WEBVTT

00:00:00.000 --> 00:00:04.890
The saved example contains fifty forecasts at fifty percent and twenty five

00:00:04.890 --> 00:00:09.780
observed successes. Observed frequency is twenty five divided by fifty, or fifty

00:00:09.780 --> 00:00:14.670
percent. Bars share seven pixels per percentage point, so forecast and observed

00:00:14.670 --> 00:00:19.560
rate have equal lengths and zero calibration gap. With forty successes instead,

00:00:19.560 --> 00:00:24.450
the observed rate is eighty percent and the gap thirty points. These

00:00:24.450 --> 00:00:29.341
are entered counts, not a measured forecasting dataset; one bucket being calibrated

00:00:29.341 --> 00:00:34.231
does not establish discrimination across cases.

00:00:34.231 --> 00:00:38.251
The source turns probabilities at least fifty percent into a yes answer.

00:00:38.251 --> 00:00:42.272
For thirty five successes in fifty forecasts, the observed rate is seventy

00:00:42.272 --> 00:00:46.293
percent. A forecast of sixty percent predicts yes and has seventy percent

00:00:46.293 --> 00:00:50.314
binary accuracy; forty percent predicts no and has thirty percent accuracy. Bars

00:00:50.314 --> 00:00:54.334
share six pixels per percentage point. The threshold changes the answer without

00:00:54.334 --> 00:00:58.355
changing outcomes; this is not a claim that higher confidence guarantees a

00:00:58.355 --> 00:01:02.376
correct forecast.

00:01:02.376 --> 00:01:07.072
For a bucket with half successes, the source Brier score averages squared

00:01:07.072 --> 00:01:11.768
probability errors. At fifty percent the mean score is zero point two

00:01:11.768 --> 00:01:16.465
five. At ninety percent on the same half-success bucket it is zero

00:01:16.465 --> 00:01:21.161
point four one. Bars share one thousand pixels per score unit. Lower

00:01:21.161 --> 00:01:25.857
is better for this entered calculation: overconfidence in this hypothetical bucket increases

00:01:25.857 --> 00:01:30.553
squared error. The source also exports a Wilson observed-rate interval, not a

00:01:30.553 --> 00:01:35.250
guarantee about future forecasts. Trial and success counts are rounded and invalid

00:01:35.250 --> 00:01:39.946
inputs block calculation and export.
