Model performance

Three different approaches to the same question

HoopModel runs three independently trained models on every game: a random forest, a gradient-boosted tree model (XGBoost), and a neural network. All three see the same inputs, rolling team form going into the game, but they learn from it in different ways. Tree models split the data into rule-like branches; the neural network learns smooth combinations of every input at once. When they agree it tends to be for good reason, and when they split, the game is usually genuinely close.

Accuracy isn't the whole story

Two models with the same winner accuracy can still differ a lot. Calibration asks whether a 70% pick actually wins about 70% of the time; a well-calibrated model's confidence means something. The ROC curve scores the probabilities rather than the picks: it shows how cleanly a model ranks likely home wins above likely home losses at every possible cutoff.

The confidence chart is the most practical view: picks the model was sure about should land more often than picks it barely made. If the bars don't climb from left to right, the model's confidence isn't telling you much.

Where the misses live

The confusion matrix splits the misses by direction. If a model has many more 'picked home, away won' errors than the reverse, it leans too hard on home-court advantage. The margin miss histogram shows the same thing for point margins: a distribution centered on zero means no systematic bias toward either side.