MLB Model Analysis – 2024/2025

Model Probability vs. Actual Win Rate

Each circle represents a bucket of bets grouped by the model's predicted win probability. The closer the circles are to the green diagonal line, the better calibrated the model is.

Model pred.

Market pred.

Perfect

Regression

Performance by Odds Range

Mkt Odds	Bets	Model	Actual	Gap
1.4–1.7	24	78.3%	70.8%	–7.5%
1.7–2.0	332	65.9%	61.7%	–4.2%
2.0–2.3	232	57.1%	52.2%	–4.9%
2.3–2.7	227	50.7%	48.0%	–2.7%
2.7+	200	41.8%	28.5%	–13.3%

        Gap = Actual – Model Predicted. Negative = model overestimates.
      

Signal Type Comparison

VALUE

50.1%

n = 667

Avg Edge: +20.5%

STRONG

50.3%

n = 348

Avg Edge: +26.3%

STRONG signals show a larger edge over the market but win at nearly the same rate as VALUE bets — the extra edge doesn't translate to extra wins.

📊 Key Findings

GOOD Favorites perform well. At odds 1.7–2.0, the model predicted 65.9% and actual was 61.7% — a small, acceptable gap.
GOOD Middle range is solid. Odds 2.0–2.7 show only 2–5% overestimation. The model's core edge lives here.
WATCH STRONG signals aren't stronger. Despite a larger predicted edge, they win at the same rate as VALUE bets (50%).
WEAK Big underdogs are a problem. At 2.7+ odds, the model predicted 41.8% wins but only 28.5% actually won — a 13.3% miss.
NOTE Model is systematically overconfident. Regression slope of 0.83 (vs ideal 1.0) means it consistently over-estimates by ~5–6% on average.

📈 Linear Regression Summary

A regression was fit between the model's predicted win probability (x) and whether the bet actually won (y = 1) or lost (y = 0).

Actual Win% ≈ 0.034 + 0.833 × Model Prob

A perfect model would give slope = 1.0 and intercept = 0. A slope of 0.83 means the model is slightly overconfident — when it says 70% chance, the real odds are closer to 62%. The R² of 0.04 is expected for game-level data (single games are noisy), but the bucket-level calibration is more meaningful.

Slope

0.83

Intercept

0.034

R²

0.043

P-value

<0.001