Hit rate, accuracy and profit are three different questions, and a model can ace one while failing the others. NBA model validation is how you find out which one you actually have. You will freeze a recipe before looking at results, build a timeline using only what was knowable at decision time, and score forecasts with Brier numbers on a tiny example. Then you will learn what closing-line checks can and cannot tell you, why a lucky streak can look like skill, and how to run a backtest that accounts for real prices. Treat every metric as a diagnostic, not proof. That mindset keeps a good-looking spreadsheet from fooling you.
Learning objectives
- Separate forecast accuracy, calibration and profitability.
- Build an honest timeline and an execution-aware backtest.
- Use closing-line value and drift monitoring as diagnostics, not proof.
