How AI Predicts Greyhound Winners

Our prediction engine uses a LightGBM gradient boosting model trained on thousands of historical greyhound races. The model analyses 23+ features per runner including recent form, track bias, trap statistics, sectional times, and trainer performance.

The training pipeline uses walk-forward validation to prevent look-ahead bias. We retrain weekly to capture the latest form changes and track conditions.

Calibration is performed using Platt scaling to ensure that when the model says a dog has a 30% chance of winning, it actually wins approximately 30% of the time across many predictions.

The result is a calibrated win probability for each runner, which feeds into our Kelly Criterion staking calculator to recommend optimal bet sizes.