Calibration & Track Record

Reliability, Brier decomposition, track record, and AI-vs-benchmark performance
Window
Category
Platform
Why the VNX ensemble — selective by design
Overall Brier
Lower is better
Reliability
↓ lower is better
Resolution
↑ higher is better
Mean CLV
Signal-time proxy
Predictions
Resolved
By Confidence — where the AI commits vs hedges
By Horizon — calibration across resolution timescales
Platform Calibration — AIA Forecaster Reliability
Brier Decomposition (Murphy)
Brier decomposition requires predictions spanning multiple probability bins across resolved markets. Data accumulating from AIA Forecaster resolved predictions.
Le 4-Component Decomp — where the residual variance lives
Skill vs Luck — bootstrap decomposition of track-record edge
Miss AnalysisLargest calibration errors — the model’s biggest misses, tracked for rigor (all-time; date filter not applied)
MarketPredicted %ActualMistake TypeSeverity
No resolved predictions with significant error