A model can explain last month perfectly and still be useless next week. Handicapping model validation is how you find out which kind you built. Start by picking one target, such as a margin, a total or a win probability, and writing it down before you look at results. Then test forward in time, never with random splits, and score the forecasts with Brier score or log loss. The model has to beat a fair benchmark, like the no-vig market price, using only information available at the same moment. If a complicated model can't beat a simple one on data it never saw, it hasn't earned a single bet.
Learning objectives
Understand the core method; reproduce the worked example; diagnose at least two failure modes; apply a testable practice workflow.
