Prediction Drift Analysis - AI March Madness 2026
Prediction drift measures how AI model picks change between the T-24h, T-6h, and T-1h collection windows before each NCAA Tournament game. A flip occurs when a model switches its predicted winner between windows. Confidence shifts track changes in stated confidence without a winner change.
Models with high flip rates in the T-24h to T-6h window tend to have lower overall accuracy. The most reliable predictors maintain stable picks at T-24h and T-6h, then update decisively at T-1h when late-breaking information like injury reports becomes available.
The drift chart visualizes each model's prediction stability across all games played. GPT-4o, Gemini 2.5, and Perplexity Sonar Pro each show distinct drift patterns based on their underlying source dependencies and reasoning approaches.
- Flip Rate
- The percentage of games where a model changed its predicted winner between collection windows. Lower flip rates generally correlate with higher accuracy.
- Confidence Shift
- A change of 10 or more percentage points in stated confidence between collection windows without changing the predicted winner.
- Prediction Stability
- The inverse of flip rate. Higher stability indicates more consistent predictions across collection windows.
