How the model works

The base engine blends four active models per half-inning: Poisson (12%), ZIP (30%), Markov 24-state chain (48%), and MAPRE (10%). Logistic Stack, NN Interaction, and Hierarchical Bayes are diagnostic outputs with zero blend weight. Dynamic Bayesian shrinkage, handedness-adjusted lineup splits, vector wind modeling, and monotonic calibration are applied before final blending. Calibration currently uses an identity mapping pending an out-of-sample refit.

P(NRFI) = clamp(0.76 × calibrated_ensemble + 0.24 × 0.5056, 0.18, 0.85)

Value bets are identified when the model's implied probability exceeds the bookmaker's by ≥3%. Kelly Criterion (25% fractional) sizes the recommended wager.