A flip occurs when a model switches its predicted winner between collection windows. Not all flips are bad - if a star player is scratched 2 hours before tipoff, you want your prediction system to update. But flips that happen without major news are a sign of an unstable prior.

In comparable prior-year data, models with higher flip rates in the T-24h to T-6h window correlate with lower overall accuracy. The best predictors tend to have stable picks at T-24h and T-6h, then update decisively at T-1h when late-breaking information is available.

We publish flip rate data for each model by round on the Prediction Drift page. Early results appear once Round of 64 games complete.

Separate from winner flips, we track confidence shifts: cases where a model keeps the same pick but drops or raises its confidence by 10+ points. These sub-flip signals often predict a pick reversal in the next window, and are visible in the drift chart per game.