Model Accuracy
Evaluate South Ark’s pregame forecasts against verified finals by sport. Explore calibration buckets, Brier scores and log loss.
College Football calibration
351 graded forecasts · Oct 9, 2026 – Nov 14, 2026
| Forecast | Brier score | Log loss |
|---|---|---|
| South Ark | 0.2372 | 0.6674 |
| Home-team baseline | 0.2463 | 0.6858 |
Predicted vs. actual win rate
| Probability bucket | Predicted win % | Actual win % | Games (n) |
|---|---|---|---|
| 50–55% | 52.4% | 40.4% | 47 |
| 55–60% | 56.5% | 57.1% | 42 |
| 60–65% | 63.2% | 57.0% | 86 |
| 65–70% | 67.0% | 58.7% | 46 |
| 70–80% | 74.5% | 64.0% | 89 |
| 80%+ | 83.7% | 80.5% | 41 |
Scoring and inclusion rules
Each game contributes one original pregame forecast paired with a verified final. Predicted win % is the average probability assigned to the model favorite in the bucket. At 50–50 the listed home team sets the orientation. Tied results are excluded and NFL probabilities are conditioned on a decisive result. The baseline gives the listed home team 55% and uses 50% when the venue is known to be neutral. Log loss clips probabilities at 0.000001 and 0.999999.
Football and college basketball use stored probabilities. When an NBA forecast contains only scores, the forecast’s margin is converted with the existing raw NBA transform, 1 ÷ (1 + exp(−margin ÷ 10)). The sample includes eligible published forecasts across model versions and can vary by sport. A lower score than the baseline is one measure of performance. The bucket gaps show calibration directly.
