Calibration
Scores you can audit
We publish overall Brier and ROC-AUC from closed windows, plus Model A risk AUC 0.6339. We do not invent per-class Brier numbers on this site.
Brier
Mean squared error of probabilities. 0 is perfect; 0.25 is a fair coin on a binary event. A wide mixed sample can sit near 0.25 — that is a measurement, not a slogan.
ROC-AUC / Model A
0.5 is chance. Short windows where every outcome is one class make AUC unreadable. Model A (risk) AUC 0.6339 is the named out-of-sample anchor — see Dual Model cards on /statistika.
What we will not print
No class Brier tables (geopolitics / supply chain / disasters) unless they come from a published calculation. No “losses prevented”.