Project field note
NBA Player Props
A forecasting system that compares live Kalshi player prop prices against probabilistic NBA player projections, surfacing the largest gaps between market and model.
Pulls live player prop prices from Kalshi and projects points, rebounds, and assists for each player using a hierarchical Bayesian model wrapped around an XGBoost feature predictor, producing full distributions with uncertainty rather than single-point predictions.
Ingests injury news from official NBA reports and beat-writer feeds, uses an LLM to classify severity and expected games missed, and routes that signal through a lineup-redistribution layer so projections shift toward the players who actually absorb the minutes.
Backtests are point-in-time correct, meaning every prediction is evaluated using only data that was available before the game, with calibration tracked over a full season of out-of-sample bets.
Surfaces the props where the model and market disagree most, ranked by the size of the disagreement and the model's confidence, so the most inspectable opportunities are easy to find.