Model intelligence lab
7 ML algorithms → 1 consensus ranking
Win probability per ball, validated with 5-fold GroupKFold by match (no leakage): 838 matches, 198,020 balls.
7
Algorithms compared
0.1772
Best Brier (lower = better)
27.6%
XGBoost skill vs coin flip
500
Bootstrap resamples
▤Brier skill vs coin flip
Higher = better
Selected for the Pressure Index: XGBoost. Logistic Regression has the best Brier but is additive: a wicket counts the same in over 1 and over 19, which distorts Leverage.
▦Algorithm scorecard
5-fold GroupKFold by match · out-of-fold
| Algorithm | Brier ↓ | Skill ↑ | AUC ↑ | Log-loss ↓ | Calib. err ↓ | Chase Brier | Role |
|---|---|---|---|---|---|---|---|
| Logistic Regression | 0.1772 | 29.1% | 0.809 | 0.520 | 0.008 | 0.1283 | Candidate |
| Ensemble (top-3 avg) | 0.1786 | 28.5% | 0.806 | 0.523 | 0.014 | 0.1291 | Candidate |
| XGBoost | 0.1809 | 27.6% | 0.802 | 0.531 | 0.025 | 0.1315 | Selected |
| LightGBM | 0.1822 | 27.1% | 0.800 | 0.536 | 0.029 | 0.1331 | Candidate |
| Random Forest | 0.1827 | 26.9% | 0.798 | 0.537 | 0.022 | 0.1324 | Candidate |
| HistGradientBoosting | 0.1841 | 26.3% | 0.798 | 0.547 | 0.038 | 0.1352 | Candidate |
| Lookup table (v1) | 0.1859 | 25.6% | – | – | – | – | Candidate |
| Neural Net (MLP) | 0.1901 | 23.9% | 0.786 | 0.567 | 0.041 | 0.1444 | Candidate |
↗Calibration: when the model says 70%, does the batting side win 70%?
Predicted vs actual win % (out-of-fold)
◈Top-10 batters
P(top 10) under each model
◈Top-10 bowlers
P(top 10) under each model
Consensus rule: a player's final rank is the mean bootstrap P(top 10) across the three pressure definitions (Lookup table, Logistic Regression, XGBoost), each resampled 500 times by match, so no single modelling choice decides the top 10.