When does 80% actually mean 80%?
A coupon model ranks shoppers from least to most likely to redeem. Five groups of 100 shoppers get an offer. Can you line up the forecasts with what actually happened?
Try 1.00×, then push toward either end. The multiplier scales each group’s log odds.
● Forecast│ Actual
Group 11% / 10%
Group 216% / 30%
Group 350% / 50%
Group 484% / 70%
Group 599% / 90%
9.3 ppMean absolute calibration gap
0.181Brier score · lower is better
The high forecasts run high and the low forecasts run low. The model still ranks all five groups in the same order. Ranking alone would miss this error.
Simulated outcomes. The gap averages the five group errors in percentage points. The Brier score averages squared probability errors across all 500 shoppers.