A 70 percent forecast that loses is not automatically wrong, so win rate makes a poor scoreboard. Betting model drift is what happens when inputs or results change over time, and spotting it early means measuring the right things. This lesson teaches the Brier score, which squares the gap between forecast and outcome and punishes confident misses. You compare against a simple base-rate baseline, then group forecasts into calibration buckets and keep sample size in view, since forty games can look off from noise alone. You also split drift into three different problems and set alerts for broken data before results arrive. The aim is telling a real model problem from plain bad luck.
Learning objectives
You will learn how to score probability forecasts, check whether they are calibrated, and tell a real model problem from ordinary bad luck. You will also set up simple alerts for broken data before results even come in.
