Walk-forward validation: the model is retrained on all seasons before year N, then predicts every year-N game blind. Benchmarked against actual closing lines — nothing in-sample is ever published.
