Vantage

A gradient-boosted machine learning model trained on historical match data. It ingests signals like team form, head-to-head records, home and away performance splits, goals scored and conceded, and rest periods between fixtures, among other variables. The model learns non-linear patterns and interaction effects that simpler statistical methods miss.

A dynamic rating system where teams gain or lose points after each result based on opponent strength and venue advantage. The rating gap between two teams translates directly into a win probability, providing a clean, time-tested baseline independent of machine learning.

The collective opinion implied by market prices, treated here purely as a forecasting signal. Markets incorporate dispersed information — injuries, tactical news, and public sentiment — that is hard to capture in structured datasets. It acts as a "wisdom of crowds" baseline to ground the engine's predictions.

The three probability sets are blended and passed through a calibration layer that corrects for overconfidence. The output is a calibrated set of probabilities for home win, draw, and away win — designed to mean what they say. An 80% prediction should correspond to roughly an 80% long-run hit rate.