Methodology

How The SABRhood defines, validates, and communicates its baseball metrics.

Open methods

Trust the number because you can inspect it

Every published product has a defined grain, information cutoff, sample threshold, and method label.

Canonical grain

Pitch

314,584

One row per pitch; no full PBP shipped to the browser.

Run environment

Empty, 0 outs

0.505

Empirical expected runs to the end of the inning.

Run environment

Loaded, 0 outs

2.448

Estimated from the same season-level state table.

Interpretation rule: fixed-weight wOBA is labeled as an estimate, barrel classification is labeled as a proxy, and the pressure measure is labeled as a leverage proxy. These fields are designed for transparent reporting, not to impersonate proprietary or official metrics.

Core rules

  1. Raw play-by-play remains outside the public website.
  2. The site receives compact derived products with one documented row grain.
  3. Pregame analysis never uses information from later in the game.
  4. Estimated and proxy statistics are labeled as such.
  5. Small samples receive reliability fields instead of false certainty.
  6. Automated story candidates remain subject to editorial review.

Current analytical layers

  • Canonical pitch, plate-appearance, batted-ball, and pitcher-appearance views.
  • Empirical 24-state run expectancy and linear run value.
  • Hitter and pitcher results, contact quality, and platoon splits.
  • Pitch shape, discipline, and within-plate-appearance sequencing.
  • Recent performance compared with a non-overlapping earlier baseline.
  • Six direction-aware player change z-scores paired with six full-season MLB percentiles. Positive change always means improvement from the player’s perspective; lower results allowed therefore count as better for pitchers.
  • Bullpen workload availability and exploratory pitcher-hook modeling.
  • Historical career profiles built from Lahman career totals, Hall of Fame voting, awards, All-Star selections, postseason volume, longevity, and record standing.
  • Recognition-aware anniversary, milestone, and career-record story candidates.

What “estimated wOBA” means here

The current site field is a transparent results-based wOBA calculation. It applies fixed linear weights to walks, hit-by-pitches, singles, doubles, triples, and home runs, then divides by wOBA opportunities. It is not xwOBA and does not currently use the launch-speed and launch-angle expected-contact model from the earlier research scripts.

That expected-contact work remains a separate future package layer. Until it is ported, validated, and versioned, the site will continue to say “estimated wOBA” rather than “expected wOBA” or “xwOBA.”

Plate-appearance matchup probabilities

Simulation Phase 1 assigns every confirmed batter-versus-probable-starter matchup one mutually exclusive distribution across walk, hit-by-pitch, strikeout, single, double, triple, home run, and other out.

The current shadow model follows four steps:

  1. Overall hitter rates are shrunk toward the MLB event distribution with a 200-plate-appearance prior.

  2. Overall pitcher rates are shrunk with a 250-batter-faced prior.

  3. The relevant hitter and pitcher handedness splits are each shrunk toward that player’s overall distribution with a 125-opportunity prior.

  4. Batter and pitcher distributions are combined event by event using multinomial log5:

    [ s_e = , p_{matchup,e} = ]

The final normalization is important: all possible plate-appearance outcomes sum to exactly 100%. Recent form can make only a restrained, reliability-weighted adjustment to strikeout, walk, hit, and home-run components. The model also publishes the player and split sample sizes, a matchup reliability score, its input status, and its version.

This layer is running in shadow mode. It will not replace the existing game probabilities until historical and forward-looking calibration demonstrate that it improves out-of-sample forecasts.

Plate-appearance game-state simulation

Simulation Phase 2 consumes the Phase 1 event distributions and plays each game one plate appearance at a time. The state carried between events is the inning and half-inning, number of outs, occupied bases, score, next lineup position, and workload of the defending probable starter. Each batting order continues from the hitter after the final out rather than restarting every inning.

The probable starter remains active until a sampled batters-faced target or pitch limit is reached. The replacement distribution is an availability-weighted blend of up to six active-roster relievers, excluding the scheduled starter. Each hitter-reliever probability is rebuilt from the same player, handedness, and recent-form contract used against the starter.

Runner advancement is deliberately inspectable in the first version:

  • walks and hit-by-pitches use forced advancement;
  • all runners score on home runs and triples;
  • runners on second score on 60% of singles;
  • runners on first reach third on 30% of singles and score on 55% of doubles;
  • strikeouts and other outs do not advance runners;
  • a runner begins on second in each half-inning after the ninth.

The engine directly returns game scores, winners, one-run and extra-inning rates, starter batters faced and pitch counts, bullpen plate-appearance share, and full-game hitter distributions for hits, walks, strikeouts, home runs, extra-base hits, and total bases. The initial output remains a shadow model. That initial Phase 2 output deliberately omitted double plays, sacrifice advancement, steals, individual reliever sequence, and park-sensitive baserunning. Phase 3 adds those mechanisms below while leaving defense as an explicit future input.

Phase 3 bullpen decisions and baserunning

Phase 3 replaces the aggregate bullpen handoff with a named-reliever decision. At every pitching change, each unused active reliever receives a selection score from current availability, event-weighted performance against the next three hitters, role, inning, score margin, and leverage. Scores are converted to a probability distribution and one arm is sampled. The selected pitcher remains in the game until a sampled batters-faced workload, 30-pitch ceiling, or inning-change hook is reached. The output retains both appearance probability and the reliever’s mean selection likelihood when considered.

Occupied bases now retain the responsible lineup position. This allows the simulation to credit runs, RBI, stolen bases, caught stealing, sacrifice flies, and grounded-into-double-play events to the appropriate hitter. Other outs can advance runners, score a runner from third, or become double plays through visible league-rate assumptions.

The first park layer is intentionally modest. Venues are assigned transparent expansive, quirky, compact, or neutral geometry tiers. Those tiers adjust only runner advancement and steal-attempt rates; they are not presented as empirical causal park factors. An eventual multi-season venue model can replace them only after home-team, runner-speed, batted-ball, and scoring-environment effects are separated.

Phase 3 calibration is chronological. The engine first needs 300 settled, pregame-eligible forecasts. The earliest 70% form the training block and the later 30% remain untouched for evaluation. A calibration layer enters shadow application only if it improves both win-probability Brier score and team-runs mean absolute error on that later block. Public promotion remains a separate decision.

Phase 4 empirical running game

Phase 4 reconstructs runner identity before and after every plate appearance in the private PBP store. The training table distinguishes scoring from second on a single, first-to-third on a single, scoring from first on a double, advancement on an out, double-play opportunities, and eligible windows for steals of second or third. Stolen-base events are matched to the runner and pitcher inside the same plate appearance.

Player rates are not used raw. Every runner tendency is pulled toward the league rate with an empirical-Bayes prior, while steal success receives a separate prior because attempts are much scarcer than eligible windows. Pitcher hold profiles measure how often runners try to go, then separately retain success allowed. The daily simulator uses the probable starter’s profile when it is stable and a team-level fallback otherwise.

The venue layer joins every completed game to the official MLB schedule venue. Park rates are estimated by runner-event type and shrunk toward the league before becoming bounded multipliers. These are genuine empirical inputs, but they remain descriptive: the current version has not fully separated runner mix, defense, batted-ball direction, or weather from the venue. That composition-adjusted model is the next validation step.

The granular opportunity table remains private. Public products contain only league rates, shrunk runner profiles, pitcher-hold summaries, bounded park factors, and a model card. Phase 4 stays in shadow mode and inherits the same chronological publication gate as Phase 3.

Phase 5 manager hook and starter workload

Phase 5 replaces the generic starter cutoff with the pooled manager hook model. After every plate appearance, the simulator recomputes removal probability from the starter’s pitch count, batters faced, times through the order, inning, fielding-score margin, and the result of the plate appearance. The hook model was trained on earlier decisions and checked on later dates before it was allowed into the shadow simulator.

Pregame starter workload blends season batters faced with the five most recent starter-like appearances, shrinks the recent component, and accounts for short rest. That expectation supplies both batters-faced and pitch-limit inputs. Temperature and the current park factor tilt event odds before each draw; wind remains descriptive until direction can be aligned to each ballpark. Phase 5 begins a new version-specific ledger and cannot inherit calibration approval from an earlier simulation version.

Run Game Engine, catcher framing, and ABS

The Run Game Engine moves the steal opportunity to pitch grain. An eligible window requires a known runner on first or second, the next base open, fewer than three outs, and a pitch delivered before the runner event. Stolen bases and caught stealings are attached to the preceding pitch so count, pitch family, plate flight time, and disengagement context describe the decision window that actually existed.

Starting catchers come from official lineups when available. Defensive substitutions and position switches create catcher stints inside the game. When a starting catcher must be inferred from roster and batting-order evidence, that assignment is labeled and contributes less reliability.

Attempt and success are separate additive logistic models. Each includes runner, pitcher, catcher, target base, count, inning, score margin, pitch family, and disengagement context, with shrunken player effects. Pitcher and catcher boards therefore describe value after the quality and behavior of opponents are considered. Count-window notes combine projected offspeed and breaking-ball mix, plate flight time, and observed running behavior; they are scouting prompts rather than automatic steal recommendations.

The framing development model uses only called pitches. Expected called-strike probability is estimated from horizontal and vertical location bins, count, batter side, pitcher, and umpire. The catcher’s shrunken residual effect produces estimated extra strikes and a transparent run estimate of 0.125 runs per strike. It is not an official Statcast framing metric.

ABS challenge products standardize the official Baseball Savant batter, catcher, pitcher, and team leaderboards. The public page preserves Savant’s opportunity, challenge, overturn, expected-value, and run-value fields. These official challenge results are kept separate from the SABRhood framing development score.

Fielding Engine and runner advancement

The Fielding Engine starts with one credited ball in play. The primary fielder and position are parsed from the official play description and matched to the season player reference. A transparent expected-out development estimate uses batted-ball trajectory, field location, exit velocity, launch angle, and distance, with detailed context cells pulled toward broader trajectory-location rates in small samples.

This estimate is not Statcast Catch Probability or Outs Above Average. Public labels explicitly note that the play-by-play feed does not contain official starting position, jump, route, wall, or opportunity-time tracking. Official Fielding Run Value is imported and displayed as a separate contract, including its range, arm, double-play, catcher throwing, blocking, and framing components.

Runner-advancement prevention reconstructs five decisions: scoring from second on a single, first-to-third on a single, scoring from first on a double, second-to-third on an out, and first-to-second on an out. A shrunken additive logistic model separates runner, credited fielder, advancement type, batted-ball context, outs, inning, and score effects. The fielder board therefore measures extra bases allowed relative to the runners and situations faced.

Gold Glove Watch standardizes official Fielding Run Value inside each league-position group and shrinks the rate by playing time. The American League and National League therefore have separate projected selections at every tracked position. It is a current-value watchlist, not a forecast of award voting. Pitcher and utility are left unfilled because the current official fielding contract does not provide comparable rows for those award positions. Play of the Day selects the strongest positive SABRhood estimated range or advancement contribution for each date and retains its development label on every surface.

History Match Engine

History Match starts from a current player or team line and selects thresholds from curated two-stat profiles rather than treating the exact current values as the only comparison. Season claims are compared with completed historical seasons. Rolling claims compare identical numbers of player games, pitcher appearances, or team games.

All rate statistics are rebuilt from additive components. The claim records its historical universe, selected thresholds, comparison directions, most recent precedent, source-through date, and method version. A claim fingerprint covers the threshold and precedent; a separate content fingerprint covers the current headline and sentence.

The Analytics Lab can display every discovered note. The broadcast packet uses the smaller packet-qualified queue and labels pending candidates for producer verification. The public History Match page accepts only claims explicitly approved for the website. If a threshold, comparison universe, or historical precedent changes, the stored approval no longer matches and the note returns to editorial review automatically.

Historical season, rolling-window, and career-path comparisons use the modeling index from 1974 through the latest completed release. The separate record-only archive scans Retrosheet player games from 1898 through the latest completed release for rare single-game feats. Older record seasons never enter career trajectories, career backtests, or current-player projections. Current-season lines use the daily completed-game pipeline and official season products where required.

Historical recognition model

The career-significance score exists to improve editorial ordering. It is not a replacement for WAR and should never be cited as a value statistic. Hall of Fame status, major awards, All-Star selections, career volume, longevity, postseason volume, and top-ten career ranks contribute to the current version.

Lahman does not include WAR. A future WAR layer must arrive as a separately sourced, versioned table and will remain visibly labeled by source and definition.

Player Change Engine

For each qualified player, the current build compares a 14-day window with the non-overlapping portion of the season that came before it. OPS, estimated wOBA, strikeout rate, walk rate, hard-hit rate, and run value per plate appearance are each converted to a z-score across the appropriate hitter or pitcher cohort. The largest absolute z-score becomes the selected change, while the card still shows all six so the selection can be audited.

Season percentiles are calculated separately. They describe the player’s full season level among the same perspective cohort and prevent an unusual short-term change from being confused with elite overall performance. The context signal weights the size of the selected change by recent-sample reliability; it remains a reporting-priority score, not a projection.

Career Path Engine validation

The hitter engine aligns players by career games and age, then measures robustly scaled distance across playing-time pace and era-indexed OBP, SLG, ISO, walk, strikeout, home-run, and speed components. Twenty completed historical neighbors form weighted outcome distributions.

Validation rolls the model forward through 1995-2022 at 100, 250, 500, 750, and 1,000 career games. A pseudo-forecast may use only players whose careers ended at least three seasons before that forecast origin. The subject’s future is used only after prediction to score remaining-games error, p10-p90 interval coverage, and 250/500/1,000-game Brier scores.

The current 545-test sample has a weighted 317-game remaining-career MAE and 83% coverage for the nominal 80% interval. A probability adjustment trained on origins through 2012 increased Brier error in the 2013-2022 holdout, including the 500-game score rising from 0.187 to 0.218. That adjustment is therefore not used. The displayed probabilities remain raw weighted-neighbor estimates, and the product remains a backtested development model rather than a calibrated forecast.

Five interpretable similarity-weight profiles were then selected separately at each career checkpoint using only 1995-2012 origins. Durability won the training comparison at 100 and 250 games, contact and discipline at 750, and power shape at 1,000; the baseline remained best at 500. On the untouched 2013-2022 holdout, however, their combined remaining-games MAE was 246.1 against 245.2 for the baseline. The tuned weights were therefore rejected.

A second refinement subtracted checkpoint-specific median forecast error learned from the early origins. Those training corrections ranged from -73 to -183 games, but the later holdout shifted in the opposite direction. Applying them increased MAE to 310.6 games and reduced p10-p90 coverage from 86% to 51%. The correction was rejected. Both experiments remain in the public model evidence so future iterations do not repeat an unsuccessful calculation.

The fourth development pass adds terminal batting-rate distributions. For each historical neighbor, the engine adds that player’s observed future hits, total bases, and times on base to the subject’s career-to-date totals. Weighted neighbor percentiles then produce projected final AVG, OBP, SLG, and OPS. On the untouched 2013-2022 holdout, median forecasts missed final AVG by .010, OBP by .010, SLG by .017, and OPS by .023.

Home-run pace is also reported per 600 future plate appearances. A transparent rate-linked scenario multiplies each neighbor’s future HR rate by the subject’s plate-appearance pace and the neighbor’s remaining games. Its final-home-run MAE was 25.1 on the later holdout, compared with 25.0 for the direct comparable career-total estimate. The rate-linked result is therefore displayed as an explanatory scenario but does not replace the slightly stronger primary home-run total.

The fifth development pass adds future offensive value and legacy-path context. Offensive value units are calculated as (offensive index - 100) × plate appearances / 600; they measure era-adjusted batting quality multiplied by opportunity and are explicitly not WAR. The engine projects each neighbor’s remaining value, adds it to the subject’s current value, and reports p10, p50, p90, and the weighted chance of adding at least 10 future units. On the 2013-2022 holdout, the median future-value estimate had a 12.0-unit MAE and the +10-upside probability had a .120 Brier score.

The Hall of Fame label is a historical lane, not an election probability. Comparable players count only when they had been retired for at least five years at that forecast date, and induction counts only when it had already happened by that date. The displayed share is the renormalized weight of those eligible neighbors who had been inducted. Labels require at least eight eligible neighbors and an effective sample of five: strong begins at 35%, credible at 15%, fringe at 5%, and lower shares are outside the current Hall of Fame lane. Position, defense, awards, postseason performance, character considerations, and voting behavior are not modeled.