Accuracy reflects how often the model’s predicted winner (based on the strength gap) matched the actual winner. It’s simply correct picks ÷ total graded games.

This view counts every game with a final score recorded, from Week 1 on.

Games are grouped by the predicted strength gap. Wider gaps indicate higher confidence.

Each band shows how often the favored team actually won and the average implied probability in that band. It’s a quick way to see if big projected edges translate to higher hit rates.

Calibration compares what the model thought would happen to what actually happened across ranges of predicted win probability.

Ideally the average predicted chance matches the true win rate in each bucket. Divergence signals that the probability curve might need tuning.

This table highlights a subset of recent games so you can spot where the model aligned with reality and where it missed.

Rows in yellow denote incorrect calls. Use the predicted margin and probability columns to gauge whether the miss was a low-confidence upset or a surprising result.