Confidence Is a Frequency, Not a Promise
When the model publishes a 70% confidence pick, the claim is precise: across many thousands of structurally similar matchups, the favoured side wins seventy per cent of the time. By definition, the other thirty per cent must lose. A model that wins every 70% call is overfit; what looks like genius is actually noise dressed as signal.
The Coin That Knew Better
Imagine a coin weighted to land heads sixty per cent of the time. Flip it ten times and you might see five heads. That does not refute the bias — ten flips is not enough sample. You need hundreds of flips before the long-run rate is visible. Picks behave the same way.
What Variance Measures
Variance is the model’s estimate of how much noise lives inside a specific matchup. Low variance means stable inputs: an ace pitcher, a clean lineup, no weather. High variance means the call is structurally fragile: a bullpen game, shifting wind, late lineup uncertainty. Variance is published alongside confidence so you can read both.
The Discipline
Do not chase. Do not double down. Trust the long run. The picks that look painful in the moment are exactly the picks that make the year.