Methodology
Walk-forward backtest 2022–2026: 57.3% of 12334 winners, Brier 0.242, log loss 0.6769. Always picking the home team: 52.9%.
| 2022 | 2470 games | 60.0% | Brier 0.2377 |
|---|---|---|---|
| 2023 | 2471 games | 57.1% | Brier 0.2439 |
| 2024 | 2472 games | 57.0% | Brier 0.2416 |
| 2025 | 2477 games | 56.5% | Brier 0.2434 |
| 2026 | 2444 games | 56.2% | Brier 0.2436 |
1. The data
Every MLB game since 2019 from the MLB Stats API (starting pitchers, lineups, batter and pitcher lines, pitch counts, home-plate umpire and game-time weather); Statcast leaderboards from Baseball Savant (previous season only, so nothing from the future leaks in); Retrosheet game logs 2010–2018 for longer park and umpire history, linked through the Chadwick Baseball Bureau register; and National Weather Service hourly forecasts for upcoming games.
2. Player projections (Marcel-style)
Each hitter's wOBA and each pitcher's strikeout, walk + hit-by-pitch, home-run and runs-allowed rates are projected from the current season and the three before it, weighted 1.25 / 1.0 / 0.8 / 0.6, measured against each season's league average and regressed toward average by adding league-average playing time: 320 plate appearances for hitters; 70, 170 and 1,000 batters faced for pitchers' strikeouts, walks and home runs.
3. The run model
Expected runs for each team come from a log-linear (Poisson) regression on the opposing starter (weighted by his expected share of innings), the opposing bullpen and its availability, the lineup's projected wOBA, Statcast contact quality and platoon advantage, defense (Outs Above Average and catcher framing), and context: ballpark, temperature, wind, umpire, home field and league scoring level. Both teams' expected runs feed a negative-binomial run distribution, which gives the win probability, one-run and two-run margins and every run total; the win probability is then Platt-calibrated.
4. The backtest
Walk-forward with no look-ahead: history is replayed one date at a time using only what was known the night before, and each test season is fit only on earlier seasons.
BullpenEdge publishes MLB statistics and model projections for analysis and entertainment. Projections are estimates, can be wrong, and are not advice of any kind. Not affiliated with Major League Baseball. See the Disclaimer and Methodology.