Walk-Forward Validation
Backtest Results
Every number on this page comes from strict walk-forward validation — the model only sees data available before each game. No look-ahead bias, no cherry-picking, no curve fitting. These are predictions the model would have made in real time.
Methodology
The backtest uses walk-forward validation — for every game, the model is trained only on data from prior games. A warm-up period of 15 games per team per season is excluded so early-season noise doesn't pollute results. This mirrors exactly how the model runs in production: it never sees the future.
Predictions are compared against opening lines from SBRO (SBR Odds), not closing lines. Opening lines are the earliest available market price, posted before sharp money has moved them. That is the moment we care about: we publish early, and the question is whether the market moves toward our number afterward, not whether we out-guess the close.
ATS (against the spread) results assume +100 (no juice) — prediction market pricing. Profit calculations use flat $100 units on every qualifying game. No kelly sizing, no parlays, no teasing.
Verify it yourself
Download the dataset as the backtest file stands today: every game, every prediction, every result. Moneyline results and spread results before 2023-24 are withdrawn while the backtest file is corrected.
Want today's edges?
The same model that produced these backtest results runs every day with live data.