A hot streak proves very little about a method. +EV backtesting asks a harder question: does a rule written in advance produce accurate probabilities on data it never saw? The lesson shows how to freeze your rule first, split data by time rather than random rows, and store quote and kickoff timestamps so history is not rebuilt after the fact. You work a small Brier score example by hand and learn what calibration bins can and cannot show. It also separates forecast quality from execution quality, and explains why closing-line value is a useful clue but not proof. The goal is a test you trust enough to act on, or to throw out.
Learning objectives
- Reproduce the lesson calculations.
- Identify the assumptions and conditions that can invalidate the modeled edge.
- Design an auditable research or practice workflow.
