Prediction accuracy
Models compared
| Model | Scored | Winner acc | Brier | Log loss | Run-margin MAE |
|---|---|---|---|---|---|
| Run margin | 2274 | 54.5% | 0.2492 | 0.6926 | 3.60 |
| Expected runs (xwOBA) | 2274 | 52.9% | 0.2500 | 0.6934 | 3.57 |
Lower Brier / log loss = better-calibrated probabilities; accuracy is a coarse tiebreak. Judge model changes by walk-forward backtest.
Detail — runs model
2274
Games scored
54.5%
Winner accuracy
0.249
Brier score
0.693
Log loss
3.60
Run-margin abs error
+30.98u
Bet P&L (7.5% ROI)
Is it beating "always pick the home team"?
+1.7pp over 2274 scored games
model, cumulative
home team every game
±2 SE on the model
2026-04-02 → 2026-09-24
Cumulative, not rolling — a 20-game window on a ~54% signal has an ~11pp standard error, so it only ever showed noise. The first 29 games are omitted while the running mean settles. The baseline still runs inside the model's ±2 SE band — at 2274 games this gap is not yet separable from chance. Live results are one sample; judge model changes by walk-forward backtest.
Calibration
| Model home win % | Games | Predicted avg | Actual home win rate | Gap |
|---|---|---|---|---|
| 0.00–0.40 | 35 | 26.7% | 51.4% | +24.7pp |
| 0.40–0.45 | 54 | 43.1% | 29.6% | -13.5pp |
| 0.45–0.50 | 358 | 48.2% | 47.5% | -0.7pp |
| 0.50–0.55 | 1331 | 52.7% | 53.4% | +0.7pp |
| 0.55–0.60 | 424 | 56.7% | 57.5% | +0.9pp |
| 0.60–0.65 | 46 | 61.6% | 63.0% | +1.4pp |
| 0.65–1.00 | 26 | 74.5% | 46.2% | -28.4pp |
Well-calibrated buckets have small gaps. Persistent one-directional gaps suggest re-examining the win-probability conversion.