Model Lab
Historical backtests for every prediction model — from dead-simple baselines to FIP-adjusted Pythagorean models. Browse season-by-season accuracy, ROI, and closing-line value (CLV). No look-ahead — priors trained strictly on prior seasons.
Tier S — Market Strategies
profitable every backtested seasonFilters fip_adjusted predictions to games where the model pick is priced as a market underdog AND the implied-probability edge exceeds 10%. Profitable in all 5 backtested seasons (2021–2025): +5.8% to +12.1% ROI, 15–23% trigger rate, consistent positive CLV (~+83 to +160 pts vs. close). Requires confirmed starting pitchers and pre-game closing odds.
Fundamentals-only moneyline model using team-prior offensive quality, starting pitcher FIP (Bayesian-shrunk), and park factors. Expected runs: LEAGUE_RPG × (team_rpg / LEAGUE_RPG) × pitching_factor × park_factor. lineup_wOBA proxy = team_rpg, upgradeable when day-of lineup data is available. Positive flat-bet ROI in all 5 backtested seasons (2021–2025).
Tier −1 — Disproven Theories
strategies that seem profitable but aren'tA live in-game monitoring model. The team that scores first to take a lead wins ~65.5% of MLB games — but filtering to games where they subsequently fall behind (before the 7th inning) selects for games where the opponent is outperforming. Real MLB data (2021–2025, 12,120 games, 1,942 triggers) shows the first-lead team wins only ~30.3% of those spots, well below the ~40.7% breakeven implied by typical live prices (+145 to +150). Flat-bet ROI: –11.6% to –26.1% across all seasons. No demonstrated positive edge. Retained for transparency and as a reference negative-result model.
Tier 0 — Baselines
Assigns 50% probability to every game. Kelly bets whenever the market prices the home team at +100 or better (i.e. the model's 50% exceeds the breakeven for that price), and skips all favorites. The Kelly curves typically collapse — aggressive sizing on underdog payouts leads to ruin. The flat-bet line bets the home team on every game at actual closing odds — favorites included.
Assigns the historical MLB home win rate (53.7%) to every home team. Shows whether home field advantage alone is sufficient to beat the market — and what Kelly does with a small, uniform edge.
Tier 1 — Team RPG
Bill James Pythagorean formula (exponent 1.83): compares each team's season runs-per-game as a proxy for overall strength. No home/away split and no park factor adjustment — a pure team-quality signal.
Pythagorean formula (exponent 1.83) with home/away RPG splits. Uses the home team's home scoring rate vs. the away team's road scoring rate — naturally embedding park effects and travel fatigue.
Expected game total = home team's home RPG + away team's road RPG. A direct read of historical scoring rates for the specific context (home team at home, away team on the road). No simulation required.
P(home covers -1.5) via a normal approximation of the run-difference distribution. Variance is estimated as 1.2× the expected total (overdispersed Poisson). Uses home/away RPG splits for accuracy.
Tier 2 — Pitcher-Adjusted
Pythagorean win probability with team run rates adjusted by opposing starter FIP relative to league average. A better-than-average starter suppresses the opponent's expected run output. Degrades to pythagorean_home when starter data is unavailable.
Expected game total from FIP-adjusted run rates for both starters. Accounts for starter quality when predicting over/under lines and run-line coverage probability.
FIP-adjusted win probability and game total, enhanced with per-starter K% and BB% to generate strikeout prop probability distributions. Uses a Poisson model with mean = K/9 × expected IP/9.