archived

Not a prediction. A probability.

A three-model ensemble (logistic regression, random forest, XGBoost) trained on 22,714 games of Statcast-era data that generated pre-game win probabilities benchmarked directly against Kalshi prediction market prices. Each model was isotonically calibrated and evaluated on Brier score, enabling an apples-to-apples comparison against market-implied probabilities to surface where the models agreed, disagreed, and where systematic edges might exist. The public app has been retired; the historical data and source code are preserved.