Trust starts with
showing the work.

Learn. Calibrate. Test.

The production score fit includes eligible completed games through its current cutoff. These results describe a retrospective preseason test, using source data that may contain later corrections. They do not measure the performance of forecasts published before those games.

Download calibration and test evidence ↗ includes every scored game, excluded game IDs, training IDs, fitted coefficients, source receipts and the model implementation hash.

Earlier v1 forecasts remain in the research ledger. Publishing v2 does not replace a game’s earliest eligible registration.

Does the efficiency correction travel?

This independent experiment adds lagged EPA per play and yards per play to the published score-only margin. Each test season is scored by a correction fitted only on earlier transition seasons. It remains a research comparison: the primary forecast, probability, interval and ledger are unchanged.

The current production scenario uses all eligible historical transition rows through 2025. Sparse or unknown teams shrink toward the league prior; no injuries, roster moves, weather or market prices enter this challenger.

Markets need a clock.

Imported lines are archival references. The archive does not provide a verified bookmaker timestamp or distinguish a closing quote. An observation made after kickoff never enters a pregame comparison. Historical forecasts and line observations are retained in D1 so future evaluation can use what was actually available before each game.

Home spread uses sportsbook convention: a negative number means the home team is favored. The model’s home margin uses the opposite convention: positive means a home win. Their difference is model home margin + market home spread. This is a disagreement in points, not a recommendation.