The tournament as it actually played out, with the model's prediction for each real matchup computed against it. A struck-through name was the model's pick for that game; the name shown above it in its place is who actually won. A box with neither is a correct prediction; a box with no shading at all is a real matchup still awaiting its result.
This season's realized numbers next to what the model's held-out backtest led us to expect beforehand (334 graded games pooled across its recent prior-season backtests - see Model Validation) - not a guarantee, but the baseline for "is this season in line with how the model typically performs."
| This season (actual) | Backtest-expected | |
|---|---|---|
| Games graded | 47 | 334 |
| Winner accuracy | 70.2% | 67.7% |
| Brier score | 0.1779 | 0.2021 |
| Margin MAE | 10.56 | 10.11 |
See the Data Dictionary for what each of these actually measures - winner accuracy is the simplest ("how many picks were right"), Brier score is the standard scoring rule for a probability forecast (rewards being right AND appropriately confident, not just bold), and margin MAE grades the point-spread call separately from the win/loss pick. "Backtest-expected" is a games-weighted average across the model's held-out prior-season backtests (Model Validation page) - not this season's own training data, so it's a genuine before-the-fact expectation, not hindsight.
| Team A | Team B | Predicted winner | Expected margin | Win probability | Actual winner |
|---|---|---|---|---|---|
| Lehigh | Prairie View | 4.4 | 65.3% | Prairie View | |
| Howard | UMBC | -2.9 | 40.6% | Howard | |
| Miami OH | SMU | -6.5 | 29.3% | Miami OH | |
| NC State | Texas | -1.2 | 46.4% | Texas |
The favorite advances at every game, computed once from each game's real win probability - a single fixed path, not a simulated range.