NBA Predictions
A multi-stage NBA forecasting system that combines four model families, calibration, walk-forward validation, and a downstream EV simulation layer.
Overview
NBA Predictions is a full forecasting and evaluation pipeline, not just a single classifier. It collects historical games, player and team statistics, injuries, odds, and market snapshots into a relational store, then builds contextual features for game, prop, and market analysis.
The modeling stack combines interpretable ratings and regressions with boosted trees, sequence and graph models, and possession-level simulation. A meta-learner and calibration layer turn those system outputs into probabilities before edge detection and portfolio logic evaluate whether a market price is actionable.
Methodology
Feature construction is organized into eleven categories and roughly 217 signals, including team efficiency, matchup efficiency, recent form, ratings, lineups, injuries, hustle, clutch play, historical matchups, situational context, and interactions. Exponential decay gives recent evidence more weight while preserving longer-term priors.
Four systems provide different inductive biases: Elo/Four Factors/hierarchical Bayesian and negative-binomial-copula components; calibrated CatBoost, LightGBM, and XGBoost learners; GNN, LSTM, and Transformer sequence models; and a possession-level Monte Carlo simulator. A 47-feature meta-learner, calibration checks, and leave-one-system-out diagnostics reconcile them before a three-stage odds/edge pipeline and correlation-aware sizing layer.
The evaluation path is time-ordered and walk-forward. Leakage and timestamp checks, realistic odds checks, no-reoptimization rules, proper scoring and calibration metrics, closing-line value, retraining monitoring, and cost-aware portfolio constraints are treated as part of the model rather than as afterthoughts. Monte Carlo is used to simulate expected value from calibrated outputs; it is one layer of the broader system.
What it does
- Relational ingestion for games, teams, players, injuries, odds, and props
- 217-feature builder with decay weighting and matchup interactions
- Four-system ensemble spanning ratings, boosted trees, deep models, and possessions
- Meta-learner with probability and point-prediction calibration
- Three-stage edge detection with overround and consensus-line handling
- Monte Carlo EV simulation, Kelly sizing, correlation adjustment, and exposure caps
- Walk-forward backtesting with leakage and reoptimization integrity checks