Model Performance
Every prediction is stored with its timestamp, model version, confidence and inputs. After each game the stored projection is graded against the actual result, which feeds the improvement loop. Model gsai-1.4.2.
Model spread vs reference spread
Absolute error of each pre-game differential against the final margin
Across 68 graded games the model's average error was 0.0 pts lower than the reference. Sample sizes are small in demo mode; treat this as a calibration readout, not a guarantee.
| Week | Model MAE | Reference MAE | Model closer | Games |
|---|---|---|---|---|
| Week 1 | 5.5 | 5.9 | 62.5% | 16 |
| Week 2 | 5.4 | 4.9 | 50% | 16 |
| Week 3 | 6.5 | 6.7 | 56.3% | 16 |
| Week 4 | 7.1 | 7.2 | 56.3% | 16 |
| Week 5 | 11.9 | 9.3 | 25% | 4 |
Reference spreads are informational context supplied by a data provider and are shown only for comparison with the GAME SPY model projection. The model spread is the model's projected point differential — it is not a line, an offer, or a recommendation to wager. GAME SPY AI does not accept or place bets.
Accuracy by week
Winner hit rate per completed week
Win probability calibration
Predicted confidence vs observed hit rate
| Bucket | Predicted | Actual | N |
|---|---|---|---|
| 50–60% | 56% | 50% | 12 |
| 60–70% | 64.4% | 82.4% | 17 |
| 70–80% | 76% | 87% | 23 |
| 80–100% | 84.4% | 87.5% | 16 |
Player projection accuracy by position
Mean absolute error in fantasy points