Sample size and calibration

With 22 forecasts, a true 55% forecaster could easily show 45% or 65% by luck alone (the 95% range is roughly ±20 points). That is why WinSight marks any live sample below 100 settled forecasts as not yet strong evidence. Calibration asks: of all outcomes given ~30%, did about 30% happen? Group forecasts into bins and compare predicted with observed frequencies. A well-calibrated model sits near the diagonal; it can be calibrated and still have modest hit rate.