How the projections work

Projection model

Marcel-style three-year weighted projection. Weights, most recent season first: 6/4/2 for hitters, 6/3/1 for pitchers. Each prior season's PA/BF is first discounted by level reliability — a 600-PA Double-A season carries less MLB-quality information than a 600-PA MLB season — then the blend is regressed toward the league mean.

Regression prior strength (PA of league-mean evidence added; bigger = more regression):

Contact stats get double the prior for players with no MLB track record, and thin MLB careers get extra regression.

MLE translation

Minor-league lines are restated as park-neutral MLB equivalents in five steps: strip the park (half its effect), apply the league or level factor per component, adjust for batted-ball mix, adjust hitters for age versus the level, and rebuild wOBA/ISO/BABIP. Hitter factors are per component (Triple-A: K ×1.20, BB ×0.69, 2B ×0.83, HR ×0.70) with league-specific rows where the data supports them. Pitcher factors are level-wide (K: AAA ×0.88, AA ×0.81, A+ ×0.77, A ×0.74, Rookie ×0.68; walks and home runs rise on promotion; Triple-A 2023+ uses a separate walk factor).

Translation is evidence, not a forecast: regression to the mean happens afterwards, when seasons are blended into a projection.

Worked example, every factor, sample sizes and backtest evidence → · Parks, leagues & translations

Scouting blend

Scouting grades from public sources feed: (1) hit-grade target K%/BABIP for prospects, (2) stuff/command-grade target K%/BB% for pitcher prospects, (3) fielding/arm-grade fielding runs when MLB innings are thin (under 400), (4) speed-grade SB/PA. Scouting weight decays as MLB experience grows: veterans project from their stats, prospects from a stats-plus-scouting blend.

Aging curves

Per-stat year-over-year aging deltas, derived from historical paired seasons and corrected for survivorship (players who dropped out are imputed as replacement-level), are applied through the forward chain. Pitchers whose velocity data shows decline age faster. Fielding ages by position (catchers and corner outfielders fastest).

Trade value

The Trade Value board ranks on Trade Value Score = Surplus × Confidence. Surplus = NPV of projected WAR × $/WAR minus NPV of owed salary over team-controlled years: $8.0M per win in 2027, +3% per year, relievers at a discount, discounted 10% per year on both value and salary. Years projecting under 1.0 WAR are excluded; deferred money is NPV-discounted; options are exercised by team-versus-player rationality. Confidence = bust-rate by MLB experience and prospect grade (see Conf% in the glossary) × a debut-distance multiplier of 0.75^(years to debut − 1).

WAR composition

Position players: BAT (wOBA-derived), RUN (SB/CS + UBR), FLD (regressed defensive metrics + scouting), POS adjustment, REP credit, at 9.8 runs per win. Pitchers: FIP-based, replacement FIP 5.40 (SP) / 4.55 (RP), 9.5 runs per win in the projection chain (display tables use a pitcher-specific figure near 9). Projection cells show park-adjusted by default; boards offer a park-neutral view.

Calibration & accuracy

The accuracy backtest scores the 2026 season's actual results against projections from the same preseason snapshot used by public industry projection systems — a check on how well each system fits the most recent reality, not a true archived-preseason test (those are paywalled). Hitter WAR RMSE 0.91; pitcher WAR RMSE 0.57; standings RMSE 5.0 wins per team. 20 sanity assertions run after every build. MLE factors have their own evidence: MLE Factors.

Recent improvements

Selected past revisions (older entries; the model is rebuilt daily):
R845 — math audit: league-abbreviation MLE routing, DSL separated from Rookie, park-factor units, half-effect park conventions. R816 — age-versus-level premium wired into translation. R782 — league/MLE recalibration for the MLB-ball era and Triple-A automated strike zone. R195 — pitcher pages on the compact header. R168 — survivorship-corrected aging curves. R156 — hitter level reliability. R153 — pitcher Marcel BF discounted by level reliability.