Calibrated Confidence
A confidence score is calibrated when its stated probability matches the observed frequency — 70% confident predictions should be right about 70% of the time.
Most models output a confidence or probability, but raw model scores are usually poorly calibrated — a model claiming 90% confidence might only be right 60% of the time. Calibration corrects this so the numbers mean what they say.
Calibration is essential when sizing positions by conviction. endeavr.ai calibrates ensemble confidence on held-out data using isotonic and Platt-style methods, so a high-confidence prediction is genuinely more reliable than a low-confidence one.