← Back to the universe 06 / 13

06 / QUANT · TEXAS UCF

Tennis fair-value model

A live win-probability model for tennis prediction markets, built to price matches in play.

Code on GitHub ↗
Open slides full screen ↗
Tennis Odds presentation
MLPScraped match dataHMMsCalibration

An MLP, trained on scraped match data, is paired with a recursive match-outcome solver that turns the live score into a fair match-win probability. I also trained HMM models, and worked with teammates on an XGBoost approach.

0.1777Brier score (book: 0.1694)
0.5293log loss (book: 0.5073)
0.008Brier behind live sportsbook odds

Reading the numbers

Measured against the bookmaker on the same matches, the model is competitive with live sportsbook odds and slightly behind them. Most of the remaining gap is calibration — its expected calibration error is 0.042 against the book’s 0.023 — which is the next thing to fix.