The numbers

Plotting every predicted margin against what actually happened, with the model's own 50, 80 and 90% intervals drawn on: roughly the right share of results fall inside each band, across the whole range of predicted margins.

The model's claimed edge is flat across lead times and roughly equal for bets that eventually won and bets that eventually lost. What moves realised profit is when the bet is entered: it climbs steadily as entry moves from tip-off back to three or four days out. Earlier, the market's price reflects less of the idiosyncratic player and team news that sits outside the backtested model's scope, therefore an early bet is much less exposed to being adversely selected on information the model has not seen. The served strategy is defined by an entry window for that reason, and the backtested profit is a lower bound: live, the model is given the same team news the market has before it prices the bet.

A calibrated probabilistic model, checked out of sample across the 2016–2025 seasons (1459 matches) including one untouched holdout, with one pocket of betting edge that clears our bar: early-entry handicap bets, bootstrap interval excluding zero on the pooled sample and on each season on its own. Every figure on these pages is reproducible from the model's own frozen predictions and the recorded market history.

Two forces work against a static model from here: the irreducible noise in the sport, and a market we expect to price these games more efficiently over the coming seasons. We expect our own edge to widen against both: the served model already carries a live information layer the backtest cannot use, and the method behind it is still being extended.