +15.9 | +13.6 | +14.4 | +19.1 | +16.1 | 2.4 | 22.8% |
| Right-handed hitters | 1,456 | −7.6 | −9.8 | −7.8 | −4.0 | −8.9 | 2.6 | 30.7% |
| Switch hitters | 278 | −0.9 | −4.6 | −4.3 | −0.4 | +4.2 | 4.1 | 28.8% |
| Left-handed pitchers | 889 | −12.1 | −14.2 | −12.4 | −11.8 | −10.3 | 1.6 | 30.2% |
| Right-handed pitchers | 2,568 | +9.6 | +8.2 | +9.4 | +9.8 | +11.0 | 1.2 | 49.4% |
| Switch pitchers | 2 | −4.5 | – | – | – | −24.6 | – | 32.1% |
Rate vs right-handed opponents minus rate vs left-handed opponents, runs per 600 PA (hitters, batting runs) or batters faced (pitchers, strikeouts and walks plus contact). Pitchers: vs RHB minus vs LHB. Switch pitchers are 2 player-seasons of 2 pitchers (53 batters faced in all), so their 'typical' gap is just their own; with batters faced from only one side in 2025, only 2026 has a single-season gap, so no SD is given.
The gaps are stable in sign and size but move by a few runs from season to season: the SD of the four seasons is 2.4 runs per 600 for left-handed hitters and 2.6 for right-handed ones, so the pooled value carries a standard error of roughly SD / √4, about 1.2 and 1.3. Over 2015–2026 (2020 aside) the left-handed hitters' gap ran from +10.6 to +20.7, the right-handed hitters' from −12.5 to −4.0, left-handed pitchers' from −18.0 to −10.3 and right-handed pitchers' from +5.4 to +11.0. The forecast pools only the last four seasons. The pitchers' share of the gap that is strikeouts and walks, s(h), is used to divide a projected pitcher's gap between his two parts (Section 12); pooled it is 0.52 for right-handers and 0.28 for left-handers, but it moves much more from season to season than the gap itself:
| Season | LHH gap | RHH gap | Switch gap | LHP gap | RHP gap | RHP K&BB share | LHP K&BB share |
|---|---|---|---|---|---|---|---|
| 2015 | +19.1 | −9.8 | −3.9 | −10.5 | +10.4 | 0.70 | 0.36 |
| 2016 | +18.7 | −9.4 | −1.9 | −13.5 | +7.5 | 0.81 | 0.14 |
| 2017 | +20.7 | −10.4 | +5.2 | −15.5 | +10.7 | 0.70 | 0.29 |
| 2018 | +17.9 | −7.7 | +3.1 | −12.3 | +9.7 | 0.67 | 0.38 |
| 2019 | +10.6 | −12.5 | +3.4 | −11.3 | +9.2 | 0.70 | 0.21 |
| 2020 | +15.9 | −7.1 | −12.4 | −10.9 | +3.7 | 1.00 | 0.32 |
| 2021 | +18.9 | −12.0 | −7.5 | −18.0 | +9.2 | 0.64 | 0.29 |
| 2022 | +14.0 | −9.5 | −6.6 | −14.3 | +5.4 | 0.79 | 0.15 |
| 2023 | +13.6 | −9.8 | −4.6 | −14.2 | +8.2 | 0.64 | 0.10 |
| 2024 | +14.4 | −7.8 | −4.3 | −12.4 | +9.4 | 0.54 | 0.38 |
| 2025 | +19.1 | −4.0 | −0.4 | −11.8 | +9.8 | 0.53 | 0.28 |
| 2026 | +16.1 | −8.9 | +4.2 | −10.3 | +11.0 | 0.41 | 0.36 |
Typical gaps by season, runs per 600, and the right- and left-handed pitchers' strikeouts-and-walks share of their gap (clamped to [0, 1]). 2020 was 60 games.
Team lineups, vs LHP and vs RHP
batting_t,s = sum over club t's PAs against hand s of bat(PA) (exact) sit_t,s, br_t,s = sum over hitters i of (his split value_s) x share(i, t) fielding_t,s = sum over hitters i of (fielding_i + catcher_i) x PA_i,s / PA_i x share(i, t) pitching_t,s = pitching_t x PA_t,s / PA_t (strikeouts & walks, contact, other) share(i, t) = PA(i, t) / PA(i)
Fielding follows each fielder's own share of plate appearances against each hand, weighted by his share of time with the club; it is not spread by the club's share, as an earlier version of this section said. Only the pitching parts are spread by the club's share of plate appearances against each hand, since the opposing pitcher's hand does not touch them. Situational outs and baserunning use each hitter's season split times his share of time with the club, so a traded player's split for a club is his season split, not his split with that club. The two sides add exactly to the season total; actual run differential is not split, because runs allowed have no opposing-pitcher hand. In the 2027 forecast the hitters' projected splits replace their actual ones, depth-chart newcomers bat at the league share against left-handed pitchers (30.0% of 2026's plate appearances), and pitching and newcomer pitchers are spread by the club's projected share.
Team staffs, vs LHB and vs RHB
BF_t,s = sum over pitchers j of BF_j,s x share(j, t) staff part_t,s = sum over pitchers j of part_j,s x share(j, t) (strikeouts & walks, contact, other) fielding_t,s = fielding_t x BF_t,s / BF_t (hitters' fielding + catcher defense)
Offense is excluded: it does not depend on the batter faced. vs LHB plus vs RHB equals the staff-and-defense total exactly. In the 2027 forecast each pitcher's projected parts by side come from Section 12 and depth-chart newcomer pitchers face the league share of left-handed batters (48.8% of 2026's batters faced). Until 2026-11-16 the projected parts were broken by side (the totals were right): the regressed gap was divided between strikeouts and walks and contact by the pitcher's own strikeouts-and-walks share of his overall rate, which explodes when the two parts have opposite signs and nearly cancel. 475 of 975 projected pitchers had that share outside [0, 1]; Sandy Alcantara (868 projected batters faced) was +1,185.0 in strikeouts and walks and −1,195.0 in contact per 600 vs LHB, and Miami's staff vs LHB +1,056.9 and −1,077.6 runs. With the fix (Section 12) no pitcher's split share is outside [0, 1] (0 of 975), the largest projected part on either side is 32.6 runs per 600, and Alcantara vs LHB reads −9.9 / −0.1. Totals, side totals and team wins did not change.
10.Team totals and reconciliation
The model's runs are built so that they add up: every component sums to zero over MLB each season (one exception: fielding −1.00 runs in 2020, from 4 fielders with innings but no plate appearance, whose centered DRS (+1.00) the model drops), and every player's runs are divided among the clubs he played for. The players' runs for a club can therefore be set against what the club actually did, and the gap is everything the model leaves out by design.
share(i, t) = PA(i, t) / PA(i) hitters; BF(j, t) / BF(j) for pitchers
players runs total_t = sum over club t's own PAs of bat(PA)
+ sum over hitters i of (situational_i + baserunning_i + fielding_i + catcher_i) x share(i, t)
+ sum over pitchers j of (strikeouts & walks_j + contact_j + running game_j + fielding_j) x share(j, t)
difference_t = (RD_t - mean RD) - players runs total_t RD = runs scored - runs allowed
% difference_t = difference_t / (RS_t + RA_t) x 100
Batting is the one part summed from the club's own plate appearances; every other part is a player's season value times his share of time with the club. That share is a share of plate appearances (hitters) or batters faced (pitchers), not of innings, so a traded catcher's framing and a traded fielder's DRS are divided by where he batted, not where he caught or fielded. The positional adjustment is not part of the total: it is for comparing players across positions, not runs a club scored or prevented. In a closed league runs scored and allowed sum to the same number, so the mean run differential is identically zero (largest absolute value over 2015–2026: 0.000), and the players' totals and the differences each sum to about zero over the 30 clubs (the largest league sum of players totals is −1.00 runs, in 2020, the fielding exception above).
Catcher defense nets out at the club. A club's framing is charged back to its own pitchers, and the catchers' throwing and blocking likewise to the pitchers' running game, so catcher value moves between a club's catchers and its pitchers without changing the club's total:
framing_t = sum over hitters i of framing_i x share(i, t)
offset_j -= sum over clubs t of framing_t x BF(j, t) / BF_t (in strikeouts & walks)
throwing + blocking: the same, into the running game
| Season | Catcher defense, sum of |runs| | SD across clubs of their catchers' runs | Largest |framing + framing offsets| at a club |
|---|---|---|---|
| 2015 | 0 | 0.0 | 0.00 |
| 2016 | 417 | 13.0 | 2.05 |
| 2017 | 344 | 10.8 | 1.63 |
| 2018 | 521 | 16.0 | 2.38 |
| 2019 | 491 | 14.1 | 1.70 |
| 2020 | 202 | 5.9 | 0.53 |
| 2021 | 403 | 10.3 | 1.59 |
| 2022 | 460 | 12.9 | 2.03 |
| 2023 | 422 | 11.6 | 2.03 |
| 2024 | 382 | 11.3 | 1.60 |
| 2025 | 377 | 12.2 | 1.42 |
| 2026 | 353 | 8.7 | 2.95 |
The identity is exact for a pitcher who stayed with one club; a traded pitcher's offset from one club is spread over both by his batters-faced shares, which leaves the small residual.
The check covers framing, the part the model stores for every pitcher; framing is most of catcher defense (277 of 353 absolute runs in 2026). The throwing and blocking offset is built the same way (ledger.py) but is stored only inside each pitcher's running game, so it is not checked here. For framing the residual is at most 2.95 runs at any club in any season, against an SD across clubs of 5.9 to 16.0 runs in the catchers' own values (all parts). Club totals therefore barely depend on whether catcher data exist (which matters for 2015, Section 11).
What the difference holds. Everything the model deliberately does not credit to a player:
- offensive and defensive timing and sequencing (the gap between run expectancy and the all-state event values, Sections 4 and 5);
- home runs and in-park outcomes beyond what exit velocity and launch angle predicted (pitchers are charged the expected value of contact);
- batted-ball conversion that DRS does not assign to a fielder;
- park effects on both sides: batting is park-adjusted, so a club's park effect on its runs scored lands here, and pitchers' parts are not park-adjusted, so contact and conversion in the park land here too;
- the play residual no player is credited with: runner advancement on hits that runner credits do not explain and, since 2026-11-16, the batter's share of fielder's choices that retire nobody, which went to his situational outs before (Section 6; league totals by season below, +179.5 to +503.6 runs, about +172 to +485 of it the fielder's choices);
- runs from the automatic extra-inning runner (2020 on): batting values are state-averaged, so the free runner's runs are credited to no one (not measured separately);
- innings played (home clubs skip the bottom of the ninth in wins);
- players with no plate appearance: the model skips hitters with 0 PA, so a pinch runner's baserunning and a defensive replacement's DRS fall into the difference (0 to 4 fielders with innings but no plate appearance a season, at most 1.00 runs of centered DRS).
| Season | Difference, RMS (runs) | Lowest (runs) | Highest (runs) | % difference, RMS | Lowest % | Highest % | Unassigned play residual (runs) | Players total over MLB (runs) |
|---|---|---|---|---|---|---|---|---|
| 2015 | 49.3 | −94.2 DET | +135.4 MIN | 3.6% | −7.2% LAD | +9.7% MIN | +387.6 | 0.0 |
| 2016 | 43.1 | −102.2 HOU | +76.2 KC | 3.0% | −7.2% HOU | +5.5% KC | +367.5 | 0.0 |
| 2017 | 50.1 | −143.5 CIN | +76.8 DET | 3.3% | −8.8% CIN | +4.7% DET | +425.5 | 0.0 |
| 2018 | 44.2 | −86.7 NYM | +72.7 KC | 3.1% | −6.3% NYM | +4.9% KC | +401.2 | 0.0 |
| 2019 | 54.0 | −150.6 HOU | +79.3 TEX | 3.5% | −9.7% HOU | +4.8% KC | +503.6 | 0.0 |
| 2020 | 19.6 | −39.3 NYM | +31.9 WSH | 3.5% | −6.6% NYM | +5.4% WSH | +179.5 | −1.0 |
| 2021 | 47.6 | −72.6 WSH | +141.1 DET | 3.2% | −4.7% WSH | +9.7% DET | +445.1 | 0.0 |
| 2022 | 27.6 | −75.6 MIL | +45.0 BOS | 1.9% | −5.4% MIL | +3.0% BOS | +419.1 | 0.0 |
| 2023 | 34.8 | −57.2 TOR | +56.3 CWS | 2.4% | −4.0% TOR | +3.8% CWS | +424.7 | 0.0 |
| 2024 | 35.7 | −100.0 BOS | +79.1 KC | 2.5% | −6.7% BOS | +5.7% KC | +419.6 | 0.0 |
| 2025 | 36.0 | −63.3 ATH | +96.9 CWS | 2.6% | −4.1% ATH | +7.0% CWS | +437.5 | 0.0 |
| 2026 | 40.7 | −108.0 ATH | +81.5 STL | 2.7% | −6.6% ATH | +5.5% STL | +482.2 | 0.0 |
RMS over the 30 clubs (the mean is zero). The unassigned residual is the league total of the play residual no player is credited with: run-expectancy changes on plays with runners that the runner credits do not explain, plus (since 2026-11-16) the batter's share of fielder's choices that retire nobody (ledger meta.unassigned_play; meta.fc_no_out holds the second part).
Why a share of runs scored and allowed. The dashboard shows the difference as a percentage so that clubs and seasons compare on one scale. Two other bases were rejected (the user's choice, recorded in the README) because both blow up when the base is near zero: as a share of the players runs total, 2026 BAL (players +0.5 runs) would read −4,643%, and 2024 TOR +5,905%; as a share of actual run differential, LAA in 2017 (run differential +1) would read +6,553%, and a club with a run differential of exactly zero (2020) would be undefined. Runs scored plus allowed is never near zero. The validation table (Section 11) used to report the spread in runs ("Unassigned SD"); the table above keeps it beside the percentage.
11.Validation
| Season | Hitters | Pitchers | Batting vs FG Batting | Baserunning vs FG BsR | Value vs FG WAR | Pitchers vs FG WAR | Team total vs run diff. | % difference SD | Difference SD (runs) | Runs / PA, season | Runs / PA, park-factor base |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2015 | 268 | 277 | 0.991 | 0.77 | 0.91 | 0.69 | 0.88 | 3.6% | 49.3 | 0.1124 | 0.1099 |
| 2016 | 268 | 280 | 0.987 | 0.82 | 0.92 | 0.64 | 0.92 | 3.0% | 43.1 | 0.1178 | 0.1151 |
| 2017 | 288 | 290 | 0.985 | 0.83 | 0.91 | 0.69 | 0.93 | 3.3% | 50.1 | 0.1219 | 0.1198 |
| 2018 | 278 | 274 | 0.987 | 0.79 | 0.91 | 0.75 | 0.95 | 3.1% | 44.2 | 0.1168 | 0.1194 |
| 2019 | 273 | 279 | 0.986 | 0.78 | 0.92 | 0.73 | 0.96 | 3.5% | 54.0 | 0.1258 | 0.1213 |
| 2020 | 267 | 244 | 0.986 | 0.65 | 0.89 | 0.71 | 0.93 | 3.5% | 19.6 | 0.1255 | 0.1257 |
| 2021 | 262 | 273 | 0.988 | 0.82 | 0.92 | 0.76 | 0.94 | 3.2% | 47.6 | 0.1211 | 0.1222 |
| 2022 | 277 | 274 | 0.990 | 0.82 | 0.93 | 0.70 | 0.98 | 1.9% | 27.6 | 0.1143 | 0.1177 |
| 2023 | 293 | 280 | 0.989 | 0.79 | 0.91 | 0.63 | 0.97 | 2.4% | 34.8 | 0.1218 | 0.1181 |
| 2024 | 286 | 277 | 0.989 | 0.81 | 0.91 | 0.67 | 0.95 | 2.5% | 35.7 | 0.1170 | 0.1194 |
| 2025 | 277 | 289 | 0.991 | 0.81 | 0.92 | 0.65 | 0.96 | 2.6% | 36.0 | 0.1182 | 0.1176 |
| 2026 | 267 | 277 | 0.986 | 0.75 | 0.89 | 0.70 | 0.92 | 2.7% | 40.7 | 0.1184 | 0.1183 |
Pearson correlations (unweighted) across hitters with 300+ PA and pitchers with 250+ batters faced (120 and 100 in the 60-game 2020 season), and across the 30 clubs. FG = FanGraphs: Batting (park-adjusted batting runs), BaseRunning (BsR), WAR. Value = runs + positional adjustment; pitchers' runs include the running game and their own fielding. Runs / PA, season = runs scored over plate appearances that season; park-factor base = the two-season figure the model stores (seasons y-1 and y pooled, the base of the park factors).
The last column used to be labeled "League runs / PA" and described as that season's figure. It is the stored meta.lg_runs_per_pa, the league rate over two seasons (the season and the one before) that the park factors are built on: 2015's 0.1099 pools 2014 and 2015, while 2015 alone scored 0.1124 runs per plate appearance. Both are now shown.
95% intervals (Fisher z for each correlation, with that season's n) and two data-coverage facts:
| Season | Batting r 95% | Baserunning r 95% | Value r 95% | Pitchers r 95% | Team r [95%] | Balls in play without exit velocity | Catcher defense, sum of |runs| |
|---|---|---|---|---|---|---|---|
| 2015 | [0.988, 0.993] | [0.72, 0.82] | [0.88, 0.93] | [0.62, 0.75] | 0.883 [0.77, 0.94] | 13.1% | 0 |
| 2016 | [0.983, 0.990] | [0.78, 0.86] | [0.89, 0.93] | [0.56, 0.70] | 0.924 [0.84, 0.96] | 12.2% | 417 |
| 2017 | [0.981, 0.988] | [0.79, 0.86] | [0.89, 0.93] | [0.62, 0.74] | 0.925 [0.85, 0.96] | 10.6% | 344 |
| 2018 | [0.984, 0.990] | [0.75, 0.83] | [0.89, 0.93] | [0.69, 0.80] | 0.953 [0.90, 0.98] | 8.8% | 521 |
| 2019 | [0.982, 0.989] | [0.73, 0.82] | [0.90, 0.94] | [0.67, 0.78] | 0.959 [0.91, 0.98] | 8.8% | 491 |
| 2020 | [0.982, 0.989] | [0.58, 0.72] | [0.86, 0.91] | [0.65, 0.77] | 0.933 [0.86, 0.97] | 0.7% | 202 |
| 2021 | [0.984, 0.990] | [0.77, 0.86] | [0.89, 0.93] | [0.71, 0.81] | 0.942 [0.88, 0.97] | 0.4% | 403 |
| 2022 | [0.988, 0.992] | [0.78, 0.86] | [0.91, 0.94] | [0.64, 0.76] | 0.982 [0.96, 0.99] | 0.5% | 460 |
| 2023 | [0.986, 0.991] | [0.74, 0.83] | [0.89, 0.93] | [0.55, 0.69] | 0.972 [0.94, 0.99] | 0.3% | 422 |
| 2024 | [0.986, 0.991] | [0.76, 0.85] | [0.89, 0.93] | [0.60, 0.73] | 0.954 [0.90, 0.98] | 0.3% | 382 |
| 2025 | [0.988, 0.993] | [0.76, 0.84] | [0.90, 0.94] | [0.58, 0.71] | 0.961 [0.92, 0.98] | 0.4% | 377 |
| 2026 | [0.982, 0.989] | [0.69, 0.80] | [0.86, 0.91] | [0.63, 0.75] | 0.923 [0.84, 0.96] | 0.4% | 353 |
Balls in play without a recorded exit velocity or launch angle are valued at their trajectory's average (prevention_decomp.expected_bip). Catcher defense is 0 for every catcher in 2015: FanGraphs has no Statcast catcher data that year.
- Batting agrees with FanGraphs' park-adjusted batting runs at r = 0.985 to 0.991 every season. Both are linear-weights run values; the remaining differences are the out-type detail, reached-on-error, FanGraphs' league adjustment, and the park factors (here two seasons of home-versus-road runs per PA, regressed 50% of the way to 1.00).
- Baserunning agrees at 0.75 to 0.83 (0.65 in 2020). The methods differ in design: our advancement is conditioned on exit-velocity band and rated per opportunity, and double-play avoidance sits in situational outs rather than baserunning.
- Value with position tracks FanGraphs WAR at 0.89 to 0.93 for hitters. For pitchers the agreement is lower (0.63 to 0.76) by design: WAR counts innings above replacement level, so a full-time league-average pitcher is worth about 1.8 WAR per 600 batters faced (the fit below) but zero here; and FanGraphs' pitcher WAR is FIP-based, ignoring contact quality, which this model credits. The WAR subsection below fits the two scales.
- Team totals track actual run differential at r = 0.88 to 0.98. 2015 is lowest, r = 0.883 with a 95% interval of [0.77, 0.94], which overlaps several other seasons' intervals, so the ranking of seasons is not established. An earlier version said 2015 is lowest "because Statcast fielding (catcher defense) and full exit-velocity tracking were not yet in place". The catcher half of that cannot be right: catcher defense is charged back to the club's own pitchers (Section 10; checked for framing, the bulk of it), so a club's total barely depends on catcher data, and the fielding in the model is DRS, which exists every season. The exit-velocity half is possible but untested: 2015, 2016, 2017, 2018 and 2019 had 8.8% to 13.1% of balls in play without exit velocity, against 0.7% or less from 2020 on, and their team correlations average 0.929 against 0.952 for 2020–2026, a difference that is borderline: comparing the eras' mean Fisher z with only sampling error (each season's z has variance 1/27) gives z = 1.96 (p = 0.050), while a Welch t on the seasons' observed spread gives t = 1.59 (9.7 degrees of freedom, p = 0.14); a direct test would value every season's contact by trajectory alone and see whether the later seasons' fit falls to the earlier level.
- 2020: the paper's samples use 120 PA and 100 batters faced. The dashboard's Method checks scale 300 PA and 250 batters faced by the season's games instead (60 games: 111 PA and 92 batters faced), so their 2020 numbers can differ slightly from these.
FanGraphs WAR alongside the model
The dashboard shows FanGraphs WAR next to the model's runs everywhere a player or team total appears: the hitters, pitchers and teams tables, the leader cards, the player and team cards (season by season, with a career total), and the Top 100 rookies. A player card lists every MLB season of his career with FanGraphs' stat line and WAR; the model's runs appear from 2015 (the first season on the dashboard) and a dash before. It is FanGraphs' own figure, shown for reference, not an output of this model. For the 2027 forecast it is FanGraphs Depth Charts' projected WAR. It is left out of the views split by hand, because WAR exists only for the full season.
team WAR = sum over hitters and pitchers of FanGraphs WAR x team share
The team share is the same share of plate appearances or batters faced that divides a traded player's runs, so team WAR and the players runs total are split the same way. (FanGraphs' own team pages split by the games a player actually played for each club, which can differ slightly for traded players.) Summed over the 30 clubs with unrounded shares this reproduces FanGraphs' own league total to within 0.07 WAR in every season, which checks that the shares account for every player; the dashboard's team WAR is rounded to 0.1 per club, so its sum (the table's fourth column) differs from FanGraphs' by up to 0.3. That total is not exactly 1,000: FanGraphs' player rows sum to 999.0 to 1,000.4 WAR in the full seasons (rounding of each player's WAR) and 369.5 in the 60-game 2020. An earlier version of this paragraph said "1,000 WAR in every full season".
| Season | FanGraphs league WAR (player rows) | Clubs summed, unrounded | Clubs summed (dashboard, 0.1 per club) | Team WAR vs wins | Players runs total vs wins | Team WAR vs players runs total | Williams t (WAR minus players) | p |
|---|---|---|---|---|---|---|---|---|
| 2015 | 999.6 | 999.6 | 999.5 | 0.87 | 0.79 | 0.94 | 2.20 | 0.036 |
| 2016 | 999.0 | 999.0 | 999.0 | 0.83 | 0.85 | 0.94 | −0.58 | 0.568 |
| 2017 | 1,000.0 | 1,000.0 | 1,000.3 | 0.91 | 0.91 | 0.94 | 0.00 | 0.997 |
| 2018 | 1,000.4 | 1,000.4 | 1,000.4 | 0.92 | 0.91 | 0.96 | 0.44 | 0.663 |
| 2019 | 999.6 | 999.6 | 999.6 | 0.93 | 0.93 | 0.95 | −0.11 | 0.910 |
| 2020 | 369.5 | 369.6 | 369.4 | 0.80 | 0.88 | 0.90 | −1.89 | 0.070 |
| 2021 | 999.6 | 999.6 | 999.6 | 0.92 | 0.86 | 0.94 | 2.52 | 0.018 |
| 2022 | 1,000.0 | 1,000.1 | 999.8 | 0.98 | 0.97 | 0.98 | 1.75 | 0.092 |
| 2023 | 1,000.0 | 1,000.0 | 1,000.1 | 0.90 | 0.90 | 0.97 | 0.10 | 0.920 |
| 2024 | 999.6 | 999.6 | 999.9 | 0.94 | 0.95 | 0.95 | −0.01 | 0.995 |
| 2025 | 1,000.0 | 1,000.0 | 1,000.1 | 0.92 | 0.91 | 0.97 | 0.55 | 0.587 |
| 2026 | 999.6 | 999.6 | 999.6 | 0.88 | 0.84 | 0.94 | 1.03 | 0.313 |
Correlations across the 30 clubs. Team WAR is every player's FanGraphs WAR times his share of time with the club; players runs total is this model's runs with the same shares. Williams' t compares the two correlations with wins, which share wins (n = 30, 27 degrees of freedom); positive = team WAR tracks wins better.
Team WAR and the model's players runs total agree closely across clubs. Which of the two tracks wins better changes from season to season: rounded to two decimals, team WAR is ahead in 6 seasons, the players runs total in 3, and they tie in 3. The two correlations share wins, so they are compared with Williams' test for dependent correlations rather than with separate intervals: team WAR tracks wins significantly better (5% level) in 2015 and 2021, the players runs total in no season (closest: 2020, t = −1.89, p = 0.070), and |t| is below 1 in 7 of the 12 seasons. Two significant results in 12 tests at the 5% level are more than the 0.6 expected by chance, but not by much (if the tests were independent, the chance of 2 or more would be 12%); the clubs' WAR, which uses FanGraphs' own fielding, baserunning and FIP, tracks wins at least as well as the model's runs. Neither total includes the timing, sequencing and park effects that make up the % difference between a club's actual run differential and its players runs total (Section 10).
Two scales. WAR counts wins above a replacement-level player; the model counts runs against an average player. Fitted across 2015–2026 with 2020 left out, by unweighted least squares through the origin (no intercept), on the dashboard's values (rounded to 0.1), FanGraphs WAR is very nearly a fixed amount per unit of playing time plus a fixed fraction of the model's value; r is the correlation of the fitted with the actual WAR:
hitters (300+ PA, 3,037 seasons): WAR = 1.98 x PA/600 + 0.083 x value with position r = 0.94 pitchers (250+ batters faced, 3,070 seasons): WAR = 1.79 x BF/600 + 0.076 x runs prevented r = 0.89
So a full-time average player (zero here) is worth about 2.0 WAR as a hitter (1.8 per 600 batters faced as a pitcher), and roughly 12 of the model's runs make one FanGraphs win for hitters (13 for pitchers). That is more than the league's own runs per win (9.54, Pythagenpat at 2026's scoring, Section 15). An earlier version put this down to attenuation: the two methods disagree on the details (fielding source, baserunning, contact quality), and disagreement in the explanatory variable flattens a least-squares slope. That is only part of it. The reverse regression, the model's value on PA/600 and WAR (also through the origin, unrounded ledger values), gives 10.1 runs per WAR for hitters and 8.6 for pitchers; with noise on both sides the true conversion lies between the two fits, 10.1–12.1 for hitters and 8.6–13.2 for pitchers. The league's 9.54 lies outside the hitters' bracket, so for hitters attenuation alone does not explain the gap; the remaining difference in scale has not been traced. Pitchers agree less because FanGraphs' pitcher WAR is FIP-based and ignores the contact quality this model credits. On 2017–2026 alone (2,501 hitter and 2,513 pitcher seasons) the fit is 2.00 / 0.083 (r = 0.94) and 1.79 / 0.076 (r = 0.89); the version published before 2015–2016 were added read 2.00 / 0.082 (r = 0.94, 2,501 seasons) and 1.79 / 0.076.
The depth charts' projected WAR for the 2027 clubs sums to 1,159, more than any played season's league total, because depth charts plan full seasons for regulars that injuries then cut. It tracks this forecast's projected team runs at r = 0.93 (on the first depth-chart pull it summed to 1,160 with r = 0.90). The dashboard shows FanGraphs' figure as published, without rescaling.
12.The 2027 forecast
Pool
Every player who appeared in MLB in 2026, plus depth-chart players back from a lost season (22 hitters, 51 pitchers: MLB time in 2024 or 2025, none in 2026, and projected playing time) and depth-chart players up from Triple-A or Double-A with no MLB time in the window (80 and 112; Section 14): 761 hitters and 975 pitchers in all. A 2026 major leaguer is left off the hitters' list when he faced more batters than he had plate appearances and had fewer than 100 plate appearances (a pitcher who batted), and off the pitchers' list in the mirror case (a position player who pitched); two-way players are on both. The rates below describe players with MLB history. For a player back from a lost season the history is just 2024 and 2025 (the 2026 weight is simply absent), which also lengthens his aging gap (Section 13). Anyone with Triple-A or Double-A time in the window also gets the minor-league translation as his starting point (Section 14); the back-tests in this section do not include it.
Rates
Each component's 2027 rate combines the player's last three seasons, weighted toward the most recent (Marcel's weights, assumed rather than fitted: hitters 5/4/3, pitchers 3/2/1), and adds k units of league-average play so that small samples are pulled toward average. Seasons he did not play are skipped, not counted as zero:
rate(c) = sum_y w_y x runs_y(c) / ( sum_y w_y x units_y + k_c ) units = PA (hitters) or BF (pitchers) then + aging for c (Section 13); pitchers' strikeouts & walks and contact use the pitch-quality prior (below) positional adjustment: sum_y w_y x pos_y / sum_y w_y x PA_y (not regressed)
Because every component is already relative to league average, "league-average play" contributes zero runs and k acts purely as a pull toward average (league centering, Section 15, then sets that added play a little below average). The positional adjustment is not regressed: it is a fixed charge for the position, not a skill.
Fitting k
Each k is fitted by back-test rather than assumed (forecast.fit_k): exhaustive search over the 21-point grid {0, 25, 50, 100, 150, 200, 300, 400, 500, 650, 800, 1,000, 1,250, 1,500, 2,000, 2,500, 3,000, 4,000, 5,000, 7,000, 10,000}, minimizing
sum over target players with units_T >= 100 and history units > 0: units_T x ( y_T(c) / units_T - rate(c) )^2 histories: the two seasons before the target (weights 5/4 or 3/2); no aging, pitch quality or minor-league translation in the fit
It is fitted once with 2023–24 predicting 2025 (409 hitters, 453 pitchers) and once with 2024–25 predicting 2026 (409 and 457). The final k is the arithmetic mean of the two grid optima, not a pooled fit (a code comment that called it pooling "by summing errors" was corrected on 2026-11-16); the forecast's out-of-sample test uses the first fit alone.
| Component | k fit on 2025 | Within 0.5% | k fit on 2026 | Within 0.5% | Mean of the two | Final k | RMSE at k = 0 / at best (2025) | |
|---|---|---|---|---|---|---|---|---|
| Hitters | Batting | 2,500 | 2,000–3,000 | 4,000 | 4,000–5,000 | 3,250 | 3,250 | 1.26 |
| Hitters | Situational outs | 7,000 | 7,000–10,000 | 7,000 | 5,000–7,000 | 7,000 | 7,000 | 1.38 |
| Hitters | Baserunning | 4,000 | 4,000–5,000 | 4,000 | 3,000–5,000 | 4,000 | 4,000 | 1.27 |
| Hitters | Fielding | 2,500 | 2,000–3,000 | 4,000 | 3,000–4,000 | 3,250 | 3,250 | 1.17 |
| Hitters | Catcher defense | 1,000 | 800–1,000 | 2,000 | 2,000–2,500 | 1,500 | 1,500 | 1.31 |
| Pitchers | Strikeouts and walks | 650 | 650–800 | 800 | 650–1,000 | 725 (not used) | 1,250 (pitch quality, pooled) | 1.36 |
| Pitchers | Contact quality | 2,000 | 1,500–3,000 | 2,000 | 1,500–3,000 | 2,000 (not used) | 5,000 (pitch quality, pooled) | 1.40 |
| Pitchers | Running game | 2,000 | 2,000–3,000 | 4,000 | 3,000–5,000 | 3,000 | 3,000 | 1.43 |
| Pitchers | Pitcher fielding | 4,000 | 3,000–7,000 | 7,000 | 4,000–10,000 | 5,500 | 5,500 | 1.31 |
'Within 0.5%': every grid value whose error is within 0.5% of the minimum. The last column is the ratio of root mean squared errors with no regression at all (k = 0) and at the fitted k, on the 2025 target (the square root of the ratio of the fit's mean squared errors). Pitchers' strikeouts and walks and contact use pitch quality's pooled fit (below) in the forecast; their plain fits are what the 'without pitch quality' comparison uses.
Three things the table shows. First, the fits are not precise: within one target a wide band of k is within 0.5% of the best, and for batting (2,000–3,000 against 4,000–5,000) and catcher defense (800–1,000 against 2,000–2,500) the two targets' bands do not even overlap, so averaging two optima is a rough compromise rather than a pooled estimate. Second, situational outs fit at 7,000 and 7,000; an earlier version said they "hit the top of the grid" and are "pulled essentially all the way to average". Extending the grid to 15,000, 20,000, 30,000, 50,000, 100,000 still picks 7,000 and 7,000, so 7,000 is an interior optimum, and the pull is strong but not total: a regular with 600 PA in each of three seasons (7,200 weighted PA) keeps 51% of his own rate (44% with the two-season histories of the fit). Catcher defense is zero for non-catchers, who add no error at any k, so its fit is in effect over catchers. Third, the fits use two-season histories while the 2027 forecast uses three, so the same k regresses a three-season player less than it was fitted to (below).
Pitch quality as the pitchers' prior
For two pitcher components, strikeouts and walks and contact quality, the k units of prior play do not sit at league average but at what the pitcher's pitches predict. This subsection gives the data, the pitch values, the stuff model, how seasons are scored, the prior, how its settings were fitted, every test, and what was tried and dropped.
Data. Every MLB pitch from 2021 through 2026 is pulled from the Stats API game feed with its tracking (fetch_pitches.py): 4.26 million pitches (2021: 709,017, 2022: 707,574, 2023: 717,003, 2024: 708,425, 2025: 708,985, 2026: 713,382). A pitch counts as tracked when its type falls in one of the four groups below and velocity, movement, spin, extension, release point and plate location are all present: 99.2% to 99.6% of each season's pitches. An earlier version of this paragraph said 2021 is "the first season of the current tracking system". It is not: Hawk-Eye has measured every MLB pitch since 2020 (Section 8). The forecast's pitch data start in 2021 because its first fit target is 2023 with two-season histories (2021–22), so 2021 is the earliest season it needs. 2020, a 60-game season, was later pulled for the dashboard's display scores only and is kept out of pitch_quality.py on purpose (it reads only data/pitches/, not data/pitches/early/); whether adding it to the stuff models' training would help has not been tested.
Pitch values in the model's runs. Each pitch is valued by the change it makes in what the plate appearance is expected to cost the pitcher in strikeouts and walks (Section 8), from the batter's view; a pitcher's values are the negatives:
V(b, s) = mean strikeouts-and-walks charge of that season's PAs that reached count b-s
value(pitch) = V(count after it) - V(count before it) every pitch but the last
= charge(PA) - V(count before it), contact value = xv - bbar the pitch that ends the PA
sum over a PA = charge(PA) - V(0-0) (contact: xv - bbar)
The count table V is built per season from the plate appearances present in both the pitch file and the play-by-play. Summed over a season the values reproduce each pitcher's strikeouts and walks in the model at r = 0.992 to 0.998 (pitchers with 250+ batters faced); the gap from 1 is mainly the framing offset, which the model charges and the pitches do not. Traps found while building it: the Stats API's pitch count is the count after the pitch (the parser carries the count before), and horizontal break and release side change sign with the pitcher's hand.
The stuff model. A ridge-regularized quadratic regression, fitted separately in each of 16 cells (4 pitch groups × same or opposite hand × two strikes or not), for the strikeouts-and-walks value and for the contact value:
groups fastball FA, FF, SI; cutter FC; breaking CS, CU, KC, SL, ST, SV; offspeed CH, FO, FS, SC; any other type is not modeled
features velocity, induced vertical break, horizontal break*, spin, extension, release height, release side*
+ for every group but fastballs: the gap in velocity, IVB and HB from the pitcher's main fastball
(* mirrored for left-handers; main fastball = his most-thrown of FF/SI/FA that season, else FC)
design [1, z, z_i z_j for all i <= j, 1(balls = 1), 1(balls = 2), 1(balls = 3), 1(strikes = 1)]
z = features standardized within the cell: 40 terms for fastballs, 70 for the other groups
estimate b = (D'D + lambda x I*)^-1 D'y, lambda = 1, I* = identity with the intercept unpenalized
Pitches that are not tracked, or of a type outside the four groups, are predicted at 0, an average pitch, while their actual values still count in the pitcher's results. Fitted on 2021–2024 and scored on 2025's pitches, the model explains 0.46% of the variance of single pitches' strikeouts-and-walks values and 0.13% of contact values; in sample the same fit explains 0.80% and 0.24% (1.7 and 1.9 times as much). Single pitches are mostly noise, but summed over a season the predictions are a stable trait (Section 8).
Leave-season-out scoring. Because the polynomial learns part of the pitchers' own results in sample, no season is ever scored by a model that saw it, and no target season is ever in training:
train(T, S) = { y in 2021..2026 : y <= T - 1 and y != S } for history season S of target T
e.g. target 2026: 2025 scored by 2021-2024; 2024 by 2021-2023, 2025; 2023 by 2021-2022, 2024, 2025
target 2023: 2022 scored by a model fitted on 2021 alone, and 2021 by 2022 alone
The first run scored every season in sample (output/pitch_quality_test_insample.json): it put the pooled leave-one-target-out miss at 12.88 against 13.07, a gain of 0.19 where leave-season-out scoring finds 0.12, and next-season correlations of 0.443 (strikeouts and walks) and 0.422 (contact) where the honest figures are 0.417 and 0.407; it also had no kq shrinkage, which was added in the second version.
Pitch quality (PQ) of a pitcher for a target season is his weighted pitch quality per batter faced over his history, in pitcher runs (positive = good), one number per component:
PQ(c) = sum_y w_y x Q_y(c) / sum_y w_y x BF_y over his history seasons that have a score; 0 if none
Q_y(c) = his summed predicted pitch values in season y; w = 3/2/1 (two-season fits: 3/2)
Does it carry information beyond results? Pooling the season pairs 2022-2023, 2023-2024, 2024-2025 and 2025-2026 (661 pairs, 335 different pitchers, 250+ batters faced in both seasons), each season's single-season PQ (scored by models fitted on seasons before it) predicts next season's contact at r = 0.407 [0.34, 0.47] while this season's contact predicts it at 0.323 [0.25, 0.39]. Williams' test for the two dependent correlations (they share next season, and PQ and this season's contact correlate at 0.450) gives t = 2.27 on 658 degrees of freedom, but it assumes independent pairs and the same pitcher appears in up to four; resampling pitchers (a cluster bootstrap, 4,000 resamples) puts the difference at +0.084 [−0.005, +0.163], positive in 96.7% of resamples. For strikeouts and walks the results win clearly: 0.583 against 0.417 (t = −4.99; cluster bootstrap −0.166 [−0.249, −0.083]). So pitch quality can only add to strikeouts and walks, and for contact it is at least as good as results: ahead by 0.084 in correlation (Williams t = 2.27, p = 0.023 if pairs were independent; ahead in 96.7% of the pitcher-cluster resamples, whose interval just reaches zero).
The prior.
rate(c) = ( sum_y w_y x runs_y(c) + k_c x beta_c x PQ(c) x U / (U + kq_c) ) / ( U + k_c ) then + aging U = sum_y w_y x BF_y (every history season, scored or not) c = strikeouts & walks, contact
beta scales pitch quality into the component's runs, kq shrinks it for a pitcher with few batters faced, and with beta = 0 this is the plain forecast. League centering (Section 15) later adds the same level m_c to this prior as to every other. The three settings are fitted together for each component by exhaustive grid search (forecast.fit_pq): k in {0, 25, 50, 100, 150, 200, 300, 400, 500, 650, 800, 1,000, 1,250, 1,500, 2,000, 2,500, 3,000, 4,000, 5,000, 7,000, 10,000, 15,000, 20,000, 30,000}, beta from 0.0 to 2.0 by 0.1, kq in {0, 50, 100, 200, 400, 800, 1,600, 3,200}: 4,032 combinations, minimizing
sum over targets T, sum over pitchers with BF_T >= 100 and U > 0: BF_T x ( y_T(c) / BF_T - rate(c) )^2 histories: T-1 and T-2 (weights 3, 2); no aging, no minor-league translation in the fit
| Fit | K&BB k | K&BB beta | K&BB kq | Contact k | Contact beta | Contact kq |
|---|---|---|---|---|---|---|
| Test: fit on 2025 (2023–24 histories) | 800 | 0.3 | 200 | 3,000 | 0.4 | 0 |
| Final: 2023–2026 pooled | 1,250 | 0.7 | 400 | 5,000 | 0.5 | 50 |
| Held out 2023 (fit on the other three) | 1,250 | 0.5 | 100 | 4,000 | 0.5 | 0 |
| Held out 2024 (fit on the other three) | 1,000 | 0.7 | 400 | 5,000 | 0.7 | 200 |
| Held out 2025 (fit on the other three) | 1,500 | 0.7 | 400 | 7,000 | 0.6 | 200 |
| Held out 2026 (fit on the other three) | 1,250 | 0.5 | 200 | 7,000 | 0.5 | 50 |
The forecast's test uses the 2025 fit; the 2027 forecast uses the pooled final fit; the held-out rows are protocol B's fits (below).
The final settings are inside the grid on every axis. Without output/pitch_quality_scores.json, or with any target from 2023 to 2027 missing from it, the forecast runs without pitch quality and prints a note, so an old file never feeds the wrong year. The weight pitch quality gets moves from season to season: fitted on 2025 alone it is small (beta 0.3 and 0.4); fitted on each set of three targets (protocol B) it is 0.5–0.7 for strikeouts and walks and 0.5–0.7 for contact, with k 1,000–1,500 and 4,000–7,000. An earlier check regressed regulars' target rates freely on their history and pitch quality: predicting 2025 the free weight on pitch quality was 0.02 (strikeouts and walks) and 0.09 (contact), predicting 2026 0.40 and 0.55. That instability is why the final settings pool all four targets with pitch data, 2023–2026, each predicted from its two seasons before (the other components average two single-target fits, above). What the settings mean for a pitcher, as the effective coefficient on his PQ, k / (U + k) × beta × U / (U + kq):
| Weighted BF (U) | Test: K&BB | Test: contact | Final: K&BB | Final: contact |
|---|---|---|---|---|
| 300 | 0.131 | 0.364 | 0.242 | 0.404 |
| 1,000 | 0.111 | 0.300 | 0.278 | 0.397 |
| 3,000 | 0.059 | 0.200 | 0.182 | 0.307 |
| 3,600 | 0.052 | 0.182 | 0.162 | 0.287 |
Effective weight of a pitcher's PQ in his projected rate, before aging and centering. A pitcher with 600 batters faced in each of three seasons has U = 3,600 in the 2027 forecast (3/2/1 weights); the fits use two seasons, U = 3,000.
The forecast's own test. Settings fitted on 2025, histories 2024–25 predicting 2026, pitchers with 250+ batters faced (250): pitch quality cuts the typical miss from 13.31 to 13.05 runs per 600 batters faced (r 0.452 to 0.480); paired bootstrap of pitchers (4,000 resamples) −0.26 [−0.43, −0.08], better in 99.7% of resamples. By component: strikeouts and walks 8.17 to 8.01, contact 8.74 to 8.53; for 250–499 batters faced (147) 15.80 to 15.57, for 500+ (103) 11.17 to 10.87. The version without pitch quality uses the plain k (strikeouts and walks 650, contact 2,000) and picks its own aging shapes (strikeouts and walks curve 2015–25, contact quality curve 2015–25, against strikeouts and walks curve 2015–25, contact quality line 2015–25 with pitch quality); with the same aging the plain version scores 13.31, so the difference is pitch quality's.
Protocol A in pitch_quality.py repeats that test for the variants. "Stuff plus location" adds the plate location's value: the mean pitch value of the pitch's location cell (0.2 ft wide × 1/8 of the batter's zone height, horizontal position mirrored for left-handed batters) for its group, count group (two strikes or not × two or more balls or not) and platoon side, shrunk toward the cell's all-group mean with a weight of 100 pitches, entering as l, l2 and l times velocity, IVB and HB. "Stuff, contact quality only" puts the prior on contact and leaves strikeouts and walks plain. The rerun of 2026-11-16 12:22 uses Opening Day ages and the forecast's current aging, and reproduces the forecast's own figures; the last column is the run of 2026-11-15 (July 1 ages), which the paper reported until this rerun.
| Variant | RMSE per 600 BF | r | Difference vs without [95%] | Better in bootstrap | July 1 ages run |
|---|---|---|---|---|---|
| Without pitch quality | 13.31 | 0.452 | 13.29 | ||
| Stuff (used) | 13.05 | 0.480 | −0.26 [−0.44, −0.09] | 99.8% | 13.03 |
| Stuff plus location | 13.10 | 0.474 | −0.21 [−0.37, −0.05] | 99.4% | 13.09 |
| Stuff, contact quality only | 13.16 | 0.467 | −0.15 [−0.28, −0.02] | 98.7% | 13.14 |
2026 test, 250 pitchers with 250+ batters faced. Paired bootstrap of pitchers from the per-pitcher errors stored in pitch_quality_test.json.
Protocol B holds out each target from 2023 to 2026 in turn, fits k, beta and kq (and the plain k of the running game and fielding) on the other three targets, two-season histories, and scores the held-out season (pitchers in the target who also pitched the season before). The arms without a pitch-quality prior search k on the plain grid (to 10,000); the arms with one on the extended grid (to 30,000):
| Typical miss per 600 BF | 2023 | 2024 | 2025 | 2026 | Pooled | Pooled difference [95%] | Better in bootstrap |
|---|---|---|---|---|---|---|---|
| Without pitch quality | 13.71 | 12.43 | 12.77 | 13.32 | 13.06 | ||
| Stuff (used) | 13.46 | 12.43 | 12.89 | 12.98 | 12.95 | −0.12 [−0.29, +0.04] | 93% |
| Stuff plus location | 13.48 | 12.42 | 12.87 | 13.07 | 12.96 | −0.10 [−0.26, +0.05] | 89% |
| Stuff, contact quality only | 13.53 | 12.48 | 12.74 | 13.16 | 12.98 | −0.09 [−0.20, +0.03] | 93% |
| Strikeouts & walks: without / with | 8.14 / 8.05 | 7.86 / 7.73 | 7.37 / 7.48 | 8.16 / 7.88 | |||
| Contact quality: without / with | 9.13 / 8.83 | 8.02 / 7.97 | 8.22 / 8.03 | 8.79 / 8.56 |
Pooled = square root of the folds' mean squared error, each fold weighted by its number of pitchers (996 pitcher-seasons, 514 pitchers). Weighting by batters faced instead gives 13.067 and 12.946. The interval is a paired percentile bootstrap of the BF-weighted pooled RMSE difference, 4,000 resamples of the pitcher-seasons. The stored 'better' share (pitch_quality.py, seed 7) resamples the 996 pitcher-seasons pooled across the four folds, not within each fold, and counts the resamples in which the BF-weighted sum of squared errors falls: 92.5% for the stuff version. The two component rows compare the plain forecast with the stuff version.
Contact improves in 4 of 4 held-out seasons and strikeouts and walks in 3 of 4. In total the stuff version wins 2023 and 2026, ties 2024 and loses 2025 (shown to two decimals). Pooled, the gain is +0.12 runs per 600 with a 95% interval of [−0.04, +0.29]: protocol B alone does not exclude zero, protocol A (one target) does. The July 1 run gave 13.07 and 12.96 (91.6%). Two caveats: the aging in every fold is the forecast's test aging, fitted on season pairs through 2024–25 that include the held-out seasons 2023–2025 as pair members (mild leakage, the same for both arms); and the 2023 fold's pitch quality comes from one-season models.
Location. Adding plate location explains far more of single pitches (19.1% of strikeouts-and-walks values and 0.62% of contact, out of sample) and its pitch quality does predict next season's strikeouts and walks better than stuff alone (r = 0.496 against 0.417; contact 0.419). An earlier version said it explains "nothing more of next season"; that is wrong. What is true is that it is largely the results again: location pitch quality correlates 0.634 with the same season's strikeouts and walks (stuff: 0.446), so once the forecast has the results it adds nothing (protocol B pooled 12.96 against 12.95 for stuff alone). Command shows up in results; the forecast does not need it twice.
Minor leaguers get the same measure from Triple-A pitches, scored by the MLB stuff model, as two more pieces of the minor-league translation (Section 14).
Splits in the forecast
The player's platoon gap is regressed toward the typical gap for his hand (Section 9), using the harmonic sample of his weighted plate appearances against each hand, and the split is built around his overall rate so that the overall projection is unchanged:
gap(i) = sum w_y runs_R / sum w_y PA_R - sum w_y runs_L / sum w_y PA_L his weighted history
gap_r(i) = g(h) + (gap(i) - g(h)) x n / (n + k_gap), n = 1 / (1/n_L + 1/n_R), n_s = sum w_y PA_s
no plate appearance against one hand (n_L or n_R < 1): gap_r = g(h)
share_L(i) = (n_L + 200 x share_L(h)) / (n_L + n_R + 200) PA_L = PA x share_L, PA_R = PA - PA_L
hitters batting_L = batting - (1 - share_L) x gap_r, batting_R = batting + share_L x gap_r
every other part is the same per PA on both sides
pitchers K&BB_L = K&BB - (1 - share_L) x gap_r x s(h), K&BB_R = K&BB + share_L x gap_r x s(h)
contact_L = contact - (1 - share_L) x gap_r x (1 - s(h)), contact_R = contact + share_L x gap_r x (1 - s(h))
running game and fielding the same on both sides
For pitchers the gap is in strikeouts and walks plus contact, and s(h) is the strikeouts-and-walks share of the typical gap for his throwing hand (Section 9): L 0.279, R 0.522, S 0.500. A hand gets an estimated share only with 1,000+ batters faced on each side over 2023–2026 and a typical gap of 1+ run per 600; otherwise its share is 0.5, an even split. Switch pitchers (2 pitchers, 17 batters faced from the left and 36 from the right) fall under that rule, so s(S) = 0.5; the 2027 forecast's switch pitcher, Jurrangelo Cijntje (92 projected batters faced), is split evenly. (An interim version of the fix, the same day, had clamped s(S) at 1.) Each part keeps its overall rate as the share-weighted mean of its two sides. This replaced, on 2026-11-16, a rule that divided the regressed gap by the pitcher's own strikeouts-and-walks share of his overall rate; that share explodes when his two parts have opposite signs and nearly cancel, and 475 of 975 projected pitchers had it outside [0, 1] (Alcantara vs LHB +1,185.0 and −1,195.0 per 600; totals and team wins were right). Now none does, the largest part on either side is 32.6 per 600, and Alcantara reads −9.9 / −0.1 vs LHB and +0.3 / +9.1 vs RHB. The share itself is uncertain: by season it ran 0.41–0.81 for right-handers and 0.10–0.38 for left-handers (2020 aside).
Fitting k_gap (forecast.fit_gap_k): the same grid as k (above), minimizing the squared error of the target season's gap, weighted by the target's harmonic sample 1 / (1/PA_L + 1/PA_R), over players with 50+ PA (batters faced) against each hand in the target and at least one in the history; histories are two seasons. The final value is the mean of the two single-target fits:
| k_gap fit on 2025 (within 0.5%) | k_gap fit on 2026 (within 0.5%) | n | Final k_gap | |
|---|---|---|---|---|
| Hitters | 4,000 (2,500–7,000) | 3,000 (2,000–7,000) | 325 / 334 | 3,500 |
| Pitchers | 1,000 (800–1,500) | 800 (500–1,250) | 428 / 436 | 900 |
In parentheses: every grid value whose error is within 0.5% of the minimum. The error curve is flat over a wide range.
For a regular with 600 PA in each of three seasons and the right-handed hitters' typical share against left-handed pitchers, n = 1,532, so he keeps 30% of his own deviation from the typical gap. The share prior of 200 is a constant, not fitted. Tested on the 2026 target (players with 100+ PA or batters faced and a split history), it is near the best of the values tried; the best was 400 for hitters and 400 for pitchers:
| Share prior (PA) | Hitters: share error (points) | Pitchers: share error (points) |
|---|---|---|
| 0 | 8.23 | 8.08 |
| 50 | 8.02 | 7.72 |
| 100 | 7.94 | 7.60 |
| 200 | 7.86 | 7.51 |
| 400 | 7.83 | 7.49 |
| 800 | 7.88 | 7.57 |
| 1,600 | 8.00 | 7.76 |
| typical share only | 9.34 | 8.82 |
Root mean squared error, weighted by the target's playing time, of the predicted against the actual share of 2026 plate appearances (batters faced) against left-handed opponents, in percentage points.
Does regressing the gap work? The paper had no split test. Predicting each player's 2026 gap from 2024–25 (the forecast's test settings), weighted by the target's harmonic sample (334 hitters and 436 pitchers with 50+ against each hand):
| k_gap | Hitters: gap RMSE | Pitchers: gap RMSE | Hitters, typical gap without 2026 | Pitchers, typical gap without 2026 |
|---|---|---|---|---|
| 0 (his own gap) | 42.19 | 32.03 | 42.19 | 32.03 |
| 500 | 31.04 | 22.99 | 31.07 | 23.03 |
| 900 | – | 22.94 | – | 22.99 |
| 1,000 | 30.27 | 22.96 | 30.30 | 23.00 |
| 3,500 | 29.88 | – | 29.93 | – |
| 4,000 | 29.89 | – | 29.94 | – |
| 10,000 | 29.99 | 23.44 | 30.04 | 23.50 |
| typical gap only | 30.15 | 23.59 | 30.21 | 23.64 |
Runs per 600. 'His own gap' is k = 0 (no regression); 'typical gap only' ignores the player's history. The forecast's test k_gap is 4,000 for hitters and 1,000 for pitchers. The last two columns build the typical gap from 2023-2025 only (the forecast pools 2023-2026, which includes the target).
Regressing matters enormously: the player's own gap misses by 42.2 runs per 600 (hitters) against 29.9 regressed. But the player's own history adds little beyond the typical gap for his hand: the fitted k beats the typical gap alone by +0.27 [−0.11, +0.65] for hitters (paired bootstrap) and +0.63 [+0.16, +1.08] for pitchers. The typical gaps and shares are pooled over 2023–2026, so they include the test season; building them without it changes the hitters' gap error from 29.89 to 29.94. They do not touch the overall test below, which scores totals.
Out-of-sample test
Settings fitted on the 2025 target, 2024–25 histories predicting 2026, rates scored at each player's actual 2026 playing time (RMSE per 600 weighted by that playing time, r unweighted):
| Players | r [95%] | RMSE per 600 | Miss minus the forecast's [95%] | |
|---|---|---|---|---|
| Hitters (300+ PA): forecast | 248 | 0.401 [0.29, 0.50] | 16.91 | |
| Hitters: no aging | 248 | 0.373 [0.26, 0.48] | 17.32 | +0.41 [+0.05, +0.76] |
| Hitters: last year's rate | 248 | 0.212 [0.09, 0.33] | 27.59 | +10.68 [+5.67, +18.01] |
| Pitchers (250+ BF): forecast | 250 | 0.480 [0.38, 0.57] | 13.05 | |
| Pitchers: without pitch quality | 250 | 0.452 [0.35, 0.55] | 13.31 | +0.26 [+0.08, +0.43] |
| Pitchers: no aging | 250 | 0.470 [0.37, 0.56] | 13.21 | +0.16 [−0.09, +0.41] |
| Pitchers: last year's rate | 250 | 0.338 [0.22, 0.44] | 17.02 | +3.97 [+2.73, +5.23] |
Intervals for r: Fisher z. Differences: paired bootstrap of players, 4,000 resamples. Last year's rate is the 2025 rate with k = 0, no aging and no pitch quality.
What this test is and is not:
- Not every choice is out of sample. k, k_gap and pitch quality's settings are fitted on 2025, but each component's aging shape is chosen by its error on 2026 itself (Section 13), so the forecast row includes a selection on the test season. How much that selection is worth depends on what it is compared with. Against one fixed shape for every component (the 2015–25 line, 16.913 for hitters and 13.051 for pitchers) the per-component pick gains only 0.000 and 0.004 runs per 600. But the pick also chooses among no aging and the 2023–25 line, and against those the choice made on 2026 is worth 0.41 and 0.21 for hitters (0.16 and 0.10 for pitchers). Made out of sample, choosing each component's shape on 2025 (shapes fitted on pairs ending by 2024) and scoring the same shape types on 2026, the pick would have been, for hitters, batting line 2023–25, baserunning no aging, fielding no aging; for pitchers, strikeouts and walks by age 2015–25, contact quality no aging; scoring 17.04 and 13.08: +0.13 [−0.04, +0.29] and +0.04 [−0.05, +0.13] against the forecast's row. So about 0.13 of the hitters' 0.41 aging gain is the in-sample choice. The typical gaps and shares are pooled over 2023–2026 and so include 2026, but they move only splits, not the totals scored here.
- It is the uncentered forecast. League centering (Section 15) shifts part-timers' projections; on the same target its back-test puts the regulars' miss at 17.09 for hitters and 13.34 for pitchers, against 16.93 and 13.13 uncentered (its pool is slightly wider: 249 and 258 regulars).
- It tests rates, not playing time, and not the minor-league translation (Section 14 has its own test).
- It uses two-season histories; the 2027 forecast uses three. Every fit and every other test in this paper uses two, a holdover from when ledgers began in 2023; the third season's weight (3 of 12 for hitters, 1 of 6 for pitchers) is tested only here, on one target. Predicting 2026 from 2023–25:
| History | Hitters RMSE (r) | Pitchers RMSE (r) |
|---|---|---|
| Two seasons (the reported test) | 16.91 (0.401) | 13.05 (0.480) |
| Three seasons, test settings | 17.23 (0.373) | 13.03 (0.483) |
| Three seasons, k refit on three-season histories | 17.10 (0.374) | 13.03 (0.483) |
| Two seasons, no aging | 17.32 (0.373) | 13.21 (0.470) |
| Three seasons, refit k, no aging | 17.55 (0.342) | 13.19 (0.472) |
Same 2026 targets as above. The refit for three seasons uses 2022–24 histories predicting 2025; hitters' k changed for batting 3,000, catcher defense 1,500; pitchers' k, beta and kq came out the same as the two-season fit.
The result is inconclusive. For hitters the point estimates are worse with the third season: +0.19 runs per 600 with k refit ([−0.11, +0.51], paired bootstrap) and +0.31 [0.00, +0.64] at the test settings, and also without aging (17.55 against 17.32); for pitchers −0.02 [−0.11, +0.07]. Every interval includes zero, so one target neither supports nor rules out the third season's weight. What it does show is that the reported test is not the configuration that produced the 2027 forecast. Ledgers exist from 2015, so the same test can be run on several targets; that is the open item.
How predictable a season is
The correlations look modest because single-season player value is noisy. Season-to-season correlation of runs per PA (per batter faced) for the same players, 300+ PA (250+ batters faced) in both seasons:
| Seasons | Hitters | Total (hitters) | Batting | Situational | Baserunning | Fielding | Pitchers | Total (pitchers) | K&BB | Contact |
|---|---|---|---|---|---|---|---|---|---|---|
| 2015–2016 | 210 | 0.47 | 0.49 | 0.22 | 0.55 | 0.54 | 181 | 0.51 | 0.65 | 0.26 |
| 2016–2017 | 209 | 0.49 | 0.42 | 0.25 | 0.64 | 0.52 | 185 | 0.53 | 0.67 | 0.31 |
| 2017–2018 | 219 | 0.54 | 0.53 | 0.33 | 0.49 | 0.48 | 175 | 0.35 | 0.60 | 0.17 |
| 2018–2019 | 187 | 0.61 | 0.60 | 0.23 | 0.52 | 0.41 | 178 | 0.49 | 0.66 | 0.24 |
| 2021–2022 | 192 | 0.41 | 0.50 | 0.20 | 0.55 | 0.38 | 167 | 0.45 | 0.63 | 0.37 |
| 2022–2023 | 202 | 0.50 | 0.56 | 0.28 | 0.43 | 0.46 | 161 | 0.35 | 0.50 | 0.35 |
| 2023–2024 | 207 | 0.44 | 0.51 | 0.24 | 0.47 | 0.49 | 165 | 0.45 | 0.57 | 0.35 |
| 2024–2025 | 199 | 0.53 | 0.55 | 0.23 | 0.45 | 0.48 | 163 | 0.45 | 0.66 | 0.26 |
| 2025–2026 | 194 | 0.40 | 0.41 | 0.33 | 0.44 | 0.43 | 172 | 0.46 | 0.60 | 0.33 |
Pearson r. Fielding is the DRS part (catcher defense apart). Pairs touching 2020 are left out. 95% Fisher intervals for 2025–2026: hitters' total [0.27, 0.51], batting [0.29, 0.52], situational [0.20, 0.45]; pitchers' total [0.33, 0.57], strikeouts and walks [0.49, 0.68], contact [0.18, 0.45].
The pair the earlier text quoted, 2025–26 (hitters 0.40), is the least repeatable of the 9 pairs for hitters: hitters' totals run 0.40 to 0.61 across them, pitchers' 0.35 to 0.53. An earlier version said that "a forecast cannot correlate with next season much better than next season correlates with itself". That is wrong. If a season's value is skill plus independent noise, its year-to-year correlation estimates the reliability ρ, and the best any forecast can do is correlate with the true skill, which caps its correlation with next season at √ρ, not ρ: √0.40 = 0.63 for hitters and √0.46 = 0.68 for pitchers on this pair. The forecast's own correlations (0.401 for hitters, 0.480 for pitchers) sit just above this pair's single-season correlations (0.395 and 0.461), by less than this pair's sampling error and on different samples (248 and 250 players with 300+ PA or 250+ batters faced in 2026, against 194 and 172 with that playing time in both seasons). Against the other pairs the forecast's r is lower than the year-to-year r in 8 of 9 hitter pairs and 3 of 9 pitcher pairs, so the forecast does not beat a single season's repeatability in general; 2025–26 was simply an unusually unrepeatable pair for hitters.
How the test evolved
| Version | Ages | Aging shapes | Pitch-quality prior | Hitters RMSE (r) | Pitchers RMSE (r) |
|---|---|---|---|---|---|
| First build, 2026-11-14 (README) | July 1 | line on 2023–25 pairs | no | 17.1 (0.39) | 13.4 (0.44) |
| The same configuration, recomputed on today's ledgers | July 1 | line on 2023–25 pairs | no | 17.08 (0.389) | 13.42 (0.445) |
| Aging refit on 2015–25 pairs, five shapes, per-part pick | July 1 | batting curve 2015–25, baserunning line 2015–25, fielding line 2015–25, strikeouts and walks curve 2015–25, contact quality curve 2015–25 | no | 16.88 (0.404) | 13.29 (0.453) |
| Pitch quality added | July 1 | strikeouts and walks curve 2015–25, contact quality curve 2015–25 | yes | same | 13.03 (0.481) |
| Opening Day ages (current) | Opening Day | batting line 2015–25, baserunning line 2015–25, fielding line 2015–25, strikeouts and walks curve 2015–25, contact quality line 2015–25 | yes | 16.91 (0.401) | 13.05 (0.480) |
The 2026 test at each stage. The first row is as reported then; the others re-run each configuration on today's ledgers (paper_checks_forecast.py), so they differ from the README's figures only where the ledgers or the README's rounding do.
13.Aging curves
The delta method
Aging is measured from the same players in consecutive seasons (forecast.fit_aging). For every pair of seasons (y, y+1) and every player with at least 200 units (PA or batters faced) in both, the change in each component's rate is recorded against his age on Opening Day of the second season (an integer, the age shown throughout the dashboard), weighted by half the harmonic mean of his two seasons' playing time:
delta = rate(y+1) - rate(y) per component, per PA or BF weight = 1 / (1/units(y) + 1/units(y+1)) age = age on Opening Day of season y+1 polynomial fits keep 20 <= age <= 41
The components aged are batting, baserunning and fielding for hitters and strikeouts and walks and contact quality for pitchers. Situational outs, catcher defense, the positional adjustment, the running game and pitchers' fielding are not aged; that choice has not been tested. Pairs touching 2020 are skipped, and no pair spans the gap (2019 to 2021). It is not that nobody reaches 200 units in a 60-game season: 124 hitters and 90 pitchers did in 2020 (2019: 360 and 366). The recorded reason is only that 2020 was a 60-game season; two concrete problems with it are that each 2020 rate rests on 60 of 162 games, and that its July 23 Opening Day adds a year to the age of every player born between late March and July 23, so a 2019–20 change would be indexed at the wrong age. Including the 2020 pairs has not been tested. The final curves use 9 season pairs, 2015-2016 through 2025-2026:
| Pairs fitted for the test | Player-pairs (polynomials) | Player-pairs (by age) | Weighted mean age | Final pairs | Final player-pairs | Final weighted mean age | Raw level per year (test pairs), per 600 | |
|---|---|---|---|---|---|---|---|---|
| Hitters | 8 (2015-2016 to 2024-2025) | 2,139 | 2,142 | 28.72 | 9 | 2,402 | 28.677 | batting −2.39, baserunning −0.16, fielding −0.42 |
| Pitchers | 8 (2015-2016 to 2024-2025) | 1,880 | 1,888 | 29.12 | 9 | 2,097 | 29.185 | strikeouts and walks −0.80, contact quality −0.75 |
Player-pairs: players with 200+ units in both seasons and a birth date; the polynomial fits also require ages 20-41. Raw level: the weighted mean change per year, the level removed below.
Removing the level
Same-player changes mix real aging with two biases: players who keep their jobs after a good season tend to fall back toward average (regression to the mean), and players who decline sharply lose their jobs and leave the sample (survivorship). The weighted mean change is negative for every component (batting −2.39, baserunning −0.16, fielding −0.42; pitchers strikeouts and walks −0.80, contact quality −0.75 runs per 600 per year). The forecast already pulls every player toward average through k, so applying the raw level as well would count regression twice. Each fitted curve therefore keeps only its shape:
fit delta = a + b x (age - 28) [+ c x (age - 28)^2] weighted least squares then a <- a - sum(weight x fitted delta) / sum(weight) the weighted mean fitted change is set to 0
| Shape | Hitters: level removed | Hitters: level kept | Pitchers: level removed | Pitchers: level kept |
|---|---|---|---|---|
| no aging | 17.32 | 17.32 | 13.21 | 13.21 |
| line 2023–25 | 17.12 | 17.66 | 13.15 | 13.18 |
| line 2015–25 | 16.91 | 17.36 | 13.05 | 13.09 |
| curve 2015–25 | 16.90 | 17.33 | 13.05 | 13.08 |
| by age 2015–25 | 16.93 | 17.39 | 13.05 | 13.08 |
| per-part best | 16.91 | 17.31 | 13.05 | 13.08 |
2026 test RMSE per 600 (AGING=absolute keeps the level). With the level kept the hitters' per-part pick becomes batting no aging, baserunning line 2015–25, fielding no aging.
Keeping the level makes every shape worse than no aging at all for hitters. On the original configuration (July 1 ages, a line on the two pairs 2023–24 and 2024–25) it scored 17.61 against 17.08 removed (reported at the time as 17.6 against 17.1; an earlier version of this paragraph called that the "three-pair fit", but the test fit used two pairs).
What removing the level costs: the peak age is not estimated. Setting the weighted mean fitted change to zero forces every straight line to cross zero exactly at the weighted mean age of its sample, whatever the data say: 28.68 for every hitter component and 29.19 for pitchers' contact (the table of final curves shows the crossings). An earlier version read "hitting peaks around 28–29 ... speed and defense decline steadily from the late 20s" off these crossings; they are a property of the method. The curves identify how much faster a young player improves than an old one, not the age at which improvement stops. That is all the forecast needs, since k and league centering set the level.
Candidate shapes and the per-component pick
Five shapes were fitted on the 8 pairs ending by 2025:
- no aging;
- a straight line in age on the 2023–25 pairs only (the original model);
- a straight line on 2015–25;
- a quadratic curve on 2015–25 (
a + b(age - 28) + c(age - 28)^2, weighted least squares); - an age-by-age average: the weighted mean change at each age, ages below 22 and above 38 pooled into the end bins, smoothed over the two ages either side with 1-2-3-2-1 weights (missing ages skipped), the level removed the same way; evaluated by linear interpolation between ages.
Each component then takes the shape that best predicts that component in 2026, the test season itself (RMSE of the component's rate per 600, pitchers with the pitch-quality prior). Earlier versions said the 2026 test "stays out of sample"; the shapes are fitted without 2026, but the choice among them is made on it (Section 12 measures what that choice is worth):
| Component | None | Line 2023-25 | Line 2015-25 | Curve 2015-25 | By age 2015-25 | Chosen | Margin to the next best | Line minus curve [95%] |
|---|---|---|---|---|---|---|---|---|
| Batting | 14.2048 | 13.9468 | 13.8705 | 13.8709 | 13.8868 | line 2015–25 | 0.0004 | −0.0004 [−0.0029, +0.0021] |
| Baserunning | 2.4737 | 2.4906 | 2.4612 | 2.4615 | 2.4620 | line 2015–25 | 0.0003 | −0.0003 [−0.0007, +0.0001] |
| Fielding | 7.6054 | 7.6318 | 7.6009 | 7.6052 | 7.6279 | line 2015–25 | 0.0043 | −0.0043 [−0.0544, +0.0464] |
| Pitchers: strikeouts and walks | 8.1454 | 8.0546 | 8.0080 | 8.0059 | 8.0274 | curve 2015–25 | 0.0020 | +0.0020 [−0.0525, +0.0551] |
| Pitchers: contact quality | 8.5543 | 8.5745 | 8.5256 | 8.5257 | 8.5484 | line 2015–25 | 0.0001 | −0.0001 [−0.0056, +0.0056] |
Component RMSE per 600 on the 2026 target. Margin: the next-best shape's RMSE minus the chosen one's. Last column: paired bootstrap of players, 4,000 resamples.
The margins are in the third or fourth decimal and every line-against-curve interval spans zero: the pick is noise. It has flipped already: with July 1 ages the picks were batting curve 2015–25, baserunning line 2015–25, fielding line 2015–25, strikeouts and walks curve 2015–25, contact quality curve 2015–25; with Opening Day ages batting went to line 2015–25 and contact quality went to line 2015–25. Choosing per component does matter for one thing, plausibility: a single quadratic for every component makes fielding U-shaped, with old players "improving":
| Quadratic, per 600 per year | 22 | 24 | 26 | 28 | 30 | 32 | 34 | 36 | 38 | 40 |
|---|---|---|---|---|---|---|---|---|---|---|
| Fielding, fit on pairs to 2025 | +1.1 | +0.5 | +0.1 | −0.2 | −0.3 | −0.3 | −0.1 | +0.2 | +0.7 | +1.3 |
| Fielding, final | +1.3 | +0.6 | +0.1 | −0.2 | −0.4 | −0.3 | −0.1 | +0.3 | +0.8 | +1.5 |
| Batting, fit on pairs to 2025 | +3.4 | +2.4 | +1.4 | +0.4 | −0.6 | −1.7 | −2.7 | −3.8 | −4.8 | −5.9 |
| Batting, final | +3.9 | +2.7 | +1.5 | +0.4 | −0.8 | −1.9 | −3.0 | −4.0 | −5.1 | −6.1 |
| Strikeouts and walks, fit on pairs to 2025 | +2.9 | +1.9 | +0.9 | +0.2 | −0.5 | −1.0 | −1.4 | −1.6 | −1.7 | −1.7 |
| Strikeouts and walks, final | +2.8 | +1.8 | +1.0 | +0.3 | −0.4 | −1.0 | −1.4 | −1.8 | −2.1 | −2.3 |
The quadratic shape for fielding, batting and pitchers' strikeouts and walks, fitted on the pairs ending by 2025 and on all pairs.
The likely reason is a selection effect, not aging: aging players who field poorly move to designated hitter, so their fielding runs per PA rise toward zero. It has not been tested (a refit on players with fielding innings in both seasons would). The line does not bend back, and for fielding the two shapes predict 2026 equally well. Total-value back-test for each choice, with each shape's gain over no aging:
| Aging | Hitters RMSE | Hitters r | Hitters: minus no aging [95%] | Pitchers RMSE | Pitchers r | Pitchers: minus no aging [95%] |
|---|---|---|---|---|---|---|
| no aging | 17.32 | 0.373 | 13.21 | 0.470 | ||
| line 2023–25 | 17.12 | 0.386 | −0.20 [−0.32, −0.08] | 13.15 | 0.475 | −0.06 [−0.12, 0.00] |
| line 2015–25 | 16.91 | 0.401 | −0.41 [−0.76, −0.05] | 13.05 | 0.480 | −0.16 [−0.38, +0.07] |
| curve 2015–25 | 16.90 | 0.402 | −0.42 [−0.80, −0.03] | 13.05 | 0.480 | −0.16 [−0.41, +0.09] |
| by age 2015–25 | 16.93 | 0.400 | −0.39 [−0.74, −0.04] | 13.05 | 0.481 | −0.16 [−0.39, +0.08] |
| per-part best | 16.91 | 0.401 | −0.41 [−0.76, −0.05] | 13.05 | 0.480 | −0.16 [−0.41, +0.09] |
2026 test, 248 hitters and 250 pitchers; paired bootstrap of players.
Gains below are no aging's RMSE minus the shape's (positive = aging helps). For hitters every 2015–25 shape beats no aging with an interval that excludes zero (the forecast's choice: +0.41 [+0.05, +0.76], better in 98.8% of resamples). For pitchers the gain is +0.16 [−0.09, +0.41], and the interval includes zero (89.9%), so the pitcher curves are not established by this test. These gains include the choice of shape made on 2026. Chosen out of sample instead (on 2025, Section 12), the hitters' gain over no aging is +0.28 [+0.09, +0.46] and the pitchers' +0.12 [−0.04, +0.29]. By component (component RMSE) the hitters' gain is almost all batting (+0.33 [+0.01, +0.65]); baserunning +0.013 [−0.003, +0.028] and fielding +0.005 [−0.026, +0.036] are tiny. Aging moves a typical player by a few runs per 600, while a season's noise is about 17 runs per 600 for hitters.
The fitted curves
Final curves, the chosen shape of each component refit on all 9 pairs through 2025-2026. Values are the change per year of age, in runs per 600 PA (hitters) or batters faced (pitchers), at each age; the last column is where the curve crosses zero:
| Component (shape) | 22 | 24 | 26 | 28 | 30 | 32 | 34 | 36 | 38 | 40 | Zero at |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Batting (line) | +3.8 | +2.6 | +1.5 | +0.4 | −0.8 | −1.9 | −3.0 | −4.1 | −5.3 | −6.4 | 28.68 |
| Baserunning (line) | +0.2 | +0.1 | +0.1 | 0.0 | 0.0 | −0.1 | −0.1 | −0.2 | −0.3 | −0.3 | 28.68 |
| Fielding (line) | +0.4 | +0.3 | +0.2 | 0.0 | −0.1 | −0.2 | −0.3 | −0.5 | −0.6 | −0.7 | 28.68 |
| Pitchers: strikeouts and walks (curve) | +2.8 | +1.9 | +1.0 | +0.3 | −0.4 | −1.0 | −1.4 | −1.8 | −2.1 | −2.3 | 28.75 |
| Pitchers: contact quality (line) | +0.8 | +0.6 | +0.4 | +0.1 | −0.1 | −0.3 | −0.6 | −0.8 | −1.0 | −1.3 | 29.19 |
Read as slopes, not levels: batting: the change per year falls by 0.564 runs per 600 for each year of age (a straight line); baserunning: the change per year falls by 0.029 runs per 600 for each year of age (a straight line); fielding: the change per year falls by 0.062 runs per 600 for each year of age (a straight line); pitchers' strikeouts and walks: a curve whose decline deepens more slowly with age (quadratic term +0.0115 per 600; the change per year bottoms out at age 43, beyond the data); pitchers' contact quality: the change per year falls by 0.118 runs per 600 for each year of age (a straight line). A player ten years older gets a per-year change that much lower; where the curve crosses zero is the sample's mean age, as explained above.
Applying the curves
A player's weighted history sits about two years before the forecast season. Each component is aged over that gap, evaluated at the midpoint age:
years = 2027 - sum_y (w_y x units_y x y) / sum_y (w_y x units_y) mid_age = clamp(age on Opening Day 2027 - years / 2, 20, 40) rate += delta_c(mid_age) x years
With equal playing time in each season, years = 1.83 for hitters (5/4/3) and 1.67 for pitchers (3/2/1); a 2026 rookie with one season has a gap of 1 year; a player back from a lost season (2024 and 2025 only) 2.43 (2.33 for pitchers). For a player with minor-league time the translation (Section 14) then scales the translated parts of his MLB rate, aging included, by (U + k) / (U + K), so their aging shrinks with his minor-league sample: batting, situational outs and baserunning for hitters and strikeouts and walks, contact quality and running game for pitchers. The other aged part, hitters' fielding, is not translated and keeps its full aging. Examples, hitters with equal playing time in 2024–2026, batting:
| Age in 2027 | Mid age | Batting change per year | Batting total | Batting total, half-year indexing | Baserunning + fielding total |
|---|---|---|---|---|---|
| 23 | 22.08 | +3.72 | +6.82 | +6.30 | +1.11 |
| 26 | 25.08 | +2.03 | +3.72 | +3.20 | +0.60 |
| 29 | 28.08 | +0.33 | +0.61 | +0.10 | +0.10 |
| 32 | 31.08 | −1.36 | −2.49 | −3.01 | −0.40 |
| 35 | 34.08 | −3.05 | −5.59 | −6.11 | −0.91 |
| 37 | 36.08 | −4.18 | −7.66 | −8.18 | −1.24 |
Runs per 600 PA. 'Half-year indexing': the same with the curve evaluated half a year older (below).
A 23-year-old with three full seasons is evaluated at 22.08, not 23: +3.7 runs per 600 per year of the gap, +6.8 in all; a 37-year-old at 36.08: −4.2 per year, −7.7 in all. (An earlier version quoted the curve at 23 and 37, +3 and −5, as the per-year change.) Pitchers' strikeouts and walks, the same ages: +4.5 and −3.1 in all.
A half-year indexing question. A delta is indexed by the age in the second season of its pair, so the change from a season played at age A − 1 to one at A is delta(A). Summing those changes over a gap of g years up to age A points to evaluating the curve at about A − g/2 + 1/2; the code uses A − g/2, half a year younger. Because the curves slope down, that adds a little to every player: on the 2027 projections the corrected indexing would move depth-chart hitters by −0.54 runs per 600 on average (weighted by projected PA, 589 players, at most −0.98) and pitchers by −0.34 (717). Nearly uniform shifts like these are largely absorbed by league centering (Section 15). On the 2026 test the corrected indexing scores 16.913 for hitters (+0.001 [−0.054, +0.053] against the current) and 13.033 for pitchers (−0.013 [−0.050, +0.023]): no measurable difference. It is recorded, not changed.
History
- The original model fitted a straight line on the two pairs then available (2023–24, 2024–25) with July 1 ages: batting +0.27 − 0.35 × (age − 28) runs per 600 (recomputed on today's ledgers; the README recorded +0.27 − 0.35 × (age − 28)). The pitchers' first version aged strikeouts and walks and contact together; baserunning and fielding slopes were about zero on so few pairs. Its 2026 test: 17.08 hitters, 13.42 pitchers.
- The refit on 2015–26 (2026-11-14) added the five shapes and the per-component pick: 16.88 and 13.29 (before pitch quality), with the picks above.
- Opening Day ages (2026-11-15) replaced July 1 everywhere (one age convention for the dashboard, the aging fits, the projection and the translation). The test moved from 16.88 to 16.91 for hitters and from 13.03 to 13.05 for pitchers (with pitch quality), and batting and contact quality changed shape, within noise.
14.Minor leaguers: Triple-A and Double-A translations
Through the 2026 season the forecast projected only players who had appeared in MLB; everyone else on a 2027 depth chart was a flat newcomer allowance. This section adds Triple-A and Double-A players. It builds the same model in the minors and measures, from the players who made the jump, how much of each piece of minor-league value shows up in MLB the next season. A minor-league line is never taken at face value: it passes through a level gap, a partial carry-over for each piece, a pull toward average for small samples, and an age adjustment. In the pooled 2025–26 back-test the translation lowers hitters' typical miss from 17.65 to 17.26 runs per 600 (−0.39 [−0.77, −0.02], better in 98.0% of 2,000 resamples); for pitchers it ranks players a little better but does not lower the typical miss (14.22 against 14.24). Both comparisons have defects of their own, measured under "What the back-test does not test".
Data
The Stats API serves minor-league play-by-play in the same shape as MLB (sports 11–14 are Triple-A, Double-A, High-A and Single-A), so the same reconstruction runs unchanged: base-out states, runner movements, batted-ball type and handedness (fetch_pbp.py YEAR LEVEL, pa_table.py YEAR LEVEL, LEVEL = aaa, aa, hia or a; about 2,200 games a season at Triple-A). Each plate appearance also keeps its park (venue), because affiliates sometimes play home games elsewhere; data/milb/teams_LEVEL_YEAR.json gives each club's league and parent organization. pa_table's self-checks (base state against the feed's runners, outs, runs against the final score) pass with a handful of base-state resyncs per 150,000 plate appearances and two score mismatches in 2021–26, one at Triple-A (2022) and one at High-A (2025); Single-A and Double-A reconcile in every game. MLB reconciles with zero. These counts are printed by pa_table.py when it runs and recorded in README.md; no output file stores them. Seasons start in 2021, when the minors were reorganized (there was no 2020 season). Exit velocity and launch angle exist only where the park has Hawk-Eye: every Triple-A park from 2023, some in 2022, none in 2021, none at Double-A or High-A, and the Florida State League's parks at Single-A. Every level has the batted-ball type and runner movements. FanGraphs' minor-league leaderboard sits behind a Cloudflare challenge, so there is no outside check of minor-league batting; the test that matters is the MLB back-test below. All four levels are built; the forecast uses Triple-A and Double-A (see "High-A and Single-A" below). Mean hitter age is 26.0–26.3 at Triple-A and 23.4–23.8 at Double-A by season.
| Season | AAA PA | AA PA | A+ PA | A PA | AAA exit velocity | A exit velocity |
|---|---|---|---|---|---|---|
| 2021 | 145,997 | 129,504 | 133,870 | 135,845 | 0% | 27% |
| 2022 | 171,720 | 156,510 | 146,833 | 149,397 | 38% | 28% |
| 2023 | 174,963 | 156,701 | 147,812 | 148,834 | 100% | 27% |
| 2024 | 172,462 | 152,818 | 147,055 | 148,374 | 99% | 29% |
| 2025 | 171,483 | 152,649 | 147,225 | 149,076 | 99% | 29% |
| 2026 | 171,799 | 156,270 | 151,429 | 151,705 | 99% | 29% |
Exit velocity = share of batted balls with exit velocity and launch angle. Double-A and High-A have none (Double-A's best season: 0.05%); Single-A has it only in the Florida State League's parks. In 2021 the leagues had interim names (Triple-A East / West, Double-A Northeast / Central / South, High-A East / Central / West, Low-A East / Southeast / West).
The minor-league model
Every number is measured against the player's own league in that season (International or Pacific Coast League; Eastern, Southern or Texas League), so the environments drop out: Triple-A's automated strike zone, the bigger bases and pickoff limits, the high-altitude Pacific Coast parks. Events are valued at the same season's MLB linear weights (LW1 less the MLB mean, Section 4), so a walk is worth the same in every league and only the player's mix of events against his league's mix counts. Batting is split into the pieces that carry over differently. For a plate appearance with result c (MLB value v_c; b_avg = MLB's average value of a ball in play that season):
K = v_K - b_avg strikeouts, valued against a ball in play BB = v_c - b_avg walks, HBP, intentional walks, interference TYPE = x_traj - b_avg batted-ball mix: MLB's mean value for the trajectory QUAL = x_evla - x_traj exit velocity x launch angle cell value beyond it RES = v_c - x_best results beyond the expected value K + BB + TYPE + QUAL + RES = the MLB model's batting value (before league centering)
The trajectory and exit-velocity values come from MLB's own batted balls that season (milb_ledger.mlb_currency). A ball is looked up from the most to the least specific cell: exit velocity x launch angle in 3 mph x 4° bins, then 6 mph x 8°, both kept separate for bunts; then the recorded trajectory; then all balls. A cell is used only if it holds 25+ MLB balls and more balls than home runs, and its value is the home-run probability split used throughout the model (Section 8): p_HR x v_HR + (1 − p_HR) x (mean value of its non-home-run balls). x_best is the finest exit-velocity cell available, else the trajectory value; without exit velocity QUAL is zero and RES = v_c − x_traj. Each walk type takes its own MLB value. One trap: the cells are keyed on the recorded trajectory, never on the result. The MLB contact model separates bunt hits from bunt outs by result, which is harmless when describing a season but would leak the outcome into a predictor.
Each batting piece is then centered on its league (the league mean per plate appearance subtracted, the league being the home club's) and park-adjusted by venue, additively, and centered again:
e(v, j) = [ mean of x_j over PA at venue v - sum_h (n_vh / n_v) x mean of x_j over club h's road PA ]
x n_v / (n_v + 10,000)
x_j <- x_j - e(venue, j), then each league re-centered on zero
n_v = PA at v over this season and the season before (both centered first); h = the clubs hosting at v
With 10,000 PA a two-season venue is regressed about halfway. 2021 has no season before, so its venue effects rest on one season. Only the five batting pieces (and their pitcher mirrors) are park-adjusted; situational outs, baserunning and the running game are not. Situational outs and baserunning use exactly the MLB method (Section 6) with the league's own run expectancy; the code is copied into milb_ledger.league_running so the validated ledger.py stays untouched (a trap: gap_decomp.decompose() rewrites rows in place, so league_running is given copies). Pitchers carry the same pieces with the sign flipped (strikeouts / walks = PK + PBB; contact = PTYPE + PQUAL; results PRES kept apart, as in MLB) and a running-game term PRUN (steals, pickoffs, wild pitches, balks). There is no minor-league fielding: the minors have no DRS. Each player record also carries his pieces summed over his odd and even plate appearances, for the reliability below.
Triple-A pitch quality
Triple-A is pitch-tracked by the same Hawk-Eye system as MLB in every park from 2023 and in about a third of 2022's pitches; Double-A and High-A have none, and Single-A only in the Florida State League (24 sampled games per level-season settled the coverage; README.md). fetch_pitches.py YEARS aaa pulls every Triple-A pitch (data/pitches/milb/), and pitch_quality.py scores scores each Triple-A season S with the MLB stuff models (Section 8) fitted on every MLB pitch season up to and including S, so the score reads on the MLB scale (what the pitch would be worth in MLB) and no model has seen a later MLB season. For each pitcher-season the predicted values of his valid-stuff pitches are summed (strikeouts / walks and contact separately, in pitcher runs), and his tracked batters faced are TBF x (tracked pitches / pitches). The two sums, less the level-season mean per tracked batter faced times his tracked units, become two more pieces: PQK in strikeouts / walks and PQC in contact. An untracked pitcher therefore counts as average. The same sums over his odd and even plate appearances give the split halves.
| Season | Tracked batters faced | Triple-A mean: K&BB / 600 | Triple-A mean: contact / 600 |
|---|---|---|---|
| 2022 | 63,990 | −5.75 | −3.32 |
| 2023 | 171,807 | −4.45 | −2.18 |
| 2024 | 170,260 | −4.03 | −2.09 |
| 2025 | 168,854 | −2.98 | −1.65 |
| 2026 | 168,809 | −3.53 | −1.85 |
The average Triple-A pitcher's pitch quality per tracked batter faced, on the MLB scale (against an average MLB pitch). Each season's mean is subtracted before the pieces enter the translation, so the level gap stays in a_L. A level-season needs 5,000+ tracked batters faced to be centered and used.
The pieces are shrunk like the others, by their own split-half reliability on tracked units: x~ = (weighted sum / weighted tracked units) x n / (n + k), n his raw tracked batters faced. They are the steadiest pieces in the minors (odd vs even plate appearances r = 0.984 and 0.954, 1,318 pitcher-seasons), because a pitcher's velocity and movement barely change from one outing to the next and the score carries no outcome noise. The formula would give k = 2.6 and 7.8; the clamp sets both to 20, so pitch quality is effectively unshrunk. Its slopes are b = 0.42 and 0.45 per shrunk run. The slopes of the other contact pieces are larger (batted-ball mix 0.62, contact quality 0.83), but those multiply heavily shrunk rates (k = 1,000 and 1,702); per full Triple-A season (450 batters faced) pitch quality carries 0.43 of a run of contact against 0.19, 0.17 and 0.01 for batted-ball mix, contact quality and results, and 0.40 of strikeouts and walks. In the back-test it lowers the pooled typical miss by −0.051 runs per 600 (95% interval [−0.107, +0.001]; better in 97.0% of resamples) and by −0.20 [−0.50, +0.13] for pitchers with no MLB time. The gain is not uniform: mostly minors in 2025 got slightly worse (14.09 → 14.12); minors only in 2026 missed less but ranked worse (r 0.07 → 0.04); pooled mostly minors ranked slightly worse (r 0.29 → 0.28). It was kept because the pooled gain has the right sign in every group and the contact piece carries more than twice as much per full season as any other contact piece; the pooled interval only just reaches zero. It is not split by batter hand (the MLB hand split of Section 8 has no Triple-A counterpart). MILB_PQ=0 turns it off; milb_pq_test.py runs both fits and compares them.
| Pitchers | n | Without pitch quality: RMSE (r) | With: RMSE (r) | With minus without (95% CI) |
|---|---|---|---|---|
| all | 928 | 14.29 (0.37) | 14.24 (0.38) | −0.05 [−0.11, 0.00], better in 97.0% of 2,000 resamples |
| minors only | 105 | 15.85 (0.11) | 15.65 (0.17) | −0.20 [−0.50, +0.13], better in 88.8% of 2,000 resamples |
| mostly minors | 281 | 15.89 (0.29) | 15.82 (0.28) | −0.07 [−0.22, +0.09], better in 79.2% of 2,000 resamples |
| mostly MLB | 542 | 13.38 (0.45) | 13.36 (0.45) | −0.02 [−0.06, +0.01], better in 91.3% of 2,000 resamples |
| all, 2025 only | 466 | 13.92 (0.42) | 13.88 (0.42) | |
| minors only, 2025 only | 56 | 15.56 (0.14) | 15.28 (0.26) | |
| mostly minors, 2025 only | 146 | 14.09 (0.31) | 14.12 (0.30) | |
| mostly MLB, 2025 only | 264 | 13.62 (0.49) | 13.59 (0.50) | |
| all, 2026 only | 462 | 14.65 (0.33) | 14.58 (0.33) | |
| minors only, 2026 only | 49 | 16.19 (0.07) | 16.08 (0.04) | |
| mostly minors, 2026 only | 135 | 17.41 (0.26) | 17.27 (0.27) | |
| mostly MLB, 2026 only | 278 | 13.15 (0.41) | 13.13 (0.41) |
The translation's own back-test (each test season fitted on earlier targets), with and without the two pitch-quality pieces; groups as in the back-test table below. Intervals: paired bootstrap per group (2,000 resamples, paper_checks_tail.py). milb_pq_test.py's own bootstrap (4,000 resamples of the pooled 928 pitchers, seed 7) found 'with' better in 96.6%.
How much each piece repeats
A piece that repeats within a season is a skill; one that does not is mostly noise. Each piece is pulled toward its league average by its own split-half reliability before it is translated:
x~ = (weighted rate) x n / (n + k_j) n = minor-league PA (or BF) in the window k_j = h (1 - r) / r r = odd-vs-even PA correlation, h = typical half size
The estimator, exactly: for each level and piece, every player-season with 100+ plate appearances (batters faced) in each half, all seasons 2021–2026 pooled, gives one pair of half rates; r is their Pearson correlation; h is the mean over players of the harmonic half size 2 / (1/n_odd + 1/n_even); a cell needs 30+ player-seasons; k = h (1 − r) / r, set to 20,000 if r ≤ 0.02 and clamped to [20, 20,000]. This is the usual stabilization logic: a mean over n units then has reliability n / (n + k). Contact quality uses only level-seasons with at least 90% exit-velocity coverage; at untracked levels it is folded into results. In x~ the rate is weighted 5/4/3 (hitters) or 3/2/1 (pitchers) over the three seasons before the target, like the MLB history, and n is the raw, unweighted minor-league units at that level.
| Piece | Hitters AAA r (k) | Hitters AA r (k) | Pitchers AAA r (k) | Pitchers AA r (k) |
|---|---|---|---|---|
| strikeouts | 0.73 (64) | 0.76 (53) | 0.59 (110) | 0.60 (104) |
| walks / HBP | 0.56 (139) | 0.56 (138) | 0.54 (133) | 0.56 (123) |
| batted-ball mix | 0.31 (384) | 0.31 (389) | 0.14 (1,000) | 0.17 (747) |
| contact quality | 0.62 (108) | – | 0.09 (1,702) | – |
| results beyond contact | 0.15 (978) | 0.34 (345) | 0.13 (1,100) | 0.09 (1,671) |
| all batting pieces | 0.26 (487) | 0.33 (357) | 0.22 (549) | 0.20 (622) |
| pitch quality: strikeouts / walks | – | – | 0.98 (20) | – |
| pitch quality: contact | – | – | 0.95 (20) | – |
Split-half correlation of a player's odd and even plate appearances (100+ in each half, every season pooled; typical half 174–178 PA for hitters, 158–162 batters faced for pitchers; 1,290–2,214 player-seasons per cell); k in plate appearances or batters faced. Exit-velocity pieces only from seasons with tracking; none at Double-A. k is clamped to [20, 20,000], so the pitch-quality k of 20 is the floor.
Strikeouts settle in about 64 plate appearances at Triple-A (53 at Double-A); results on balls in play need about 978 for Triple-A hitters and 1,100–1,671 for pitchers. A hitter's total batting value over half a Triple-A season repeats at only r = 0.26, which is why the translation works piece by piece. For pitchers, contact quality allowed hardly repeats at all, as in MLB.
The translation
The projection from level L is a level gap plus the carried-over pieces plus an age term; levels are mixed by discounted playing time:
milb_L = a_L + sum_j b_Lj x~_Lj + g_L x (age - 26) milb = sum_L s_L milb_L s_L = d_L U_L / sum(d U) U = weighted units
It enters the forecast as the starting point a player's MLB record is weighed against. R and U_mlb are his MLB runs and units weighted 5/4/3 (pitchers 3/2/1), k is the MLB forecast's own regression constant (Section 12), and m is the minor-league share of his history:
rate = (R + K x m x milb) / (U + K) K = k + sum_L d_L U_L m = sum(d U) / (sum(d U) + U_mlb)
Written out as milb_translate.row_for builds each player's regression row (per part c: batting, situational outs and baserunning for hitters; strikeouts / walks, contact and the running game for pitchers; fielding and catcher defense are not translated):
s = sum_L d_L x U_L U_L = his 5/4/3 (pitchers 3/2/1) weighted units at level L K = k + s , m = s / (s + U_mlb) , f = K x m / (U_mlb + K) rate = R / (U_mlb + K) + f x sum_L (d_L U_L / s) x [ a_L + g_L x age_rel + sum_j scale_L x b_j x x~_Lj ] age_rel = clamp(age on Opening Day of the target season, 19, 33) - 26
With no minor-league time, K = k and m = 0: exactly the MLB forecast, pulled toward average. With no MLB time, f = 1 and it is exactly the translation. A veteran's rehab stint has a tiny m and barely moves him. Given d, the scales and (for the single-piece parts) the minor-league k, the rate is linear in a, b and g, so they come from weighted least squares, weighted by the player's MLB units in the target season (any MLB time counts in the fit; the 100+ minimum applies only to the evaluation), with a ridge of 1e-09 times the largest diagonal element for numerical safety. The target is the player's MLB rate of that part in the target season; for pitchers' contact the minor-league PRES is one of the predictors although it is not part of minor-league contact. The rest is a grid search, scored by the weighted mean squared error of the fit on its own training targets: d_AAA and d_AA over {0.25, 0.5, 0.75, 1, 1.5, 2, 3} with d_AA ≤ d_AAA; the Double-A scale of the shared carry-overs over {0.4, 0.55, 0.7, 0.85, 1}; the single-piece k over {300, 1,000, 3,000, 10,000}. Several fits sit on the grid's edges: hitters batting: Triple-A discount d = 3 (grid top); hitters situational outs: Triple-A discount d = 3 (grid top); hitters situational outs: Double-A discount d = 3 (grid top); hitters situational outs: single-piece k = 300 (grid bottom); hitters baserunning: Triple-A discount d = 3 (grid top); hitters baserunning: Double-A discount d = 3 (grid top); hitters baserunning: single-piece k = 300 (grid bottom); pitchers strikeouts / walks: Double-A scale = 1 (grid top); pitchers contact allowed: Triple-A discount d = 3 (grid top); pitchers contact allowed: Double-A discount d = 0.25 (grid bottom); pitchers contact allowed: Double-A scale = 1 (grid top); pitchers running game: single-piece k = 300 (grid bottom); pitchers bottom line: Triple-A discount d = 0.25 (grid bottom); pitchers bottom line: Double-A discount d = 0.25 (grid bottom). At the edges the grid, not the data, sets the value, so a wider grid could move those parameters. Batting, strikeouts / walks and contact use one carry-over per piece for both levels, with the Double-A piece scaled by the fitted factor; the single-piece parts keep a slope per level. The fit uses 2,698 hitter and 3,407 pitcher player-seasons (targets 2022–2026). The "bottom line" comparison fits one piece per level: the sum of K, BB, TYPE, QUAL, RES, sit, br for hitters and of PK, PBB, PTYPE, PQUAL, PRES, PRUN for pitchers (pitch quality is not in it), shrunk by the split-half k of the total of K, BB, TYPE, QUAL, RES (pitchers: PK, PBB, PTYPE, PQUAL, PRES) only, so situational outs, baserunning and the running game ride along without a reliability of their own, and with the mean of the parts' MLB k.
| Piece | Levels | Carry-over b | k | Full Triple-A season | |
|---|---|---|---|---|---|
| Hitters | strikeouts | both (Double-A × 0.70) | 0.83 | 64 | 0.73 |
| Hitters | walks / HBP | both (Double-A × 0.70) | 0.76 | 139 | 0.58 |
| Hitters | batted-ball mix | both (Double-A × 0.70) | 1.47 | 384 | 0.79 |
| Hitters | contact quality | Triple-A only (no exit velocity at Double-A, where it is folded into results) | 0.94 | 108 | 0.76 |
| Hitters | results beyond contact | both (Double-A × 0.70) | 1.16 | 978 | 0.37 |
| Hitters | situational outs | Triple-A | 0.53 | 300 | 0.32 |
| Hitters | situational outs | Double-A | 0.59 | 300 | 0.35 |
| Hitters | baserunning | Triple-A | 0.79 | 300 | 0.47 |
| Hitters | baserunning | Double-A | 0.55 | 300 | 0.33 |
| Pitchers | strikeouts | both (Double-A × 1.00) | 0.54 | 110 | 0.44 |
| Pitchers | walks / HBP | both (Double-A × 1.00) | 0.24 | 133 | 0.18 |
| Pitchers | pitch quality: strikeouts / walks | Triple-A only (no pitch tracking at Double-A) | 0.42 | 20 | 0.40 |
| Pitchers | batted-ball mix allowed | both (Double-A × 1.00) | 0.62 | 1,000 | 0.19 |
| Pitchers | contact quality allowed | Triple-A only (no exit velocity at Double-A, where it is folded into results) | 0.83 | 1,702 | 0.17 |
| Pitchers | results beyond contact | both (Double-A × 1.00) | 0.04 | 1,100 | 0.01 |
| Pitchers | pitch quality: contact | Triple-A only (no pitch tracking at Double-A) | 0.45 | 20 | 0.43 |
| Pitchers | running game | Triple-A | 0.61 | 300 | 0.37 |
| Pitchers | running game | Double-A | 0.34 | 300 | 0.20 |
b multiplies the reliability-shrunk rate. The last column, b x 450 / (450 + k), is what a run above the league over a full season (450 PA or batters faced, the only minor-league time) is worth in MLB the next season. Single-piece parts (situational outs, baserunning, running game) take their k from the grid {300, 1,000, 3,000, 10,000}, the same at both levels.
| Part | AAA gap / 600 | AA gap / 600 | AAA age / yr | AA age / yr | d (AAA / AA) | MLB k in K | |
|---|---|---|---|---|---|---|---|
| Hitters | batting | −16.2 | −20.4 | −1.54 | −2.36 | 3.00 / 1.50 | 3,250 |
| Hitters | situational outs | +0.1 | +0.2 | −0.05 | −0.07 | 3.00 / 3.00 | 7,000 |
| Hitters | baserunning | −0.1 | −0.6 | −0.04 | −0.24 | 3.00 / 3.00 | 4,000 |
| Pitchers | strikeouts / walks | −5.1 | −6.7 | −0.58 | −0.25 | 0.50 / 0.50 | 1,250 |
| Pitchers | contact allowed | −1.0 | +0.7 | +0.39 | −0.04 | 3.00 / 0.25 | 5,000 |
| Pitchers | running game | −0.4 | −0.4 | −0.06 | −0.05 | 1.50 / 1.50 | 3,000 |
Gap = what an average player at the level, aged 26, projects to in MLB the next season. Age = change per year older than 26 (negative: the young project better). d = how much a minor-league plate appearance counts against an MLB one when both exist. MLB k in K = the forecast's final k for that part (Section 12), which K adds to the discounted minor-league units.
For hitters, strikeouts (0.73), contact quality (0.76) and batted-ball mix (0.79) carry over at roughly 0.7–0.8 of a full-season run, walks at 0.58 and results beyond contact at 0.37. For pitchers strikeouts carry over best among the result pieces (0.44); walks carry over weakly (0.18) and contact allowed hardly at all (0.19, 0.17, 0.01); Triple-A pitch quality (0.40 and 0.43) and the Triple-A running game (0.37) carry over far better than the contact pieces. Two hitter slopes are above 1 (batted-ball mix b = 1.47, results b = 1.16). Because b multiplies the shrunk rate, that means the split-half shrinkage pulls those pieces in further than next-season MLB outcomes support; the full-season carry-over stays below 1. Pitchers' Triple-A contact age slope is positive (+0.39 per year: older pitchers project slightly better contact), which the paper has no explanation for; it is small.
Earlier versions and what was dropped (numbers from README.md; their outputs were not kept). The first version used one regression constant for every piece and gave slopes above 1 and a Double-A exit-velocity slope of −12.7 on almost no data. Separate slopes per level fit no better (17.24 against 17.25 runs per 600 at the time) and gave unstable, sometimes negative, Double-A slopes. Minor-league strikeouts as a predictor of MLB contact allowed were dropped (14.11 against 14.12). Those comparisons were made against the first production fit (hitters 17.25, pitchers 14.12); the current fit reads 17.27 and 14.24 because three things changed since: Triple-A pitch quality entered, the forecast's pitchers' k was refit together with the MLB pitch-quality prior (strikeouts / walks 725 → 1,250, contact 2,000 → 5,000), and ages moved from July 1 to Opening Day. The first final fit put the batting gap at −15.7 (Triple-A) and −20.9 (Double-A) per 600; it now reads −16.2 and −20.4.
The translation in the forecast
forecast.add_minors applies the final fit to every player with Triple-A or Double-A time in 2024–2026:
rate_c = (MLB-only rate_c, with aging and, for pitchers' strikeouts/walks and contact,
the MLB pitch-quality prior) x (U + k_c) / (U + K_c) + translated part_c
centering weight w_c = k_c / (U + K_c)
Expanded, that is (R + k β PQ' + K m milb) / (U + K) + aging x (U + k) / (U + K): the MLB pitch-quality prior's weight shrinks from k / (U + k) to k / (U + K) and the MLB aging term is scaled by the same factor while the translation adds its own age term. Players without minor-league time are unchanged (checked on 75 of 75 hitters and 54 of 54 pitchers when it was wired in; README.md). Depth-chart players with minor-league time and no MLB time in the window are added with zero fielding and catcher defense before centering, the 2026 average positional adjustment per PA for their listed position (positions with 500+ PA; OF and IF are the means of LF/CF/RF and 2B/3B/SS), the typical platoon gap and left-handed share for their hand, and right-handed if the hand is unknown. Centering (Section 15) then gives them the full below-average prior level in the parts that were not translated: on this pull the 19 translated catchers carry −1.00 fielding and −3.62 catcher defense per 600, and other translated hitters −0.98 fielding. "No fielding value" therefore means none from the data; they still carry the centering level.
Back-test
| Group | n | Pieces: RMSE (r) | Bottom line: RMSE (r) | Before: RMSE (r) | Bias: pieces / before | |
|---|---|---|---|---|---|---|
| Hitters | all | 767 | 17.26 (0.34) | 17.42 (0.33) | 17.65 (0.28) | −1.6 / +0.3 |
| Hitters | minors only | 99 | 20.33 (0.25) | 20.92 (0.21) | 21.58 (−0.10) | −6.1 / −7.6 |
| Hitters | mostly minors | 241 | 17.67 (0.24) | 17.66 (0.23) | 18.57 (0.06) | −2.2 / +2.3 |
| Hitters | mostly MLB | 427 | 16.55 (0.42) | 16.72 (0.41) | 16.53 (0.42) | −0.6 / +0.7 |
| Pitchers | all | 928 | 14.24 (0.38) | 14.64 (0.33) | 14.22 (0.37) | −1.4 / −1.1 |
| Pitchers | minors only | 105 | 15.65 (0.17) | 16.21 (−0.07) | 15.54 (−0.02) | −5.1 / −4.8 |
| Pitchers | mostly minors | 281 | 15.82 (0.28) | 16.11 (0.24) | 15.75 (0.27) | −2.2 / −0.9 |
| Pitchers | mostly MLB | 542 | 13.36 (0.45) | 13.79 (0.42) | 13.39 (0.45) | −0.7 / −0.6 |
2025 and 2026 test seasons pooled; a, b, g, d and the scales fitted only on earlier targets (2022, 2023, 2024 for 2025, 2022, 2023, 2024, 2025 for 2026). Players with minor-league time in the three seasons before and 100+ MLB PA or batters faced in the test season; weighted by playing time, runs per 600. 'Before' is the comparison arm: MLB history pulled to average by the forecast's k (no aging, no pitch-quality prior, no centering), and the newcomer level for players with no MLB time. 'Minors only': no MLB time in the window; 'mostly minors': more weighted minor-league than MLB time. Bias = projected minus actual.
The evaluation, exactly: each test season T is projected from the three seasons before it with the parameters fitted on earlier targets; the total is batting + situational outs + baserunning for hitters and strikeouts / walks + contact + running game for pitchers (no fielding); the typical miss is the root mean squared error per 600 weighted by MLB units in T; r is unweighted. 'Before' is Σ_c R_c / (U + k_c), the forecast's MLB history with the final k and nothing else, and, for players with no MLB time in the window, the newcomer level of that test's own fit targets: what players with no MLB time in the three seasons before produced in 2022, 2023, 2024 (for 2025) or 2022, 2023, 2024, 2025 (for 2026), offense only: hitters −7.36 (2025) and −7.37 (2026) per 600 PA; pitchers −4.62 (2025) and −4.53 (2026) per 600 batters faced. The same quantity on the final fit's targets (2022–2026, which include both test seasons), −5.81 and −4.37, is stored but used by no test. Both differ from the forecast's newcomer level in Section 15 (players not in MLB the season before, all components). Because this level is one constant per test season, the 'before' arm cannot rank minors-only players: its correlation within a season is undefined (shown as 0.00 in the season table), and its pooled r for that group only reflects the difference between the two seasons' constants.
| Group | n | Pieces minus before (95% CI) | Pieces minus bottom line (95% CI) | Pieces r (95% CI) | Before r (95% CI) | |
|---|---|---|---|---|---|---|
| Hitters | all | 767 | −0.39 [−0.77, −0.02], better in 98.0% of 2,000 resamples | −0.16 [−0.32, −0.01], better in 98.4% of 2,000 resamples | 0.34 [+0.28, +0.40] | 0.28 [+0.22, +0.35] |
| Hitters | minors only | 99 | −1.25 [−3.14, +0.51], better in 91.6% of 2,000 resamples | −0.59 [−1.26, +0.07], better in 96.1% of 2,000 resamples | 0.25 [+0.05, +0.42] | −0.10 [−0.29, +0.10] |
| Hitters | mostly minors | 241 | −0.90 [−1.88, +0.14], better in 96.0% of 2,000 resamples | +0.01 [−0.32, +0.33], better in 46.7% of 2,000 resamples | 0.24 [+0.12, +0.35] | 0.06 [−0.07, +0.18] |
| Hitters | mostly MLB | 427 | +0.02 [−0.17, +0.21], better in 41.9% of 2,000 resamples | −0.16 [−0.30, −0.04], better in 99.4% of 2,000 resamples | 0.42 [+0.34, +0.49] | 0.42 [+0.34, +0.49] |
| Pitchers | all | 928 | +0.02 [−0.09, +0.13], better in 37.8% of 2,000 resamples | −0.40 [−0.55, −0.25], better in 100.0% of 2,000 resamples | 0.38 [+0.32, +0.43] | 0.37 [+0.31, +0.42] |
| Pitchers | minors only | 105 | +0.11 [−0.55, +0.73], better in 36.0% of 2,000 resamples | −0.56 [−1.14, +0.05], better in 96.0% of 2,000 resamples | 0.17 [−0.02, +0.35] | −0.02 [−0.21, +0.17] |
| Pitchers | mostly minors | 281 | +0.07 [−0.23, +0.35], better in 33.3% of 2,000 resamples | −0.29 [−0.54, −0.04], better in 99.0% of 2,000 resamples | 0.28 [+0.17, +0.39] | 0.27 [+0.16, +0.38] |
| Pitchers | mostly MLB | 542 | −0.03 [−0.08, +0.04], better in 80.0% of 2,000 resamples | −0.43 [−0.61, −0.26], better in 100.0% of 2,000 resamples | 0.45 [+0.38, +0.52] | 0.45 [+0.38, +0.51] |
Differences in typical miss, runs per 600: paired bootstrap over players, resampling each player's projections together (2,000 resamples, paper_checks_tail.py); 'better in' = share of resamples in which the first arm misses less. Correlations with Fisher intervals (r is unweighted).
| Test season | Group | n | Pieces: RMSE (r) | Before: RMSE (r) | Pieces minus before (95% CI) | |
|---|---|---|---|---|---|---|
| Hitters | 2025 | all | 378 | 16.98 (0.34) | 17.32 (0.31) | −0.35 [−0.80, +0.10], better in 93.5% of 2,000 resamples |
| Hitters | 2025 | minors only | 49 | 22.70 (0.06) | 21.71 (0.00) | +0.99 [−1.38, +3.43], better in 18.2% of 2,000 resamples |
| Hitters | 2025 | mostly minors | 122 | 17.40 (0.27) | 18.58 (0.10) | −1.18 [−2.14, −0.16], better in 98.8% of 2,000 resamples |
| Hitters | 2025 | mostly MLB | 207 | 15.73 (0.44) | 15.93 (0.42) | −0.21 [−0.47, +0.04], better in 94.2% of 2,000 resamples |
| Hitters | 2026 | all | 389 | 17.53 (0.34) | 17.95 (0.26) | −0.42 [−1.03, +0.15], better in 92.9% of 2,000 resamples |
| Hitters | 2026 | minors only | 50 | 17.90 (0.44) | 21.46 (−0.00) | −3.56 [−5.99, −1.31], better in 100.0% of 2,000 resamples |
| Hitters | 2026 | mostly minors | 119 | 17.95 (0.23) | 18.57 (0.01) | −0.61 [−2.42, +1.19], better in 76.1% of 2,000 resamples |
| Hitters | 2026 | mostly MLB | 220 | 17.29 (0.40) | 17.08 (0.41) | +0.22 [−0.09, +0.52], better in 7.8% of 2,000 resamples |
| Pitchers | 2025 | all | 466 | 13.88 (0.42) | 14.05 (0.39) | −0.17 [−0.32, −0.03], better in 98.9% of 2,000 resamples |
| Pitchers | 2025 | minors only | 56 | 15.28 (0.26) | 15.64 (0.00) | −0.36 [−1.21, +0.55], better in 77.5% of 2,000 resamples |
| Pitchers | 2025 | mostly minors | 146 | 14.12 (0.30) | 14.51 (0.21) | −0.39 [−0.82, +0.02], better in 96.4% of 2,000 resamples |
| Pitchers | 2025 | mostly MLB | 264 | 13.59 (0.50) | 13.63 (0.49) | −0.05 [−0.12, +0.03], better in 88.4% of 2,000 resamples |
| Pitchers | 2026 | all | 462 | 14.58 (0.33) | 14.39 (0.35) | +0.19 [+0.03, +0.34], better in 0.9% of 2,000 resamples |
| Pitchers | 2026 | minors only | 49 | 16.08 (0.04) | 15.41 (0.00) | +0.66 [−0.23, +1.46], better in 6.7% of 2,000 resamples |
| Pitchers | 2026 | mostly minors | 135 | 17.27 (0.27) | 16.83 (0.33) | +0.44 [+0.03, +0.80], better in 2.0% of 2,000 resamples |
| Pitchers | 2026 | mostly MLB | 278 | 13.13 (0.41) | 13.13 (0.41) | 0.00 [−0.09, +0.09], better in 53.7% of 2,000 resamples |
Each test season on its own (fits as in the pooled table).
For hitters the translation beats the forecast without it in all, minors only, mostly minors but not in mostly MLB (mostly MLB: 16.55 against 16.53), and the 95% interval excludes zero only for all; the piece-by-piece version beats translating total value with an interval clear of zero for all, mostly MLB; for minors only, mostly minors the two are not distinguishable. Players coming straight from the minors gain the most in typical miss, from 21.6 to 20.3 runs per 600, but with 99 players the interval on the gain is wide. The translation does rank them: r = 0.25 [+0.05, +0.42] pooled, 0.06 in 2025 (n = 49) and 0.44 in 2026 (n = 50) by season; the 'before' arm's −0.10 is not a ranking (it is one constant per season). Season by season the picture moves (minors only 2025: 22.70 against 21.71; mostly MLB 2025: 15.73 against 15.93; minors only 2026: 17.90 against 21.46; mostly MLB 2026: 17.29 against 17.08). For pitchers the translation ranks players a little better but does not shrink the typical miss: among pitcher pieces only strikeouts (0.44), pitch quality (0.40 / 0.43) and the Triple-A running game (0.37) carry over at more than about a third of a run per full season. Both arms under-project call-ups who reach 100 PA or batters faced: hitters with no MLB time by +6.1 (translation) and +7.6 (before) runs per 600, pitchers by +5.1 and +4.8: players who perform keep playing, so the test group is selected toward success. The case for pitchers is weak. The pooled typical miss is slightly worse than without the translation, and the minors-only ranking, r = 0.17 [−0.02, +0.35] pooled (0.26 in 2025 (n = 56) and 0.04 in 2026 (n = 49) by season), has an interval that includes zero and is compared with an arm that cannot rank. The pitchers' translation went into production with the hitters' on the user's go-ahead for both kinds together (README.md). No test here shows it lowering pitchers' typical miss: the corrected comparisons below put its fitted path behind the forecast without the minors, and its production path level with it (−0.06 [−0.19, +0.07], better in 82.0% of 2,000 resamples).
What the back-test does not test
Five properties of the back-test, measured on its own rows by paper_checks_tail.py:
- The pitchers' "before" arm is too weak. It uses the forecast's final k for strikeouts / walks and contact (1,250 and 5,000), but those values were fitted together with the pitch-quality prior (β = 0.7 and 0.5, kq = 400 and 50), and the back-test applies no prior: the k units of prior play sit at zero, so MLB history is regressed harder than the forecast regresses it. With the prior put back, the pooled pitchers' "before" improves from 14.22 to 13.99 (−0.23 [−0.41, −0.06], better in 99.6% of 2,000 resamples); with the plain no-pitch-quality k (725 and 2,000) it is 14.09. Those final pitch-quality settings were pooled over the 2023–26 targets, which include both test seasons; refit on the targets before each test season only (2025: strikeouts / walks k 1,500, β 0.6, kq 400, contact k 10,000, β 0.6, kq 200; 2026: strikeouts / walks k 1,250, β 0.5, kq 200, contact k 7,000, β 0.5, kq 50), the corrected "before" is 14.05 (−0.17 [−0.35, −0.01], better in 98.4% of 2,000 resamples against the published arm). Against the corrected baseline the translation's fitted path (14.24) is worse for pitchers, not tied: +0.25 [+0.04, +0.46], better in 0.9% of 2,000 resamples with the final settings, +0.19 [−0.01, +0.39], better in 3.1% of 2,000 resamples with the out-of-sample ones. The direction holds; the out-of-sample interval just reaches zero.
- The translation's fixed part has the same gap. row_for's MLB part is R / (U + K) with no prior, so a, b and g were fitted as if the prior play sat at zero, while production adds k β PQ' / (U + K) on top. In the 2027 forecast that untested term averages 1.59 runs per 600 in absolute value over the 808 projected pitchers with both MLB and minor-league time (largest 8.90; 452 at 1+), and 1.67 over the 665 of them with a depth-chart role (386 at 1+).
- Production combines the translation with aging and the prior; the back-test scores neither. Scoring the production path on the same rows (aging and the prior added and scaled as add_minors does) gives 17.13 for hitters and 14.06 for pitchers against the fitted path's 17.26 and 14.24 (hitters −0.13 [−0.31, +0.05], better in 92.1% of 2,000 resamples; pitchers −0.18 [−0.43, +0.07], better in 92.6% of 2,000 resamples). Comparing production with production, translation against none: hitters −0.60 [−1.01, −0.20], better in 99.9% of 2,000 resamples, pitchers −0.06 [−0.19, +0.07], better in 82.0% of 2,000 resamples. So the untested path is not worse than the tested one, and the translation's value for hitters survives the correction. These production columns use the forecast's final aging and pitch-quality settings, partly fitted on the test seasons.
- k is not out of sample. The MLB k inside K, in both arms, is the forecast's final k, fitted on targets that include 2025 and 2026 (Section 12). a, b, g, d and the scales are out of sample; k is not. The reliability k_j pool every minor-league season, including the histories of later targets; they use no MLB outcome.
- Situational outs. The minor-league situational-out slope (b = 0.53 at Triple-A with k = 300, d = 3.00) keeps 0.32 of a full Triple-A season's rate, while the MLB forecast, whose k for situational outs is an interior optimum of its grid (7,000 weighted units), keeps 0.24 of the same 450 PA as the last MLB season (5 x 450 weighted units), 0.30 of a 600-PA season and 0.51 of three full seasons. Minor-league situational outs are trusted 1.30 times as much as the same amount of MLB time: a small tension, not a large one, and not tested on its own.
| Group | n | Before, as published | Before + pitch-quality prior | Before, no-PQ k | Before, production path | Pieces, fitted path | Pieces, production path | |
|---|---|---|---|---|---|---|---|---|
| Hitters | all | 767 | 17.65 (0.28) | – | – | 17.72 (0.29) | 17.26 (0.34) | 17.13 (0.35) |
| Hitters | minors only | 99 | 21.58 (−0.10) | – | – | 21.58 (−0.10) | 20.33 (0.25) | 20.33 (0.25) |
| Hitters | mostly minors | 241 | 18.57 (0.06) | – | – | 19.10 (0.07) | 17.67 (0.24) | 17.63 (0.24) |
| Hitters | mostly MLB | 427 | 16.53 (0.42) | – | – | 16.41 (0.44) | 16.55 (0.42) | 16.36 (0.44) |
| Pitchers | all | 928 | 14.22 (0.37) | 13.99 (0.39) | 14.09 (0.38) | 14.12 (0.37) | 14.24 (0.38) | 14.06 (0.38) |
| Pitchers | minors only | 105 | 15.54 (−0.02) | 15.54 (−0.02) | 15.54 (−0.02) | 15.54 (−0.02) | 15.65 (0.17) | 15.65 (0.17) |
| Pitchers | mostly minors | 281 | 15.75 (0.27) | 15.46 (0.29) | 15.67 (0.26) | 15.60 (0.26) | 15.82 (0.28) | 15.47 (0.29) |
| Pitchers | mostly MLB | 542 | 13.39 (0.45) | 13.15 (0.46) | 13.20 (0.45) | 13.29 (0.45) | 13.36 (0.45) | 13.24 (0.45) |
Typical miss per 600 (correlation) on the same back-test rows. 'Before + pitch-quality prior': the MLB history with the forecast's final k and the prior play at the pitcher's pitch quality, as the forecast itself does. 'No-PQ k': plain fit_k values without pitch quality, averaged over the 2025 and 2026 fits (as the forecast averages its k). 'Production path': adds the pitch-quality prior and the forecast's aging, scaled as add_minors scales them. The production-path columns use the forecast's final aging and pitch-quality settings, which were fitted with the 2025 and 2026 targets in the sample, so they flatter those columns somewhat.
Do prospect grades help?
FanGraphs' Future Value and tool grades were tested as additions to the translation. The grades come from The Board, using each season's end-of-season list (2021–26, data/prospects/fg_board_2021_2026.json, 5,403 graded player-seasons, joined to MLB ids through FanGraphs' minor-league leaderboards, which carry the minor-league id, the FanGraphs id and the MLBAM id; 620–1,088 players per list match, everyone who reached Triple-A or Double-A). For target season T the T−1 list is used, which is what was known when a forecast for T would have been made. The lists and leaderboards sit behind a Cloudflare challenge (curl gets the challenge page), so they were read in a browser and moved to disk through the clipboard; two traps came with that: pbpaste without a UTF-8 locale writes MacRoman, and the Board's level fields (mlevel, llevel) give the player's level today, not at list time (Rutschman shows "MLB" on the 2021 list), so they are never used. These facts are from README.md; the data file is the record.
The grades enter as extra terms in the minor-league projection, weighted like the stats: each feature x is added as f x x, where f = K m / (U_mlb + K) is the minor-league share defined below, and its coefficient is fitted with the others by the same weighted least squares. The features are Future Value as (FV − 45) / 5 plus an indicator for players not on the list; the tools as (grade − 45) / 5 for future hit and game power (batting) and speed (baserunning); and, for pitchers, the best future pitch grade as (max − 55) / 5 and future command as (grade − 45) / 5, both in strikeouts / walks, the best pitch also in contact. 'Grades only' keeps FV, the unlisted indicator, the level gap and age and drops every minor-league stat. A fifth variant coded in milb_translate.py (FV fading with the minor-league sample, FV x 500 / (n + 500)) was not run. A second test caps every player's weight at 150 PA or batters faced, in the fit as well as in the evaluation, and counts every call-up with 30+. That guards against a selection trap: teams keep playing high-graded prospects who struggle and drop low-graded players who struggle, so weighting by playing time could hide what the grades know. Before the test, the expectation (README.md) was that grades would help pitchers and short samples most.
| Test | Group | n | Stats only | + FV | + tools | Grades only | Before | + FV minus stats only (95% CI) | |
|---|---|---|---|---|---|---|---|---|---|
| Hitters | playing-time weights, 100+ | all with minor-league time | 767 | 17.26 (0.34) | 17.26 (0.34) | 17.26 (0.34) | 17.59 (0.29) | 17.65 (0.28) | 0.00 [−0.11, +0.10], better in 50.9% of 2,000 resamples |
| Hitters | playing-time weights, 100+ | no MLB time | 99 | 20.33 (0.25) | 20.60 (0.22) | 20.44 (0.24) | 21.14 (0.04) | 21.58 (−0.10) | +0.27 [−0.37, +0.86], better in 21.2% of 2,000 resamples |
| Hitters | playing-time weights, 100+ | mostly minors | 241 | 17.67 (0.24) | 17.49 (0.26) | 17.52 (0.26) | 18.42 (0.08) | 18.57 (0.06) | −0.19 [−0.41, +0.03], better in 95.2% of 2,000 resamples |
| Hitters | playing-time weights, 100+ | mostly MLB | 427 | 16.55 (0.42) | 16.58 (0.41) | 16.60 (0.41) | 16.60 (0.41) | 16.53 (0.42) | +0.03 [−0.02, +0.08], better in 11.8% of 2,000 resamples |
| Pitchers | playing-time weights, 100+ | all with minor-league time | 928 | 14.24 (0.38) | 14.24 (0.38) | 14.22 (0.38) | 14.34 (0.36) | 14.22 (0.37) | 0.00 [−0.03, +0.03], better in 52.0% of 2,000 resamples |
| Pitchers | playing-time weights, 100+ | no MLB time | 105 | 15.65 (0.17) | 15.74 (0.15) | 15.66 (0.18) | 15.78 (−0.06) | 15.54 (−0.02) | +0.09 [−0.03, +0.20], better in 7.8% of 2,000 resamples |
| Pitchers | playing-time weights, 100+ | mostly minors | 281 | 15.82 (0.28) | 15.82 (0.28) | 15.80 (0.28) | 16.04 (0.26) | 15.75 (0.27) | 0.00 [−0.07, +0.08], better in 49.1% of 2,000 resamples |
| Pitchers | playing-time weights, 100+ | mostly MLB | 542 | 13.36 (0.45) | 13.35 (0.45) | 13.34 (0.46) | 13.40 (0.45) | 13.39 (0.45) | −0.01 [−0.03, 0.00], better in 99.0% of 2,000 resamples |
| Hitters | weights capped at 150, 30+ | all with minor-league time | 984 | 21.73 (0.38) | 21.72 (0.38) | 21.71 (0.38) | 22.11 (0.34) | 23.11 (0.26) | −0.01 [−0.06, +0.03], better in 71.7% of 2,000 resamples |
| Hitters | weights capped at 150, 30+ | no MLB time | 163 | 24.83 (0.27) | 24.92 (0.26) | 24.92 (0.26) | 25.35 (0.23) | 25.62 (−0.10) | +0.09 [−0.07, +0.23], better in 13.7% of 2,000 resamples |
| Hitters | weights capped at 150, 30+ | mostly minors | 346 | 23.84 (0.32) | 23.77 (0.33) | 23.74 (0.33) | 24.50 (0.23) | 26.58 (0.05) | −0.07 [−0.16, +0.02], better in 94.6% of 2,000 resamples |
| Hitters | weights capped at 150, 30+ | mostly MLB | 475 | 19.32 (0.42) | 19.32 (0.42) | 19.33 (0.42) | 19.43 (0.40) | 19.83 (0.42) | 0.00 [−0.02, +0.02], better in 48.8% of 2,000 resamples |
| Pitchers | weights capped at 150, 30+ | all with minor-league time | 1287 | 17.94 (0.28) | 17.94 (0.28) | 17.94 (0.28) | 17.97 (0.28) | 18.00 (0.28) | 0.00 [−0.02, +0.02], better in 45.1% of 2,000 resamples |
| Pitchers | weights capped at 150, 30+ | no MLB time | 208 | 20.69 (0.10) | 20.72 (0.10) | 20.65 (0.12) | 20.85 (0.06) | 20.62 (0.05) | +0.03 [−0.03, +0.09], better in 19.1% of 2,000 resamples |
| Pitchers | weights capped at 150, 30+ | mostly minors | 415 | 18.88 (0.17) | 18.88 (0.17) | 18.90 (0.16) | 18.87 (0.21) | 19.05 (0.20) | 0.00 [−0.05, +0.04], better in 53.0% of 2,000 resamples |
| Pitchers | weights capped at 150, 30+ | mostly MLB | 664 | 16.62 (0.39) | 16.62 (0.39) | 16.62 (0.39) | 16.65 (0.38) | 16.64 (0.38) | 0.00 [−0.02, +0.01], better in 72.9% of 2,000 resamples |
Typical miss in runs per 600 (correlation), 2025 and 2026 pooled, each fitted on earlier seasons only. '+ FV' adds Future Value (per 5 points) and an indicator for players not on the list; '+ tools' adds future hit, game power and speed grades (pitchers: best future pitch and command); 'Grades only' drops the minor-league stats and keeps FV, age and level; 'Before' is the translation back-test's comparison arm (MLB history over its units plus k, no aging, no pitch-quality prior; the newcomer level for players with no MLB time). Last column: paired bootstrap over players of the difference in typical miss (2,000 resamples, paper_checks_tail.py). Results file of 2026-11-16 17:59; production translation of 2026-11-16 17:46.
Neither version improved on the stats. Adding Future Value or the tool grades moved the typical miss by at most 0.27 runs per 600 in any group of either test, and the only paired intervals that exclude zero in the grades' favor are pitchers, mostly MLB, weighted, + fv: −0.014 [−0.027, −0.002]; pitchers, mostly MLB, weighted, + tools: −0.022 [−0.037, −0.007], all a few hundredths of a run per 600. Once the stats are in, the fitted FV coefficient is −0.02 runs per 600 per 5 FV points for hitters' batting and +0.75 / −0.85 for pitchers' strikeouts / walks and contact (opposite signs that nearly cancel); the unlisted indicator is +3.78 for hitters, so a player missing from the list projects slightly better, not worse, once his stats are known. On their own, grades beat the old flat newcomer level for hitters but fall short of the stats. Future Value tracks age-for-level closely (correlation −0.57 to −0.44 across the four level-kind groups, 378–658 graded player-seasons with 200+ PA or batters faced at the level: the higher graded are the younger) and the same season's minor-league production (0.19 to 0.34), and the translation already uses both. The translation therefore stays stats only. Grades may still carry information this test cannot see. One example is who reaches MLB and gets playing time, which the forecast takes from FanGraphs' depth charts. Another is the low minors, where there are no Triple-A or Double-A stats. milb_grades_test.py reruns the test (8 fits in parallel; its run time is recorded in paper_checks_tail.json).
| Weights | Mode | Part | Fitted coefficients (runs per 600) | |
|---|---|---|---|---|
| Hitters | weighted | fv | batting | FV (per 5 points from 45) −0.02; not on the list +3.78 |
| Hitters | weighted | tools | batting | future hit (per 5 from 45) +0.80; future game power (per 5 from 45) +1.45; FV (per 5 points from 45) −1.40; not on the list +4.58 |
| Hitters | weighted | tools | baserunning | future speed (per 5 from 45) +0.35 |
| Hitters | weighted | fvonly | batting | FV (per 5 points from 45) +0.93; not on the list +2.80 |
| Pitchers | weighted | fv | strikeouts / walks | FV (per 5 points from 45) +0.75; not on the list +0.86 |
| Pitchers | weighted | fv | contact allowed | FV (per 5 points from 45) −0.85; not on the list +0.56 |
| Pitchers | weighted | tools | strikeouts / walks | best future pitch (per 5 from 55) +0.13; future command (per 5 from 45) +0.27; FV (per 5 points from 45) +0.49; not on the list +1.41 |
| Pitchers | weighted | tools | contact allowed | best future pitch (per 5 from 55) +1.28; FV (per 5 points from 45) −1.49; not on the list +2.37 |
| Pitchers | weighted | fvonly | strikeouts / walks | FV (per 5 points from 45) +1.93; not on the list −1.48 |
| Pitchers | weighted | fvonly | contact allowed | FV (per 5 points from 45) −0.83; not on the list +0.63 |
| Hitters | capped | fv | batting | FV (per 5 points from 45) −0.12; not on the list +2.88 |
| Hitters | capped | tools | batting | future hit (per 5 from 45) +0.59; future game power (per 5 from 45) +1.44; FV (per 5 points from 45) −1.37; not on the list +3.59 |
| Hitters | capped | tools | baserunning | future speed (per 5 from 45) +0.30 |
| Hitters | capped | fvonly | batting | FV (per 5 points from 45) +0.62; not on the list +2.65 |
| Pitchers | capped | fv | strikeouts / walks | FV (per 5 points from 45) +0.57; not on the list +0.11 |
| Pitchers | capped | fv | contact allowed | FV (per 5 points from 45) −0.80; not on the list +0.45 |
| Pitchers | capped | tools | strikeouts / walks | best future pitch (per 5 from 55) +0.27; future command (per 5 from 45) +0.17; FV (per 5 points from 45) +0.39; not on the list +0.58 |
| Pitchers | capped | tools | contact allowed | best future pitch (per 5 from 55) +1.51; FV (per 5 points from 45) −1.48; not on the list +2.56 |
| Pitchers | capped | fvonly | strikeouts / walks | FV (per 5 points from 45) +1.60; not on the list −2.00 |
| Pitchers | capped | fvonly | contact allowed | FV (per 5 points from 45) −0.82; not on the list +0.62 |
Final fit (targets 2022–26). A coefficient multiplies f x feature, f being the minor-league share of the projection (1 for a player with no MLB time), so it is runs per 600 for a minors-only player per unit of the feature.
Earlier run. The test was first run on 2026-11-14 14:53, before Triple-A pitch quality entered the translation, before the forecast's pitchers' k was refit together with the MLB pitch-quality prior, and before ages moved to Opening Day. Its numbers are kept in paper_checks_tail.json (history); the conclusion did not change:
| Weights | Stats only, earlier run | Stats only, current | Before, earlier run | Before, current | Largest grade effect, earlier | Largest grade effect, current | |
|---|---|---|---|---|---|---|---|
| Hitters | weighted | 17.25 | 17.26 | 17.62 | 17.65 | 0.02 | 0.00 |
| Pitchers | weighted | 14.12 | 14.24 | 14.09 | 14.22 | 0.02 | 0.02 |
| Hitters | capped | 21.81 | 21.73 | 23.11 | 23.11 | 0.02 | 0.02 |
| Pitchers | capped | 17.90 | 17.94 | 17.93 | 18.00 | 0.01 | 0.00 |
All players with minor-league time, typical miss in runs per 600. 'Largest grade effect' = the largest change in typical miss from adding FV or tools in that row's test, in any direction. The pitchers' 'Before' moved because the forecast's k for strikeouts/walks and contact moved (see 'What the back-test does not test' above).
High-A and Single-A
High-A and Single-A were rebuilt the same way (2021–26, about 150,000 plate appearances per level per season, the same reconciliation and the same per-league model; leagues South Atlantic / Midwest / Northwest at High-A and Carolina / Florida State / California at Single-A, interim names in 2021) and added to the translation as two more levels (LEVELS=aaa,aa,hia,a). Each gets its own gap, age term, playing-time discount and scale of the shared carry-overs, never above the level above it: after the Triple-A / Double-A grid, one coordinate round per lower level searches d over the discount grid at or below the level above and the scale over {0.1, 0.2, 0.3, 0.4, 0.55, 0.7, 0.85, 1} at or below the level above. A level is treated as untracked if no season reaches 90% exit-velocity coverage; there its contact quality is folded into results, in the history sums and in the split halves, so Single-A's Florida State League tracking (27%–29% of balls by season) is not lost. The fold also applies to Double-A (a few dozen tracked balls in 2024) and moved the production coefficients by at most 0.16 runs per 600 with the back-test unchanged (README.md).
| Players | n | Four levels | Two levels (production setting) | Before the minors | Four minus two levels (95% CI) | |
|---|---|---|---|---|---|---|
| Hitters | Triple-A or Double-A time | 767 | 17.36 (0.33) | 17.26 (0.34) | 17.65 (0.28) | +0.10 [−0.05, +0.25], better in 7.5% of 2,000 resamples |
| Hitters | Triple-A or Double-A time, minors only | 99 | 20.63 (0.21) | 20.33 (0.25) | 21.58 (−0.10) | |
| Hitters | Triple-A or Double-A time, mostly minors | 251 | 17.51 (0.24) | 17.39 (0.25) | 18.28 (0.08) | |
| Hitters | High-A / Single-A time only | 10 | 17.20 (0.36) | 17.77 (0.39) | 17.77 (0.39) | −0.57 [−1.26, +0.41], better in 90.1% of 2,000 resamples |
| Hitters | any minor-league time | 777 | 17.36 (0.33) | 17.27 (0.34) | 17.65 (0.28) | +0.10 [−0.04, +0.24], better in 8.9% of 2,000 resamples |
| Pitchers | Triple-A or Double-A time | 928 | 14.24 (0.37) | 14.24 (0.38) | 14.22 (0.37) | +0.01 [−0.04, +0.05], better in 37.4% of 2,000 resamples |
| Pitchers | Triple-A or Double-A time, minors only | 105 | 15.66 (0.12) | 15.65 (0.17) | 15.54 (−0.02) | |
| Pitchers | Triple-A or Double-A time, mostly minors | 291 | 15.65 (0.27) | 15.63 (0.28) | 15.57 (0.27) | |
| Pitchers | High-A / Single-A time only | 17 | 16.88 (0.24) | 16.08 (0.55) | 16.08 (0.55) | +0.80 [−0.54, +2.94], better in 12.2% of 2,000 resamples |
| Pitchers | any minor-league time | 945 | 14.30 (0.37) | 14.28 (0.38) | 14.26 (0.37) | +0.03 [−0.03, +0.08], better in 20.6% of 2,000 resamples |
Typical miss in runs per 600 (correlation), 2025 and 2026 pooled, each fitted on earlier seasons only, players with 100+ MLB PA or batters faced. 'Two levels' gives players with only High-A / Single-A time today's treatment (MLB history pulled to average, or the newcomer level). The groups count minor-league units at all four levels, so 'mostly minors' holds more players than in the production back-test. Last column: paired bootstrap (2,000 resamples). Results file of 2026-11-16 18:08.
They do not help next-season projections. Nearly everyone who reaches MLB has Triple-A or Double-A time in the three seasons before, so the lower levels add only older, noisier information about the same players, and more parameters. Only 10 hitters and 17 pitchers in the test had nothing above High-A. The four-level fit put the batting gap at −16.6 (Triple-A), −20.1 (Double-A), −22.7 (High-A) and −35.1 (Single-A) runs per 600, with a Single-A age slope of −5.94 per year: the few players who go straight from A-ball to the majors are exceptional, and the fit chases them. The test's own two-level fit reproduces the production back-test row by row (identical to 0.0001 runs per 600 for hitters and pitchers). The test was first run on 2026-11-14 15:17, before Triple-A pitch quality, the pitchers' k refit and Opening Day ages; rerun on the current settings the numbers moved (hitters four 17.33 → 17.36, two 17.25 → 17.26; pitchers four 14.14 → 14.24, two 14.12 → 14.24) and the decision did not. The forecast therefore uses Triple-A and Double-A. The lower levels stay on disk for projections further ahead, where they would carry the information; milb_levels_test.py reruns the comparison (--compare reuses the saved fits in output/milb/levels_test/).
| Part | AAA gap / age | AA gap / age | A+ gap / age | A gap / age | d (AAA / AA / A+ / A) | Scale (AA / A+ / A) | |
|---|---|---|---|---|---|---|---|
| Hitters | batting | −16.6 / −1.66 | −20.1 / −1.22 | −22.7 / −1.35 | −35.1 / −5.94 | 3.00 / 1.50 / 1.50 / 1.50 | 0.85 / 0.85 / 0.55 |
| Hitters | situational outs | +0.1 / −0.05 | +0.3 / +0.04 | +3.6 / +0.82 | −10.7 / −2.91 | 3.00 / 3.00 / 1.00 / 1.00 | 1.00 / 1.00 / 1.00 |
| Hitters | baserunning | −0.1 / −0.01 | −1.2 / −0.56 | +1.8 / +0.62 | −4.6 / −0.35 | 3.00 / 3.00 / 3.00 / 0.25 | 1.00 / 1.00 / 1.00 |
| Pitchers | strikeouts / walks | −5.2 / −0.68 | −7.8 / −0.72 | −2.6 / +3.43 | −4.0 / −1.78 | 0.50 / 0.50 / 0.50 / 0.50 | 1.00 / 1.00 / 0.55 |
| Pitchers | contact allowed | −1.1 / +0.38 | +1.8 / +0.29 | −1.8 / +0.22 | +0.3 / −0.37 | 3.00 / 0.25 / 0.25 / 0.25 | 1.00 / 1.00 / 1.00 |
| Pitchers | running game | −0.4 / −0.06 | −0.3 / −0.01 | −2.0 / −0.56 | +2.3 / +1.62 | 2.00 / 2.00 / 1.00 / 0.25 | 1.00 / 1.00 / 1.00 |
The four-level final fit: level gap per 600 at age 26, age slope per year older, the playing-time discount d and the scale of the shared carry-overs (1.00 for the single-piece parts).
What it changes in 2027
The forecast now includes 80 hitters and 112 pitchers coming up from Triple-A or Double-A and 22 hitters and 51 pitchers back from a lost season, all with 2027 depth-chart roles. It also uses the minors as the starting point for 606 hitters and 808 pitchers already in the pool who had minor-league time in 2024–2026. That matters most for fringe major leaguers with long Triple-A records, whom the forecast without the minors pulled toward MLB average: among the 606 pool hitters with minor-league time the minors moved the projection by a median of −3.41 runs per 600 (mean −6.48); the largest moves in projected runs were Will Banfield −13.7, Wade Meckler −13.5, TJ Rumfield −13.2, Victor Mesa Jr. −13.1, Tristan Peters −12.5. Pool pitchers moved a median of −0.21 (mean −1.13); 731 depth-chart pitchers carry a Triple-A pitch-quality shift and 106 of them by 2+ runs per 600. Translated players get the positional adjustment of their listed position, the typical platoon gap for their handedness, and no fielding value from the data (the centering level only, above). One listing trap: a Stats API primary position of DH (Sebastian Walcott) brings the designated hitter's adjustment.
| Player | Age | Club | Latest level | Window PA (AAA / AA) | Window runs / 600 vs league | 2027 MLB / 600 | Projected PA |
|---|---|---|---|---|---|---|---|
| Aidan Miller | 22 | PHI | 2026 AA | 37 / 522 | +31.9 | +0.8 | 343 |
| Sebastian Walcott | 21 | TEX | 2026 AAA | 9 / 747 | +23.4 | +2.5 | 322 |
| Owen Ayers | 25 | CHC | 2026 AAA | 208 / 278 | +26.1 | −10.0 | 321 |
| Seaver King | 23 | WSH | 2026 AAA | 356 / 522 | +15.7 | −2.5 | 266 |
| Ralphy Velazquez | 21 | CLE | 2026 AAA | 369 / 295 | +35.2 | +8.6 | 245 |
| Murf Gray | 23 | PIT | 2026 AA | 0 / 195 | +22.0 | −8.1 | 196 |
| Nelson Rada | 21 | LAA | 2026 AAA | 786 / 889 | +0.9 | −11.8 | 182 |
| Noah Cardenas | 27 | MIN | 2026 AAA | 404 / 432 | −4.1 | −28.4 | 179 |
Players up from the minors with the most 2027 depth-chart playing time (offense only: batting, situational outs, baserunning; after centering). 'Latest level' is the highest level of his latest minor-league season; the window columns sum every Triple-A and Double-A line of 2024–2026 without weights (the projection weights them 5/4/3 and shrinks each piece). The projection keeps only part of a minor-league edge, after the level gap, with credit for youth. Age on Opening Day, March 25, 2027.
15.Playing time, newcomers and projected wins
Playing time and teams
Projected plate appearances, batters faced and games started come from FanGraphs Depth Charts, whose rosters and playing time follow FanGraphs' RosterResource depth charts. Only playing time and team are taken from it; every rate is this model's own. fetch_depth_charts.py requests the projections endpoint (type fangraphsdc) with curl, because FanGraphs refuses Python's client, and keeps xMLBAMID, PlayerName, Team, PA, G, WAR for hitters and xMLBAMID, PlayerName, Team, TBF, IP, G, GS, WAR for pitchers; the endpoint has no season parameter and always serves the coming season, so each file is stamped with its pull date. On the pull of 2026-11-16: 669 hitters and 829 pitchers in the pool had a 2027 role (89 and 124 of them unsigned free agents, who keep FanGraphs' projected playing time and count for no team until they sign); 92 hitters and 146 pitchers were on no depth chart and get zero playing time, keeping their per-600 rates. A pitcher's role is a label only (it changes no rate): starter = gs >= 8 or (g and gs / g >= 0.5). One gap in that pipeline: fetch_depth_charts.py keeps only rows with an MLBAM id. Fetched again on 2026-11-16 18:11, the feed had 671 hitter and 836 pitcher rows, of which 1 and 2 have no MLBAM id (Teruaki Sato (FA, 623); Kaima Taira (FA, 607); Hiromi Itoh (FA, 447); PA or batters faced). All are unsigned, so today they cost no club anything, but when one signs his playing time vanishes from his club and from the leftover newcomer time until FanGraphs adds the id: 623 PA and 1,054 batters faced at stake. The slim files on disk hold 670 and 834 rows.
The feed's team assignments were checked once against all 30 RosterResource pages during development: 1,276 players on the same club in both, none on a different club, no feed free agent on a RosterResource roster, and 9 feed players on no RosterResource page. That check is not stored in any output or in README.md and was not repeated on later pulls, so it cannot be reproduced from the repository. If the depth-chart files are missing, playing time falls back to Marcel's rule, with its own starter definition from the last season's starts and batters faced, and the forecast then skips league centering and team totals entirely:
hitters: marcel = (0.5 * last["u"] + 0.1 * HS[y2].get(pid, {}).get("u", 0) + 200) if not origin else None
pitchers: starter26 = gs_last >= 10 or (gs_last >= 3 and gs_last * 20 >= 0.5 * last["u"])
marcel = (0.5 * last["u"] + 0.1 * PS[y2].get(pid, {}).get("u", 0) + (260 if starter26 else 110)) if not origin else None
no depth charts, no centering or team totals:
centering = center(hitters, pitchers, dc_b, dc_p, years) if dc_b is not None else None
Players outside the 2026 majors
Depth charts also give 2027 playing time to players who did not appear in MLB in 2026. Since the minor-league stage they are projected individually, like everyone else:
- Back from a lost season (22 hitters, 51 pitchers with a 2027 role): MLB time in 2024–2025 but none in 2026. They are projected from their own MLB history with the same regression and aging, with any Triple-A or Double-A rehab time entering through the translation. The history weights are positional (5/4/3 hitters, 3/2/1 pitchers for 2026, 2025, 2024) and the empty 2026 slot is not re-weighted, so in effect their seasons count 4/3 and 2/1: the same k is a larger share of a smaller weighted history and they are pulled further toward average, and their aging gap (forecast year minus the weighted mean season) is longer: 2.4 years for equal time in the two seasons, (4 x 2025 + 3 x 2024) / 7, against 1.8 for a regular with equal time in all three.
- Up from Triple-A or Double-A (80 hitters, 112 pitchers): no MLB time in 2024–2026, projected from the translation alone (Section 14).
| Back from a lost season: players / units | Up from the minors: players / units | Leftover newcomer units | All depth-chart time outside the 2026 majors | Projected individually | |
|---|---|---|---|---|---|
| Hitters (PA) | 21 / 2,981 | 80 / 5,299 | 70 | 8,350 | 99.2% |
| Pitchers (BF) | 44 / 9,127 | 112 / 9,266 | 216 | 18,609 | 98.8% |
Depth-chart time on the 30 clubs (unsigned free agents excluded), pull of 2026-11-16.
What is left is depth-chart time for players with no MLB, Triple-A or Double-A record since 2024 (overseas signings, jumps from A-ball): 70 PA and 216 batters faced across the league. Before the minor-league stage the same allowance covered about 8,400 PA and 18,900 batters faced (README.md: 8,441 and 18,869 on the 2026-11-14 pull). It counts at the measured newcomer level: what players who were not in MLB the season before actually produced, pooled over the season pairs 2025–2026, 2024–2025, 2023–2024 (the newcomers of 2024–2026), every component, runs over units:
newcomer level = sum of runs of players absent the season before / their PA (BF)
= −6.8 runs per 600 PA (hitters), −2.5 per 600 BF (pitchers)
This is not the newcomer level of the translation's back-test (Section 14: players with no MLB time in the three seasons before, offense only, measured on each test's own fit targets: hitters −7.36 (2025 test) and −7.37 (2026 test) per 600; pitchers −4.62 (2025 test) and −4.53 (2026 test) per 600). The group it is measured on also differs from the group it now values: the measured group includes regulars back from a lost season (Musgrove, Darvish and Steele were among them; README.md), who are now projected individually, while the remainder is overseas signings and A-ball jumps. At 70 PA and 216 batters faced the choice is immaterial: valuing the remainder at replacement instead would change the league total by −1.28 runs for hitters and −3.93 for pitchers.
The conventional alternative, replacement level, is far harsher. It follows the convention that a replacement-level team wins about 48 games (1,000 WAR a season, 57% to position players, 43% to pitchers), converted to runs at 2026's runs per win, both divided by the league's plate appearances (batters faced equal plate appearances league-wide):
rep_h = -0.57 * 1000 * rpw / lg_pa # runs per PA vs average replacement = −17.8 per 600 PA (hitters), −13.4 per 600 BF (pitchers)
Measured newcomers play well above replacement because the group includes regulars back from a lost season, not just call-ups. The dashboard's team popups show both values.
League centering
Runs against average add up to zero in every real season. The projections did not: on the 2026-11-16 pull the 30 depth charts summed to +1,402 runs (+46.7 a club; hitters +668, pitchers +734). The back-test at actual playing time shows the same lean and where it sits: regulars were projected about right (2026 bias +0.32 runs per 600 for hitters, +0.61 for pitchers), part-time players well above what they produced (hitters by actual PA, 2026: 0-99 PA: +35.4; 100-299 PA: +14.7).
How the sum got there (the run totals are from README.md): +2,127 runs before the minors went in (+71 a club) and +1,221 after (hitters +688, pitchers +534), when centering was added on 2026-11-14, a drop of 906. README.md splits the drop into the minor-league prior on 2026 players with minor-league time (−941) and the individual newcomers (−132) who replaced the flat allowance. The allowance itself was 8,441 PA x (−6.77) / 600 + 18,869 BF x (−2.47) / 600 = −173.0 runs (README.md's units, today's newcomer levels); README.md prints it as +168, a sign slip. So the newcomer swap added −132 − (−173) = +41 runs, and −941 +41 = −900, within rounding of the −906 drop: the minor-league prior is larger than the whole drop. The MLB pitch-quality prior then raised the pitchers' sum: average pitch quality sits slightly above zero and its larger k leans harder on the prior. Today's sum is the one above.
The fix moves the pull toward average. Each projection adds k units of league-average play to the player's history (Section 12). That added play now counts at a level m a little below average, one per component, solved so the depth-chart league, newcomers included, sums to zero:
rate(c) = ( sum_y w_y x runs_y(c) + k_c x m_c ) / ( sum_y w_y x units_y + k_c )
= uncentered rate + w_c x m_c , w_c = k_c / (U + k_c) (translated players: k_c / (U + K_c))
m_c = - ( sum_i rate_i(c) x units_i + N(c) ) / ( sum_i w_i(c) x units_i x mask_i(c) )
i = depth-chart players on the 30 clubs (pt = dc, team not FA), at depth-chart units
N = leftover newcomer runs at the measured level, counted in batting (hitters) or
strikeouts and walks (pitchers) only
mask = 0 for fielding of designated hitters and for catcher defense of non-catchers
m per 600, hitters: batting −3.6, situational outs −0.3, baserunning −0.5, fielding −1.0, catcher defense −3.6
m per 600, pitchers: strikeouts and walks −3.6, contact −1.1, running game +0.1, pitcher fielding 0.0
The shift w x m is then applied to every record, not just those in the solve: the 89 hitters and 124 pitchers who are unsigned free agents, and the 92 and 146 with no role, move by their own w x m too. Players with no MLB history have w = 1 in the parts that were not translated (fielding and catcher defense, pitcher fielding), so translated catchers carry −1.00 fielding and −3.62 catcher defense per 600 from centering alone. The newcomers stay at the measured level. A regular with years of history leans little on the added play: hitters with 400+ projected PA moved −2.05 per 600 on average (−6.05 to −1.35), those under 150 PA −2.83; pitchers with 400+ batters faced −1.87, under 150 −3.00; players up from the minors −3.47 (hitters) and −3.32 (pitchers). Because the shift follows each roster's mix, it is not the same for every club: from −53.0 runs (SD) to −42.7 (CLE), a spread of 1.1 wins. Teams leaning on part-timers and rookies give back the most.
| Depth-chart playing time or origin | Players | Mean shift / 600 | Range | |
|---|---|---|---|---|
| Hitters (PA) | 0-149 | 224 | −2.83 | −7.70 to −1.53 |
| Hitters (PA) | 150-399 | 168 | −2.51 | −6.35 to −1.51 |
| Hitters (PA) | 400+ | 277 | −2.05 | −6.05 to −1.35 |
| Hitters (PA) | in MLB in 2026 | 567 | −2.29 | |
| Hitters (PA) | back from a lost season | 22 | −2.26 | |
| Hitters (PA) | up from the minors | 80 | −3.47 | |
| Pitchers (BF) | 0-149 | 236 | −3.00 | −4.34 to −1.77 |
| Pitchers (BF) | 150-399 | 391 | −2.60 | −4.36 to −1.49 |
| Pitchers (BF) | 400+ | 202 | −1.87 | −4.52 to −1.26 |
| Pitchers (BF) | in MLB in 2026 | 666 | −2.37 | |
| Pitchers (BF) | back from a lost season | 51 | −3.04 | |
| Pitchers (BF) | up from the minors | 112 | −3.32 |
The centering shift (cen600) of every depth-chart player, free agents included, pull of 2026-11-16.
Free-agent drift. m is solved without the unsigned free agents, but they keep FanGraphs' projected playing time: 89 hitters with 27,410 PA and 124 pitchers with 43,584 batters faced, worth −352.7 and −336.3 centered runs (−7.72 and −4.63 per 600). When one signs, his runs join a club at the old m while depth charts take time from others, so the league drifts off zero by his runs less those of the players he displaces, until the next weekly refresh re-solves m. If every one of them signed and displaced players of exactly average value (zero runs), the league would move by −689.0 runs (−23.0 a club); displacing above-average players makes the drift larger, displacing below-average ones smaller. Wins are measured against the league mean, so the drift moves no club's projected wins until it changes the clubs' relative totals.
| Target | Uncentered league | All: none | All: uniform | All: prior-weighted | Regulars: none | Regulars: uniform | Regulars: prior-weighted | |
|---|---|---|---|---|---|---|---|---|
| Hitters | 2026 | +2.57 | 21.79 | 21.41 | 21.27 | 16.93 | 17.11 | 17.09 |
| Hitters | 2025 | +3.12 | 21.71 | 21.44 | 21.25 | 16.38 | 16.78 | 16.70 |
| Pitchers | 2026 | +1.91 | 16.20 | 16.09 | 16.12 | 13.13 | 13.20 | 13.34 |
| Pitchers | 2025 | +1.77 | 16.26 | 16.17 | 16.09 | 12.70 | 12.87 | 12.83 |
Runs per 600 PA (hitters) or batters faced (pitchers). Each projection is put at the player's actual playing time in the target season; 'uncentered league' is how far the projected league, newcomers at the measured level, came out above zero. Typical miss is weighted by playing time. All = everyone projected who played; regulars = 300+ PA or 250+ batters faced. 2026 is out of sample; 2025 uses the same settings, fitted on it. Uniform = one shift per unit for every player.
The back-test (forecast.centering_test) projects every target-season player who has MLB history in the two seasons before, with the forecast's test settings (k, split k and aging fitted for the 2026 test, pitch quality's test parameters), puts each projection at his actual playing time, values players without that history at the measured newcomer level of the seasons before (−11.47 per 600 PA and −2.30 per 600 BF for 2026), and solves m on that league. It contains no minor-league translation, which production centering does. Its m differs from production's (2026 test, hitters: batting −4.4, situational outs 0.0, baserunning −0.3, fielding −1.1, catcher defense −1.9).
Over everyone who played, prior-weighted centering lowers the typical miss in both seasons for hitters and pitchers, with paired intervals clear of zero against no centering for hitters 2026 (−0.52 [−0.79, −0.26]), hitters 2025 (−0.46 [−0.77, −0.14]); not for pitchers 2026 (−0.08 [−0.26, +0.12]), pitchers 2025 (−0.17 [−0.34, +0.01]) (table below). Against a single uniform shift it wins 3 of the 4 everyone-cells and 3 of the 4 regular cells; the cells it loses are pitchers 2026 everyone, pitchers 2026 regulars. The only out-of-sample pitcher cells are among them, so the choice of the prior-weighted form for pitchers rests on the in-sample 2025 cell and on the hitters. For regulars alone centering costs 0.13 to 0.32 runs per 600, and it leaves them under-projected: bias after centering hitters 2026 −2.25, hitters 2025 −3.39, pitchers 2026 −1.30, pitchers 2025 −1.92 (before: +0.32, −0.27, +0.61, −0.12). Over everyone the bias goes from +4.32, +3.39, +1.93, +1.74 to +1.47, −0.02, −0.21, −0.26 (same order: hitters 2026, 2025, pitchers 2026, 2025); it does not reach zero because the evaluated players' actual production differs from the newcomers' measured level. The regular group is chosen by what happened (players who play well keep their jobs), which flatters an uncentered projection, so the cost is probably overstated.
| Target | Actual PA / BF | Players | Units | None: bias / miss | Uniform: bias / miss | Prior-weighted: bias / miss | |
|---|---|---|---|---|---|---|---|
| Hitters | 2026 | 0-99 | 136 | 5,763 | +35.4 / 57.4 | +32.6 / 55.7 | +30.9 / 54.7 |
| Hitters | 2026 | 100-299 | 160 | 32,213 | +14.7 / 26.8 | +11.8 / 25.3 | +11.0 / 25.2 |
| Hitters | 2026 | 300+ | 249 | 128,003 | +0.3 / 16.9 | −2.5 / 17.1 | −2.2 / 17.1 |
| Hitters | 2025 | 0-99 | 145 | 5,873 | +33.7 / 56.4 | +30.3 / 54.4 | +28.3 / 53.3 |
| Hitters | 2025 | 100-299 | 145 | 26,541 | +15.3 / 30.1 | +11.9 / 28.5 | +10.9 / 28.3 |
| Hitters | 2025 | 300+ | 264 | 135,033 | −0.3 / 16.4 | −3.7 / 16.8 | −3.4 / 16.7 |
| Pitchers | 2026 | 0-99 | 214 | 8,903 | +22.5 / 37.5 | +20.3 / 36.2 | +19.4 / 35.7 |
| Pitchers | 2026 | 100-249 | 199 | 35,108 | +1.3 / 16.8 | −0.9 / 16.8 | −1.4 / 16.8 |
| Pitchers | 2026 | 250+ | 258 | 120,267 | +0.6 / 13.1 | −1.5 / 13.2 | −1.3 / 13.3 |
| Pitchers | 2025 | 0-99 | 225 | 9,451 | +17.3 / 36.6 | +15.3 / 35.7 | +14.5 / 35.3 |
| Pitchers | 2025 | 100-249 | 185 | 32,094 | +4.2 / 18.3 | +2.2 / 18.0 | +1.6 / 18.0 |
| Pitchers | 2025 | 250+ | 268 | 120,947 | −0.1 / 12.7 | −2.1 / 12.9 | −1.9 / 12.8 |
The same back-test by the player's actual playing time in the target season (runs per 600; bias = projected minus actual, weighted by playing time). Recomputed player by player by paper_checks_tail.py with the forecast's test settings; it reproduces the forecast's own figures above to the hundredth.
| Target | Players | n | Prior-weighted minus uniform (95% CI) | Prior-weighted minus none (95% CI) | |
|---|---|---|---|---|---|
| Hitters | 2026 | all | 545 | −0.14 [−0.24, −0.04], better in 99.8% of 2,000 resamples | −0.52 [−0.79, −0.26], better in 100.0% of 2,000 resamples |
| Hitters | 2026 | regulars | 249 | −0.02 [−0.15, +0.11], better in 62.2% of 2,000 resamples | +0.16 [−0.17, +0.49], better in 18.4% of 2,000 resamples |
| Hitters | 2026 | part-timers | 296 | −0.36 [−0.55, −0.18], better in 100.0% of 2,000 resamples | −1.79 [−2.14, −1.43], better in 100.0% of 2,000 resamples |
| Hitters | 2025 | all | 554 | −0.19 [−0.31, −0.07], better in 100.0% of 2,000 resamples | −0.46 [−0.77, −0.14], better in 100.0% of 2,000 resamples |
| Hitters | 2025 | regulars | 264 | −0.08 [−0.24, +0.08], better in 85.6% of 2,000 resamples | +0.32 [−0.07, +0.72], better in 5.9% of 2,000 resamples |
| Hitters | 2025 | part-timers | 290 | −0.45 [−0.68, −0.22], better in 100.0% of 2,000 resamples | −2.08 [−2.51, −1.61], better in 100.0% of 2,000 resamples |
| Pitchers | 2026 | all | 671 | +0.04 [−0.03, +0.10], better in 15.0% of 2,000 resamples | −0.08 [−0.26, +0.12], better in 78.6% of 2,000 resamples |
| Pitchers | 2026 | regulars | 258 | +0.14 [+0.04, +0.23], better in 0.4% of 2,000 resamples | +0.21 [−0.06, +0.48], better in 5.9% of 2,000 resamples |
| Pitchers | 2026 | part-timers | 413 | −0.12 [−0.22, −0.01], better in 98.5% of 2,000 resamples | −0.55 [−0.82, −0.26], better in 100.0% of 2,000 resamples |
| Pitchers | 2025 | all | 678 | −0.08 [−0.14, −0.01], better in 98.8% of 2,000 resamples | −0.17 [−0.34, +0.01], better in 96.7% of 2,000 resamples |
| Pitchers | 2025 | regulars | 268 | −0.04 [−0.13, +0.05], better in 81.3% of 2,000 resamples | +0.13 [−0.10, +0.39], better in 14.8% of 2,000 resamples |
| Pitchers | 2025 | part-timers | 410 | −0.14 [−0.23, −0.05], better in 99.9% of 2,000 resamples | −0.66 [−0.92, −0.43], better in 100.0% of 2,000 resamples |
Paired bootstrap over players of the difference in typical miss (2,000 resamples). Regulars = 300+ PA or 250+ batters faced; part-timers = the rest.
Team totals and projected wins
team runs = sum of projected (centered) runs of every player on the club's depth chart
(2026 major leaguers, players back from a lost season, players up from AAA / AA):
hitters bat + sit + br + fld + catcher defense, pitchers kbb + con + running game + fielding,
no positional adjustment
+ leftover PA x newcomer hitter level + leftover BF x newcomer pitcher level
wins = 81 + (team runs - league mean team runs) / RPW (the mean is zero after centering)
R = L["meta"]["lg_runs_per_pa"] * lg_pa / (2 * games) # runs per team-game
rpw = 4 * R / (2 * R) ** 0.287 # Pythagenpat runs per win
= 9.54 R = 4.476 runs per team-game in 2026 (2,429 games, the sum of the standings' wins)
Projected wins sum to 2,430 (30 x 81 by construction). On the 2026-11-16 pull the top clubs were LAD 95.7, MIL 93.3, BOS 92.3, NYY 90.4, SEA 85.5; the bottom were SD 73.0, CIN 72.0, COL 66.4. RPW is linear and the 2026 run environment is assumed for 2027. A club-specific Pythagenpat (each club's own runs scored and allowed, from its offense and its pitching and defense around the league's R) would move clubs by at most 0.74 wins. Team splits in the forecast follow the same rules as for actual seasons. Translated players get the typical platoon gap for their handedness; leftover newcomer hitters are assigned the league share of plate appearances against left-handers and leftover newcomer pitchers the league share of left-handed batters.
Team-level back-test. No team-level test existed until this paper (README.md records none). paper_checks_tail.py builds one from the centering back-test: the same player projections at each player's actual target-season playing time, split by his share of time with each club, centered, summed by club, and compared with the club's actual players runs total, run differential and wins:
| Target | Compared with | Model RMSE | All-average RMSE | Previous-season RMSE | Model minus all-average (95% CI) | Model minus previous season (95% CI) | r (95% CI) | SD projected / actual |
|---|---|---|---|---|---|---|---|---|
| 2026 (out of sample) | players runs total | 72.9 | 101.7 | 99.3 | −28.7 [−48.3, −9.2], better in 99.7% of 2,000 resamples | −26.4 [−50.8, −3.2], better in 98.6% of 2,000 resamples | 0.71 [+0.47, +0.85] | 57.9 / 101.7 |
| 2026 (out of sample) | run differential | 85.8 | 104.9 | 107.5 | −19.2 [−38.5, +1.6], better in 96.8% of 2,000 resamples | −21.7 [−48.2, +3.5], better in 95.5% of 2,000 resamples | 0.58 [+0.27, +0.78] | 57.9 / 104.9 |
| 2026 (out of sample) | wins | 8.38 | 10.85 | 10.77 | −2.47 [−4.37, −0.45], better in 99.1% of 2,000 resamples | −2.39 [−4.37, −0.26], better in 98.9% of 2,000 resamples | 0.64 [+0.36, +0.81] | 6.06 / 10.83 |
| 2025 (settings fitted on it) | players runs total | 91.9 | 130.2 | 89.7 | −38.3 [−54.8, −20.0], better in 100.0% of 2,000 resamples | +2.3 [−22.3, +23.3], better in 41.6% of 2,000 resamples | 0.80 [+0.62, +0.90] | 55.8 / 130.2 |
| 2025 (settings fitted on it) | run differential | 89.4 | 124.6 | 93.4 | −35.2 [−51.8, −17.0], better in 100.0% of 2,000 resamples | −4.1 [−36.5, +24.2], better in 61.5% of 2,000 resamples | 0.77 [+0.56, +0.88] | 55.8 / 124.6 |
| 2025 (settings fitted on it) | wins | 8.08 | 11.50 | 9.28 | −3.42 [−5.48, −1.05], better in 99.9% of 2,000 resamples | −1.19 [−3.55, +0.94], better in 85.9% of 2,000 resamples | 0.75 [+0.53, +0.87] | 5.89 / 11.50 |
Clubs (30) as the unit. Each target-season player's projection (the centering back-test's, prior-weighted) times his actual playing time, split by his share of time with each club, plus players without MLB history at the newcomer level; runs, except the wins rows. 'All-average' predicts zero runs (.500) for every club; 'previous season' predicts the club's own previous-season figure (wins scaled to the target's games). Intervals: bootstrap over clubs, paired; r with a Fisher interval.
Out of sample (2026) the projected club totals beat the two naive forecasts: the miss in wins is 8.4 (95% interval [6.5, 10.0]) against 10.8 for every club at .500 and 10.8 for last season's record, and the projected wins correlate with actual wins at r = 0.64 [+0.36, +0.81]. The projections are compressed: their spread is 6.1 wins against 10.8 actual, as regressed forecasts must be, and the actual players runs total rises 1.25 runs per projected run (slope through the origin). Uncentered, the club miss against the players runs total would have been 86.7 runs instead of 72.9. It is a lower bound on the error of a preseason forecast for one reason: it uses actual playing time, so it carries none of the depth charts' playing-time error. The missing minor-league translation is no such reason; in Section 14's back-test the translation lowers hitters' error, so adding it would more likely shrink this miss than grow it. Read as a normal error, a miss of 8.4 wins puts a 95% band of at least ±16 wins around each 2027 projection. A test with preseason depth charts needs archived preseason pulls; data/depthcharts_history/ starts with the 2026-11-14 pull, so the first such test is the 2027 season itself.
Top 100 rookies
The 2027 forecast season has a Top 100 rookies view: the rookie-eligible players with a depth-chart role (unsigned free agents with depth-chart playing time included), ranked by projected runs above replacement. That is runs against average (hitters with the position adjustment) plus the playing time valued at the replacement gap above, so a regular outranks a cameo; ranked against average alone, a rookie with ten innings would beat an everyday rookie a run below average:
runs above replacement = sum of his centered component runs (+ position adjustment for hitters)
- projected units x replacement level / 600
replacement level = −17.8 per 600 PA (hitters), −13.4 per 600 BF (pitchers), the same for starters and relievers
377 of the 422 rookie-eligible players in the forecast have a 2027 role. One replacement level serves starters and relievers alike; the usual convention gives relievers a higher replacement level, so the list leans toward starters through volume.
Eligibility follows MLB's rule: a player is a rookie unless, in earlier seasons, he exceeded 130 at-bats or 50 innings in the majors, or spent more than 45 days on an active roster during the period of the 26-player limit (Opening Day to August 31), injured-list time excluded (fetch_rookie_status.py, about 3 minutes, run by every weekly refresh). At-bats and innings come from the Stats API's season-by-season MLB stats (sport 1 only). Active days are counted only for players still under both stat limits. For every season with MLB's daily active rosters on disk (2015–2026, fetch_active_rosters.py, data/active_days_YEAR.json: every club, every day from its first game to August 31, about 28,000 requests, cached because finished seasons do not change) the count is exact. Only for seasons before 2015 is it rebuilt from the player's transaction log: on at a call-up, selected contract or waiver claim to an MLB club (CLW, CU, SE) or an "activated" status change by an MLB club; off at an option, designation, outright, release or retirement (DES, DFA, OPT, OUT, REL, RET), an injured-list, paternity, bereavement, restricted, suspended or military placement, or an assignment to the minors or to a rehab stint; spans are half-open (the day he leaves does not count) and clipped to Opening Day through August 31.
The method changed twice. The people endpoint's roster entries came first and cannot be used: they record organization stints, not active-roster days (Kyle Hurt shows "Active, Dodgers" continuously from 2023 through a dozen options), and they agreed with MLB's 2026 rookie list for 1,373 of 1,470 players. The transaction rebuild that replaced them agreed for 1,436 of 1,470 but missed players who make an Opening Day roster with no transaction after a spring-camp assignment (Peyton Pallette, a 2026 Rule 5 pick: 2 days counted against about 55 actual; Bubba Chandler and Sal Stewart: 0 against 159). Among 2026 rookies, 40 differed by more than 3 days and 17 crossed the 45-day line, so on 2026-11-15 the daily rosters replaced it (these figures are from README.md). Checked now against MLB's own 2026 rookie list, the rule agrees for 1,435 of 1,470 players who appeared that season. Of the 25 disagreements the check stores, 23 are players we count as eligible and MLB does not (Drew Anderson (44.33 IP, 39 days); Jhonathan Díaz (46.33 IP, 20 days)) and 2 the reverse (Caleb Kilian (48 days); Kyle Hurt (84 days)); with exact roster days the cause is not a missed transaction, and it is not established (injured-list accounting is the leading guess). So MLB's lists settle each player's history: a player whose last MLB season is in 2024–2026 and who was not on MLB's rookie list that season is not a rookie now, and the counts decide only the rest. The lists can remove eligibility, never grant it. Reasons for ineligibility in the forecast pool: 1,134 not on MLB's rookie list for his last season, 65 over 50 IP, 64 over 130 AB, 50 over 45 active days. Players with 35–45 counted days are flagged as close calls (30 of the eligible).
| # | Rookie | Age | Club | Pos | Proj. PA / BF | Runs vs avg | Runs above repl. |
|---|---|---|---|---|---|---|---|
| 1 | Kaelen Culpepper | 24 | MIN | SS | 602 | +3.7 | +21.5 |
| 2 | Walker Jenkins | 22 | MIN | CF | 490 | +4.3 | +18.8 |
| 3 | Kade Anderson | 22 | SEA | SP | 558 | +5.0 | +17.5 |
| 4 | Leo Bernal | 23 | STL | C | 434 | +1.2 | +14.0 |
| 5 | River Ryan | 28 | DET | SP | 530 | +1.9 | +13.7 |
| 6 | Aidan Miller | 22 | PHI | SS | 343 | +2.9 | +13.1 |
| 7 | Hagen Smith | 23 | CWS | SP | 437 | +3.2 | +12.9 |
| 8 | Ethan Salas | 20 | SD | C | 459 | −0.8 | +12.8 |
| 9 | Joshua Báez | 23 | STL | LF | 518 | −3.7 | +11.6 |
| 10 | Jonah Tong | 23 | NYM | SP | 349 | +2.4 | +10.2 |
The top 10 of the dashboard's Top 100 rookies view (2027 forecast season). Age on Opening Day, March 25, 2027, as on the dashboard.
Weekly refresh
refresh_forecast.py 2027 runs as a scheduled task every Monday morning (07:00 local, in the Claude desktop app; it runs only while the app is open, and a missed run fires at the next launch; README.md). In order it: copies the current depth-chart files to data/depthcharts_history/ under their pull date (4 files kept so far, the first from 2026-11-14); re-pulls the depth charts; reruns forecast.py; re-checks rookie eligibility for everyone in the forecast (fetch_rookie_status.py, about 3 minutes of Stats API calls); rebuilds the dashboard and this paper; and prepends to output/refresh_log.md what changed: players who changed clubs or signed, players newly on or off a depth chart, clubs whose projected wins moved by 0.5 or more, and the top five clubs. Any failed step exits non-zero and the scheduled task stops; otherwise it checks both pages before republishing them: the dashboard (output/run_ledger_artifact.html) exists, is over 1 MB, is pure ASCII and contains "2027F"; this paper (output/white_paper_artifact.html) exists, is over 30 KB, is pure ASCII and contains "BaseState Player Value Model White Paper" (read from the task's SKILL.md). The latest entry: "2026-11-16 07:11 refresh (depth charts 2026-11-15 -> 2026-11-16)". A full run took 4 minutes 18 seconds when last timed (README.md).
The refresh changes playing time, clubs and team totals. It also changes every player's rate slightly, because league centering re-solves m for the new depth-chart mix (in the refresh test of 2026-11-15 rates moved by at most 0.07 runs per 600; README.md). forecast.py itself refits k, the split k, the aging shapes and pitch quality's k, β and kq on every run, but from inputs that change only once a season (ledgers, pitch-quality scores), so the values come out the same; the pitch-quality scores and the minor-league translation are not recomputed (Section 18), and neither the paper's tests nor paper_checks_*.py are rerun by it. Because the paper is rebuilt every week, every pull-dependent number in Sections 14 and 15 is as of the pull named in the text; the archived pulls make each week's figures reproducible.
16.How this compares with other methods
Almost every ingredient of the model comes from earlier public work. What is particular to it is how the pieces are combined: one run currency for both sides of every plate appearance, books that balance against actual run differential, a tested rule for which situational play counts as skill, and a forecast whose settings are fitted on past seasons and checked on held-out ones.
What is borrowed
| Piece | Borrowed from | What the model changes |
|---|---|---|
| Run expectancy and event run values | Lindsey's run expectancy tables; Palmer's linear weights; Tango, Lichtman and Dolphin's The Book; FanGraphs wOBA and wRAA | Outs are valued by batted-ball type and reaching on an error is its own event (wOBA counts it as an out). Batting still agrees with FanGraphs at r = 0.985–0.991 in every full season 2015–2026 (Sections 4, 6). |
| Baserunning | Lichtman's UBR; FanGraphs BsR | What an average runner does is conditioned on the exit-velocity band of the ball, not just the event and state; rated per time on base and applied at the league-average number of chances (Section 6). |
| Fielding | Sports Info Solutions' DRS (non-catchers); Statcast framing, blocking and throwing (catchers) | Used as published. DRS was chosen because it tracks team batted-ball outcomes best (Section 7). |
| Positional adjustment | FanGraphs | Shown for comparing players, kept out of team sums (Section 7). |
| Pitcher contact quality | Statcast expected statistics (xwOBA, xERA) | Expected value of each batted ball from exit velocity and launch angle, in the hitters' run currency. It replaces FIP's actual home runs (Section 8). |
| Park factors | The usual home/road comparison | Runs per PA in the home club's games against its road games, two seasons, regressed 50% toward neutral, keyed by home club rather than venue (Section 6); the minor leagues use additive venue effects per batting piece (Section 14). |
| Platoon splits | The Book's regressed platoon gap | The overall rate is preserved, so the two splits add to the total (Sections 9, 12). |
| Forecast weights | Tango's Marcel (5/4/3 hitters, 3/2/1 pitchers) | Marcel's weights are kept; its fixed amounts of league-average regression (1,200 PA for hitters) are replaced by a k fitted per component on past targets (Section 12). |
| Reliability and regression to the mean | Tango's regression constant; split-half stabilization points (Carleton) | k = h (1 − r) / r from odd/even plate appearances, per minor-league piece and level, before the carry-overs are fitted (Section 14). |
| Pitch quality | Stuff+ and similar pitch-tracking models | Each pitch is valued in the model's own runs (the count-based change in expected strikeout and walk value, plus the contact value on the ball-in-play pitch). The stuff models are degree-2 polynomial ridge regressions per pitch group, batter hand and two-strike count. Three uses, scored differently: the pitchers' forecast prior, leave-season-out (each history season by models fitted on the seasons before the target other than itself; weight fitted, Section 12); a descriptive measure for every season 2015–2026 on the dashboard, by pitch type and by batter hand (each Hawk-Eye season by models fitted on the other Hawk-Eye seasons; 2015–2019 after translation to Hawk-Eye's scale and 2020 as is by models fitted on 2021–2026; Section 8); and Triple-A pitchers in the translation, scored by MLB models fitted on MLB seasons up to and including that season, so the same season's MLB pitches (not the Triple-A ones) are in the training data (Section 14). |
| Pitch tracking before Hawk-Eye | Public PITCHf/x and TrackMan data and the calibration problems known for both | 2015–2019 are put on the Hawk-Eye scale before scoring: a fitted PITCHf/x movement map, park recalibration, system bridges from the same pitchers, horizontal measurements per pitcher hand and 2017 bridged on its own; 2020 is Hawk-Eye already and is scored as is (Section 8). |
| Aging | The delta method | Shape only: the level shift in same-player deltas double-counts regression to the mean and is removed; the shape (straight line, quadratic or age-by-age average) is chosen per component by the held-out season (currently batting straight line, baserunning straight line, fielding straight line, pitchers strikeouts / walks quadratic, pitchers contact allowed straight line; Section 13). |
| Minor-league translation | Bill James's major-league equivalencies | Piece by piece in MLB runs, each piece shrunk by its own reliability, entered as the prior in the forecast's own formula; Triple-A pitchers also carry their MLB-scale pitch quality (Section 14). |
| Replacement level | The FanGraphs / Baseball-Reference convention (1,000 WAR a season, 57% to position players, 43% to pitchers) | Used only to rank the Top 100 rookies and as a comparison for leftover newcomer time (Section 15); every other value is against average. |
| Rookie eligibility | MLB's rule; MLB's rookie lists and daily active rosters (Stats API) | Rebuilt from MLB's own daily rosters, checked against MLB's lists, which settle a player's history (Section 15). |
| Playing time and teams | FanGraphs Depth Charts (RosterResource) | Playing time and club only; every rate is the model's (Section 15). |
| Runs per win | Pythagenpat | Turns projected team runs into wins, linearly at the last season's run environment (Section 15). |
| Splitting each plate appearance among batter, runners, pitcher and fielders | openWAR (Baumer, Jensen and Matthews, 2015) | The closest relative; the differences are below. |
What is different
- The books balance. Every plate appearance is valued once for the offense and once for the defense with the same run values. Every component sums to zero across MLB, players sum to teams, and whatever is not credited to a player is kept as an explicit difference from actual run differential (the % difference, Section 10), so team totals reconcile with it (players' totals track it at r = 0.88–0.98 across 2015–2026; Sections 10–11). WAR is built to sum to a fixed league total (1,000 wins above replacement), not to reconcile with what happened. openWAR also conserves runs, but it divides each play's actual change in run expectancy, so the timing of events stays in the players' values.
- Context is measured, then sorted. RE24 minus plain event values splits into detail (out types, reaching on error), timing (the same event valued in the state it came in) and the play (what happened beyond the event in that state). Most systems take one side: WAR's batting is fully context-neutral and treats all outs alike; RE24 and WPA keep everything. The model keeps detail, credits the repeatable part of the play layer to the player responsible, and leaves timing out. The rule comes from what repeats year to year across teams (Section 5): timing does not, double-play avoidance does, scoring the runner from third does not.
- Opportunities are neutralized. FanGraphs already credits double-play avoidance (wGDP, part of BsR) on each player's own opportunities. Here situational outs and baserunning are rated per opportunity and then applied at the league-average number of opportunities, so a hitter is neither helped nor hurt by how often the hitters ahead reach base (Section 6).
- No double counting across positions. Catcher framing is taken out of the pitchers' strikeouts and walks by batters faced; FanGraphs credits framing to catchers, and its FIP-based pitcher WAR, as far as we can tell, is not reduced for it. Pitchers get the expected value of the contact they allow; turning batted balls into outs belongs to the fielders, and home runs beyond what the contact predicted go to the % difference. Steals, pickoffs and wild pitches are split between pitcher and catcher (Section 8). Baseball Prospectus's DRA makes some similar adjustments for the catcher, through a mixed statistical model rather than an accounting split.
- The forecast is fitted, then tested, but not everything is out of sample. The forecast's own 2026 test uses k, the split k and pitch quality's k, β and kq fitted on 2025 only, so those are out of sample there; its aging shape, though, is picked by how well each shape predicts 2026, so that test is in sample for the aging choice. The final settings are refits that include the test seasons: k averages the 2025 and 2026 fits, and pitch quality's settings are pooled over the 2023–26 targets (final β 0.7 / 0.5). Any test that borrows final settings is therefore partly in sample: the minor-league translation's tests use the final k (Section 14, where its carry-overs, discounts and scales are fitted on earlier targets only), the production-path checks of Section 14 use the final aging and pitch-quality settings, and the centering test's 2025 cell uses settings fitted on 2025 (Section 15). Sections 12–14 say which is which. Centering the league to zero is prior-weighted, so part-timers move more than regulars; against uniform scaling it wins 3 of 4 everyone-cells and 3 of 4 regular cells, but the out-of-sample pitcher cells go to uniform (everyone 16.12 against 16.09; regulars 13.34 against 13.20; Section 15). Inputs that did not help, prospect grades, High-A and Single-A stats and pitch location, were tested and left out rather than assumed in (Sections 12, 14).
- One pitch-quality scale across three tracking systems. The dashboard's pitch quality for 2015–2019 is not a separate model (2020 is Hawk-Eye already): each older season is translated into Hawk-Eye's measurements from the same pitchers' changes across each system boundary and then scored by the Hawk-Eye models, with the calibration steps tested against season-to-season repeatability and against results (Section 8).
What other systems do that this one does not
- A replacement level. Values are against average, so there is no credit for volume alone; that is why FanGraphs WAR is shown beside the runs rather than replaced by them (Section 11). Only the Top 100 rookies list ranks by runs above replacement.
- Pitch-level valuation in the player values. DRA and pitch-level run values credit each pitch (count, location, umpire). The model computes count-based pitch values in its own currency, but only as the target of the stuff models; the player and team values still start at the batted ball and the plate-appearance result, and pitch tracking enters the forecast (as the pitchers' prior) and the dashboard's descriptive pitch quality, not the runs (Sections 8, 12).
- Nonlinear pitch models and more pitch features. Public Stuff+-type models use gradient boosting and features such as approach angle, arm angle and spin axis; this model uses polynomial ridge regressions on velocity, induced vertical break, horizontal break, spin, extension, release height and side, and the gaps from the pitcher's main fastball.
- Leverage. Win-probability systems weight late, close plays more; the model counts a run as a run.
- Projection ranges and playing-time models. ZiPS and PECOTA publish percentile ranges and model injuries and playing time; this forecast is a point projection with FanGraphs' playing time taken as given. Its team-level error is measured once, at actual playing time (Section 15: 8.4 wins in 2026).
- A head-to-head accuracy test. The forecast has not been compared with Steamer, ZiPS, THE BAT or FanGraphs Depth Charts on the same players and seasons; the only outside comparison is the r between FanGraphs' projected WAR and this forecast's team runs (Section 11).
- A batted-ball prior for hitters. Pitchers' forecasts lean on pitch quality; hitters' forecasts use results only (exit velocity and launch angle enter the minor-league translation's pieces, not the MLB hitter forecast).
- An own fielding model, and park factors by venue and by handedness (Section 17).
"Different" here means we know of no public system that combines these choices this way; each piece may have precedents beyond those named.
17.Limitations and next steps
- Fielding is borrowed. Individual fielding comes from DRS, not from the play-by-play, because the feed has no spray angles or fielder positions. A tracking-based fielding model built on Statcast data would remove the dependence and the partial circularity of validating DRS against outcomes.
- One park factor per home club, not per park.
ledger.park_factorskeys runs per PA by the home club's id over two seasons and regresses 50% of the way back to neutral; the MLB plate-appearance tables carry no venue. A club that changed parks (the Athletics' move to Sacramento in 2025, the Rays' 2025 season at Steinbrenner Field) gets a factor mixed across two parks for two seasons, and neutral-site games count as the home club's park. That is the trap the minor-league ledger avoids by keying on venue (Section 14). Parks also affect left- and right-handed hitters differently (short right fields, for example); venue- and handedness-specific park factors would sharpen batting and its splits. - Expected contact ignores spray angle and sprint speed. A pulled fly ball and an opposite-field fly ball with the same exit velocity and launch angle get the same expected value, as do grounders from fast and slow runners.
- Pitch quality is not part of the run values. It is the pitchers' forecast prior and, on the dashboard, a descriptive measure for every season 2015–2026 (by pitch type and by batter hand), but it never enters the runs. The models are one degree-2 polynomial ridge regression (λ = 1, intercept unpenalized) per pitch group (fastball, cutter, breaking, offspeed) x same or opposite batter hand x two strikes or not, separately for strikeouts / walks and contact, on velocity, induced vertical break, horizontal break, spin rate, extension, release height, release side (horizontal break and release side mirrored for left-handers) and, for non-fastballs, velocity gap, induced-vertical-break gap, horizontal-break gap from the pitcher's main fastball, plus count indicators. Hawk-Eye covers 2020–2026; the stuff models are fitted on the 2021–2026 pitch files only. They know nothing about command beyond what results already show: plate location was tested and added nothing for forecasting (pooled leave-one-season-out typical miss 13.06 without pitch quality, 12.95 with stuff, 12.96 with stuff and location). Its value moves from season to season (by target, without → with stuff: 2023: 13.71 → 13.46; 2024: 12.43 → 12.43; 2025: 12.77 → 12.89; 2026: 13.32 → 12.98), so its weight is fitted on all four seasons together. Minor leaguers get it only from Triple-A (fully tracked from 2023, about a third of 2022), through the translation; Double-A and High-A have no pitch tracking, so a pitcher coming straight from Double-A has none (34 of the 112 2027 pitchers with no MLB time), and Triple-A pitch quality is not split by batter hand. The dashboard's 2015–2019 pitch quality comes from older tracking systems put on the Hawk-Eye scale (2020 is Hawk-Eye already and scored as is; Section 8): 2015–16 without spin or extension; horizontal break and release side bridged per pitcher hand, because a calibration bias in the catcher's frame flips sign under the left-handers' mirroring; and 2017 bridged on its own, because TrackMan was recalibrated after it. Season-to-season repeatability (strikeouts and walks / contact, pitchers with 150+ batters faced in both seasons) is 2015–2016 0.893 / 0.885; 2016–2017 0.848 / 0.798; 2017–2018 0.885 / 0.828; 2018–2019 0.882 / 0.873; 2019–2020 0.886 / 0.813; 2020–2021 0.904 / 0.815, against 0.886–0.912 / 0.855–0.869 for the Hawk-Eye pairs from 2021–2022: 2016–17 contact repeats clearly less, 2017–18 sits at the bottom of the Hawk-Eye range, and the 2020–21 pair, across the 60-game season, is lower still in contact. Staff pitch quality agrees with staff results at r = 0.38 [+0.14, +0.58] under PITCHf/x, 0.62 [+0.47, +0.73] under TrackMan and 0.65 [+0.56, +0.72] under Hawk-Eye; only PITCHf/x differs from Hawk-Eye (z test p = 0.014; TrackMan p = 0.68; all older seasons pooled against Hawk-Eye p = 0.13).
- Season splits are unregressed. They describe what happened; only the forecast regresses them.
- Aging is one curve per component for everyone. It does not separate power hitters from speed players or starters from relievers, and does not yet use early-warning skill signals (bat speed, fastball velocity, sprint speed).
- Baserunning and fielding aging is weakly identified (very small back-test gains).
- Playing time follows one source and is not tested. Depth Charts is a single editorial view of rosters; injuries and depth changes between weekly pulls are not reflected. It plans full seasons for regulars that injuries and slumps then cut: on the current pull 188 hitters are projected for 502+ PA and 50 pitchers for 162+ innings, against 129–145 and 39–58 actual qualifiers a season in 2021–2026 (2020 left out), and its projected WAR on the 30 clubs sums to 1,159 against 1,000 in a played season. Team totals and the centering solve inherit that. Rows without an MLBAM id are dropped at the fetch (1 hitter with 623 PA and 2 pitchers with 1,054 batters faced on 2026-11-16 18:11, all unsigned today). Unsigned free agents are left out of the centering solve but keep their playing time (−689.0 runs between them), so signings move the league off zero until the next refresh (Section 15).
- Minor-league translations use stats only. FanGraphs Future Value and tool grades were tested and added nothing beyond the stats, age and level (Section 14). The test covers only players who reached MLB; grades may still say who gets there, and they may matter for projections that start below Double-A. High-A and Single-A stats exist and are built with the same model; they simply did not help next-season projections.
- Only Triple-A and Double-A in the forecast. High-A and Single-A are built and were tested as extra levels; they did not improve next-season projections (Section 14) and would matter only for projections further ahead. The minors have no DRS, so translated players carry no fielding value from data, only the centering level.
- Survivorship in the translation. It is fitted on players who reached MLB, weighted by their MLB playing time, so it answers how a player will do if he plays. Call-ups with no MLB time who reached 100 PA or batters faced out-performed it by +6.1 runs per 600 (hitters) and +5.1 (pitchers), because those who struggle lose their playing time.
- The translation's back-test is not the production path. Its comparison arm and fixed part use the forecast's final k without the pitch-quality prior those k values were fitted with (pitchers' "before" 14.22 as published, 13.99 with the prior); production adds the prior and aging on top of the fitted translation, a path the back-test never scored until paper_checks_tail.py (Section 14); and k is fitted with the test seasons in the sample. Fitting the translation with the prior in its fixed part and with aging would close the gap.
- Centering fixes the level, not the order. It makes the projected league sum to zero by setting the pull toward average a little below average, which moves part-timers more than regulars. It does not say which part-timers are overrated, and in the back-test it costs regulars 0.13 to 0.32 runs per 600 and leaves them under-projected by 1.3 to 3.4.
- Projected wins have only a lower-bound error. The team-level back-test misses by 8.4 wins (Section 15). Because it uses actual playing time it carries none of the depth charts' playing-time error, so a preseason forecast misses by more; its lack of the minor-league translation is not part of that argument (the translation lowers hitters' error). No projection ranges are published.
- Few test seasons. The forecast's own test, the translation's tests and the centering test score 2025 and 2026 only (minors-only groups of 99 hitters and 105 pitchers); the pitch-quality prior's leave-one-season-out test covers 4 target seasons (2023, 2024, 2025, 2026); the aging shapes are scored on 2026 alone.
- Rookie eligibility disagrees with MLB's own 2026 list for 35 of 1,470 players even with exact daily rosters; the cause is not established.
- Two-way players are valued separately as a hitter and as a pitcher; their combined value is the sum of the two rows.
- 2015 has no Statcast catcher data, and 2020's short season is excluded from aging and noisier everywhere.
- The weekly refresh runs only while the desktop app is open (a missed run fires at the next launch).
Next steps
- Use each pitcher's stuff platoon edge as the prior for his platoon gap, in place of the typical gap for his hand. The stuff edge repeats at 0.79 (strikeouts and walks) and 0.85 (contact) on average over season pairs against 0.26 and 0.16 for the actual edge, and predicts next season's actual edge at r = 0.26 / 0.28 against 0.26 / 0.16 (Section 8). The forecast now regresses a pitcher's gap toward his hand's typical gap with k_gap = 900.
- Refit the translation with the MLB pitch-quality prior in its fixed part and with aging, so the back-test scores the production path; widen the grids where the fit sits on an edge (14 parameters, Section 14).
- Park factors by venue and by batter hand.
- A team-level test with preseason depth charts, from the archived pulls, after the 2027 season; projection ranges from the test errors.
- A batted-ball (exit velocity, bat speed) prior for hitters, tested like pitch quality.
- A head-to-head test against public projection systems.
- A tracking-based fielding model.
18.Appendix: scripts, data and constants
Scripts (run from run_value_model/scripts)
| Script | Does | Writes |
|---|---|---|
fetch_pbp.py YEAR [aaa|aa|hia|a] | Pulls a season of play-by-play (16 threads, slimmed fields); with a level, the minor-league feed (sports 11–14), keeping the venue | data/pbp_YEAR.jsonl.gz, data/milb/pbp_LEVEL_YEAR.jsonl.gz |
pa_table.py YEAR [LEVEL] | Rebuilds every plate appearance, runner and runner event; runs the three reconciliation checks (printed, not stored) | data/pa_, advance_, runner_events_YEAR.csv.gz, data/milb/..._LEVEL_YEAR.csv.gz |
gap_decomp.py YEAR | Run expectancy, LW0/LW1/LW2, offense layers by team | output/gap_decomp_YEAR.json |
brewers_detail.py 2025 2026 | The Brewers study (Section 5): persistence, error-scorer check, situations; printed | screen output |
fetch_fg_team.py YEARS | FanGraphs team batting, pitching and fielding totals (curl) | data/fg_team_{bat,pit,fld}_YEAR.json |
prevention_decomp.py YEARS | Run-prevention layers by team; expected value of contact | output/prevention_YEAR.json |
fetch_fg_players.py YEARS | FanGraphs player batting, fielding and pitching rows (DRS and parts, Statcast fielding, positional, WAR); 3 s between files by default, 20 s with --only bat,pit (the 1990–2014 career rows), and back-off on HTTP 429 | data/fg_players_*_YEAR.json |
ledger.py YEAR | The player and team run values (needs YEAR and YEAR−1 play-by-play) | output/ledger_YEAR.json |
milb_ledger.py LEVEL YEAR | all | Minor-league run values in MLB runs against each league, park-adjusted by venue, with split-half sums | output/milb/ledger_LEVEL_YEAR.json |
milb_translate.py | Reliability, translation fit and back-test, 2027 minor-league projections (switches below) | output/milb_translation.json, output/milb_projections_2027.json |
milb_grades_test.py | Prospect grades as additions to the translation: 8 test fits in parallel | output/milb_grades_test.json, fits in output/milb/grades_test/ |
milb_levels_test.py [--compare] | High-A / Single-A as two more levels: 2 test fits | output/milb_levels_test.json, fits in output/milb/levels_test/ |
milb_pq_test.py | The translation with and without Triple-A pitch quality | output/milb_pq_test.json, output/milb_translation_{pq,nopq}.json |
fetch_pitches.py YEARS [aaa] | Every MLB pitch with its tracking (velocity, movement, spin, extension, release, location); before 2021 into data/pitches/early/ so pitch_quality.py cannot read it; with aaa, Triple-A | data/pitches/pitches_YEAR.csv.gz, data/pitches/early/, data/pitches/milb/ |
pitch_quality.py scores | Pitch values, the stuff models, each pitcher's pitch quality: forecast scores leave-season-out, display scores, Triple-A scores (once a season) | output/pitch_quality_scores.json, caches pq_YEAR.npz |
pitch_quality.py | The pitch-quality back-test: stuff vs stuff plus location, two protocols, bootstrap | output/pitch_quality_test.json |
pitch_types.py | Pitch quality by pitch type and by batter hand for 2021–2026, scored as the cards' season totals are (types and sides add up to them) | output/pitch_types.json |
pitch_quality_early.py | 2015–2020 put on the Hawk-Eye scale (movement map, park recalibration, system bridges, per pitcher hand for horizontal measurements, 2017 bridged on its own) and scored for the dashboard only, with types and hands | output/pitch_quality_early.json, caches data/pitches/early/raw_YEAR.npz |
pq_platoon_staff.py | The platoon edge in pitch quality and staff pitch quality against results; the park-recalibration test; the pitch-group cells | output/pq_platoon_staff.json |
opening_day.py | Opening Day dates 1990–2027 and age_on: the age used everywhere (aging, projections, the translation's age term, the dashboard) | module |
fetch_depth_charts.py 2027 | FanGraphs Depth Charts playing time and teams (rows without an MLBAM id are dropped) | data/fg_depthcharts_{bat,pit}_2027.json |
forecast.py 2027 | Fits k, splits and aging by back-test; projects players; centering; team totals | output/forecast_2027.json |
fetch_active_rosters.py YEARS | MLB's daily active rosters, every club, first game to August 31 (12 threads, cached) | data/active_days_YEAR.json |
fetch_rookie_status.py 2027 | Rookie eligibility entering the season for everyone in the forecast, and the check against MLB's list (run by the weekly refresh) | data/rookie_status_2027.json |
fetch_graduates.py | Graduates by season for the dashboard's Segmentation filter | data/graduates_*.json |
fetch_fg_prospects_2015_2016.py | Prospect grades for 2015–2016 from FanGraphs' articles (cached) | data/prospects/fg_reports_2015_2016.json |
build_dashboard.py | Dashboard for 2015–2026 and the 2027 forecast, with FanGraphs WAR, pitch quality and the segment pools (MLB qualifier and rookie lists) | MLB_Run_Ledger.html, output/run_ledger_artifact.html, data/segments_YEAR.json |
pinch_fix_history.py BEFORE_DIR [TABLES_DIR] | Once, 2026-11-16: compares the outputs before the pinch-runner fix (snapshots) with the rebuilt ones; the snapshots were not kept, so it cannot be rerun | output/pinch_fix_history.json |
auto_fix_history.py SNAP_DIR | Once, 2026-11-16: compares the outputs before the automatic-runner fix (snapshots) with the rebuilt ones, and probes the play-by-play for the runner_placed event (that part reruns from the play-by-play alone) | output/auto_fix_history.json |
fc_fix_history.py SNAP_DIR | Once, 2026-11-16: compares the outputs before the fielder's-choice fix (snapshots) with the rebuilt ones | output/fc_fix_history.json |
paper_checks_core.py | Reproducible checks behind Sections 1–8 (reconciliation rerun, contact cells, runner replica, year-to-year intervals, Brewers tables and more) | output/paper_checks_core.json |
paper_checks_forecast.py | Reproducible checks behind Sections 9–13 (forecast replication, aging, pitch-quality and split tests, validation) | output/paper_checks_forecast.json |
paper_checks_tail.py [--rerun-tests] | The bootstraps, intervals, defect measurements, centering bands, team-level back-test and other figures Sections 14–18 quote; reruns the grades and levels tests on request | output/paper_checks_tail.json |
white_paper_core.py, white_paper_forecast.py, white_paper_minors.py, white_paper_tail.py, white_paper_pq.py, white_paper_tips.py | Sections 1–8; Sections 9–13; Section 14 and the rookies text; Sections 15–18; Section 8's pitch-quality subsections; the tooltips | this paper |
build_white_paper.py | This paper, with its tables computed from the outputs | MLB_Run_Ledger_White_Paper.html, output/white_paper_artifact.html |
refresh_forecast.py | Weekly: depth charts, forecast, rookie status, dashboard, paper, change log | output/refresh_log.md, data/depthcharts_history/ |
Data and records
| File | Holds | Source |
|---|---|---|
output/pitch_quality_early_history.json | The season-to-season checks of all three versions of the 2015–2020 calibration (pooled offsets, 2017-2019 one system; mirrored measurements per pitcher hand; + 2017 its own bridge (published)), plus platoon_kbb_stuff_edge, staff_system_r, cells_version2, version1_pfx_fastball_hb_bridge, version1_offsets; read by white_paper_pq.py | Assembled by hand on 2026-11-16 from runs of the three versions; no script writes it (checked: none of the scripts opens it for writing), so it cannot be regenerated |
output/pinch_fix_history.json | The pinch-runner defect as measured before the 2026-11-16 fix (season by season), the tables' row counts then, and the fix's measured effects on the ledgers, the 2026 tables, the translation test and the 2027 forecast; read by white_paper_core.py | pinch_fix_history.py, run once on the pre-fix snapshots, so it cannot be regenerated |
output/auto_fix_history.json | The automatic-runner defect as measured before the 2026-11-16 fix, the runner_placed probe, every MLB and minor-league table compared file to file, and the fix's measured effects on the MLB and minor-league ledgers, the translation test and the 2027 forecast; read by white_paper_core.py | auto_fix_history.py, run once on the pre-fix snapshots |
output/fc_fix_history.json | The fielder's-choice defect as measured before the 2026-11-16 fix and its measured effects on the MLB and minor-league ledgers, the translation and the 2027 forecast; read by white_paper_core.py | fc_fix_history.py, run once on the pre-fix snapshots |
data/standings_YEAR.json | Wins and losses (team totals' games and runs per win; the team-level back-test) | Stats API, fetched once |
data/milb/teams_LEVEL_YEAR.json | Minor-league clubs with league and parent organization | Stats API |
data/prospects/fg_board_2021_2026.json, fg_graduates_2021_2026.json, fg_reports_2017_2020.json, fg_reports_2015_2016.json | FanGraphs prospect lists and reports (grades only) | Read in a browser (Cloudflare); clipboard transfer with a UTF-8 locale |
data/active_days_YEAR.json, data/rookie_status_2027.json, data/segments_YEAR.json | Daily-roster days, rookie status and the check, MLB's qualifier and rookie pools | Stats API |
data/depthcharts_history/ | Every earlier depth-chart pull, by date | refresh_forecast.py |
output/milb/grades_test/, output/milb/levels_test/, output/milb_translation_{pq,nopq}.json | The test fits behind Section 14's tests, with their per-player rows | the test scripts |
output/paper_checks_core.json, output/paper_checks_forecast.json | The checks behind Sections 1–8 and 9–13 | paper_checks_core.py, paper_checks_forecast.py |
output/paper_checks_tail.json | Section 14–18 checks; its history block keeps the earlier grades and levels test results | paper_checks_tail.py |
studies/pitchability/ | A separate blog study that only reads the model's data; outside the model and this paper | its own scripts |
Run order
from scratch (once): forecast.py 2027 -> milb_translate.py -> forecast.py 2027
(milb_translate reads the forecast's final k; forecast imports milb_translate
inside add_minors and skips the minors if the translation file is missing)
once a season (after it ends; YEAR = the season just finished):
1 data fetch_pbp.py YEAR and pa_table.py YEAR, for MLB and each level (aaa, aa, hia, a);
fetch_fg_players.py YEAR (ledger.py reads its DRS, fielding parts, positional
adjustment and WAR); fetch_fg_team.py YEAR; fetch_pitches.py YEAR and YEAR aaa;
fetch_active_rosters.py YEAR
2 values gap_decomp.py YEAR; prevention_decomp.py YEAR; ledger.py YEAR; milb_ledger.py all
3 pitches pitch_quality.py scores -> pitch_types.py -> pitch_quality_early.py ->
pq_platoon_staff.py; pitch_quality.py (the back-test; LOSO=0 for the in-sample copy)
4 forecast forecast.py 2027 -> milb_translate.py -> forecast.py 2027
(milb_translate.py reads the forecast's final k, so the forecast runs on both sides)
5 tests milb_pq_test.py; milb_grades_test.py; milb_levels_test.py; fetch_rookie_status.py 2027;
fetch_graduates.py
6 checks paper_checks_core.py; paper_checks_forecast.py; paper_checks_tail.py
(without them the paper shows its "Stale checks" notes)
7 pages build_dashboard.py -> build_white_paper.py
weekly (scheduled): refresh_forecast.py 2027 (depth charts, forecast.py, fetch_rookie_status.py,
dashboard, paper; forecast.py refits k, k_gap, aging and pitch quality's
k / beta / kq from unchanged inputs, so the values do not move)
after a test change: the test script, then paper_checks_*.py -> build_white_paper.py
Environment switches
| Script | Variable | Default |
|---|---|---|
forecast.py | AGING | relative |
milb_translate.py | LEVELS | aaa,aa |
milb_translate.py | MILB_PQ | 1 |
milb_translate.py | MIN_EVAL | 100 |
milb_translate.py | WCAP | 0 |
milb_translate.py | GRADES | none |
milb_translate.py | TAG | |
pitch_quality.py | LOSO | 1 |
Read from each script's source. milb_translate.py: LEVELS = the levels fitted (aaa,aa,hia,a for the four-level test); MILB_PQ = 0 leaves Triple-A pitch quality out; GRADES = none, fv, fvw, tools or fvonly (none, fv, fvw, tools, fvonly); TAG writes milb_translation<TAG>.json and stops (tests); WCAP caps each player's weight; MIN_EVAL is the evaluation minimum (one value for hitters and pitchers). forecast.py: AGING=relative keeps only the shape of the aging curves. pitch_quality.py: LOSO=0 scores every season with models fitted on it (the in-sample comparison) and writes output/pitch_quality_test_insample.json, which the Section 12 checks read. paper_checks_tail.py takes flags, not variables: --rerun-tests, --rerun-only, --snapshot-only, --no-fetch.
Key constants
| Constant | Value | Where it comes from |
|---|---|---|
| Run expectancy table | per season, innings 1–8 | gap_decomp.decompose (Section 4) |
| Park factor regression | 50% of the way to 1.00, two seasons, by home club | ledger.park_factors |
| Advancement cell minimum | 30 runners (else drop the exit-velocity band) | ledger.build (Section 6) |
| Contact cell minimum and bins | 25 balls; 3 x 4 (mph x degrees), then 6 x 8 | milb_ledger (as the MLB contact model, Section 8) |
| Forecast weights | hitters 5/4/3, pitchers 3/2/1 | forecast.W (Marcel) |
| k grid | 0, 25, 50, 100, 150, 200, 300, 400, 500, 650, 800, 1,000, 1,250, 1,500, 2,000, 2,500, 3,000, 4,000, 5,000, 7,000, 10,000 | forecast.K_GRID |
| Final k | hitters bat 3,250; hitters sit 7,000; hitters br 4,000; hitters fld 3,250; hitters cat 1,500; pitchers kbb 1,250; pitchers con 5,000; pitchers run 3,000; pitchers fld 5,500 | forecast_2027.json report (Section 12) |
| Split gap k | hitters 3,500, pitchers 900 | forecast_2027.json report |
| Fit minimum | 100 units in the target season | forecast.MIN_FIT |
| Split share prior | 200 PA at the league share for the hand | forecast.project |
| Pitch-quality grids | k = k grid + 15,000, 20,000, 30,000; beta 0 to 2 by 0.1; kq 0, 50, 100, 200, 400, 800, 1,600, 3,200; first target 2023 | forecast.PQ_K, PQ_BETA, PQ_KQ, PQ_FROM |
| Pitch quality (final) | kbb: k 1,250, beta 0.7, kq 400; con: k 5,000, beta 0.5, kq 50 | fitted pooled over 2023–2026 targets |
| Stuff models | ridge λ 1; groups fastball, cutter, breaking, offspeed; features velocity, induced vertical break, horizontal break, spin rate, extension, release height, release side (+ velocity gap, induced-vertical-break gap, horizontal-break gap) | pitch_quality |
| Aging | 200+ units in both seasons; seasons from 2015; 2020 skipped; binned curve ages 22–38; applied at the mid age clamped to 20 to 40 | forecast |
| Evaluation and centering-test minimums | 300 PA / 250 batters faced | forecast.centering_test |
| Depth-chart starter label | starter = gs >= 8 or (g and gs / g >= 0.5) | forecast.main |
| Replacement level | 1000 WAR, 57% hitters / 43% pitchers; −17.8 / 600 PA, −13.4 / 600 BF | forecast.team_totals |
| Runs per win | 9.54 (exponent 0.287) | Pythagenpat, Section 15 |
| Newcomer level | −6.8 / 600 PA, −2.5 / 600 BF | Measured, Section 15 |
| Minor-league venue effects | two seasons, regressed by n / (n + 10,000 PA) | milb_ledger.PARK_N0 |
| Minor-league reliability | split-half k = h(1 − r) / r, 100+ per half, k in [20, 20,000] | milb_translate.reliability |
| Translation grids | d 0.25, 0.5, 0.75, 1, 1.5, 2, 3; Double-A scale 0.4, 0.55, 0.7, 0.85, 1; lower levels 0.1, 0.2, 0.3, 0.4, 0.55, 0.7, 0.85, 1; single-piece k 300, 1,000, 3,000, 10,000 | milb_translate |
| Translation age and reporting | age centered at 26, clamped 19–33; full season 450; evaluation 100+ MLB units | milb_translate |
| Triple-A pitch-quality centering | 5,000+ tracked batters faced per level-season | milb_translate.attach_pq |
| Rookie limits | 130 AB, 50 IP, 45 days; daily rosters from 2015 | fetch_rookie_status |
| Refresh change log | win swings of 0.5+ | refresh_forecast.py |
| Bootstraps | paper_checks_tail.py 2,000 paired resamples, seed 20261116; paper_checks_forecast.py 4,000, seed 20261116; paper_checks_core.py 2,000 (player-clustered, seed 7; 1,000 in some year-to-year calls) and 4,000 (paired differences, seed 11); milb_pq_test.py 4,000, seed 7; pitch_quality.py 4,000, seed 7 | each script's source |
| Pitcher split share | strikeouts-and-walks share of each throwing hand's typical gap: L 0.279, R 0.522, S 0.500; clamped to [0, 1]; 0.5 if |gap| x 600 < 1 or either side has under 1,000 batters faced | forecast.league_gap_shares (2026-11-16) |
Values are imported from the modules named, or read from the literal in the function named, when the paper is built; fitted values come from the outputs.
Traps found while building
- FanGraphs' RE24 is park-adjusted and its wRAA is not; never subtract them.
- FanGraphs' player exports lump traded players under "- - -"; team totals must come from the team endpoint or be allocated by playing time.
- In the FanGraphs fielding API,
fFRPis framing andcFRPis first-base receiving. - FanGraphs refuses Python's HTTP client; use curl. Its prospect Board and minor-league leaderboards sit behind a Cloudflare challenge (read in a browser); a fast run of player pulls gets HTTP 429.
- The team endpoint needs
team=0%2Cts; an unencoded comma returns players. - Split batted-ball residuals by home-run probability, never by the actual result; key the minor-league trajectory cells on the recorded trajectory, never on the result.
- Apply a play's runner movements as one batch, and let a moving runner clear his starting base (pinch runners).
- Write a plate appearance's runner rows only after putting each pinch runner's id on his base, from the play's movers and the feed's postOn ids: the tracker does not see substitutions, and until 2026-11-16 a pinch runner's first move on a plate-appearance-ending play was written as the replaced runner staying put. The occupancy check cannot catch a wrong id.
- Detect the automatic runner from the data (2020–21 seven-inning doubleheaders), and seat him by the id of the
runner_placedevent on the half-inning's first play: until 2026-11-16 he was a placeholder whose stays before his first move were never written. Once (2023-10-01) the feed ends a run at "4B". gap_decomp.decompose()rewrites rows in place; give copies to anything else that reads them.- The feed's
fielders_choiceretires nobody (fielders_choice_outretires one), but LW1 classes it as an out by its trajectory: never hand its play residual to the batter. Until 2026-11-16 the ledgers did (about +400 runs a season to batters' situational outs). - Until 2026-11-16
ledger.fetch_peoplerewrotedata/people.jsonon every run, in place, so two seasons' ledgers run in parallel could read it half-written (a JSONDecodeError in 2021 and 2022); it now writes only when it fetched someone, to a temporary file renamed over the cache. - The Board's level fields (mlevel, llevel) are today's level, not the level at list time; pbpaste needs a UTF-8 locale.
- The people endpoint's rosterEntries are organization stints, not active-roster days; the transaction log misses Opening Day rosters with no transaction after camp. Use the daily rosters.
- The Stats API pitch count is the count after the pitch; breakHorizontal and x0 change sign with the pitcher's hand, so a catcher-frame calibration bias flips sign under mirroring; a measurement a baseline lacks must not be averaged as zero.
- Scoring a season with a model fitted to it inflates pitch quality; score out of season.
- forecast.py and milb_translate.py import each other (the import sits inside add_minors); from scratch, run the forecast, the translation, then the forecast again.
- The translation's back-test borrows the forecast's final k, which for pitchers' strikeouts / walks and contact was fitted with the pitch-quality prior; the back-test applies no prior (Section 14).
- fetch_depth_charts.py drops rows without an MLBAM id; FanGraphs' depth-chart feed itself has no season parameter.
- 2020's July 23 Opening Day makes some ages jump two years from 2019; a Stats API primary position of DH brings the DH adjustment to a translated player.
- Never divide a pitcher's regressed platoon gap between strikeouts / walks and contact by his own strikeouts-and-walks share of his overall rate: the share explodes when the two parts have opposite signs and nearly cancel. Until 2026-11-16 it did, and 475 of 975 projected pitchers had a share outside [0, 1] (Sandy Alcantara vs LHB +1,185 in strikeouts and walks and −1,195 in contact per 600; totals were right). forecast.league_gap_shares now uses the share of each hand's typical gap instead (constants above).