Public scoreboard
Calibration.
How accurate are our forecasts? Every prediction we publish is locked at publish time. We grade ourselves in public.
Time range
|
Category
Last update 2026-05-19 11:42:01Z
Reliability curve
Predicted probability vs. realized frequency · decile buckets
Realized frequency
Perfect calibration
95% confidence interval
Headline numbers
Forecasts tracked412
Brier score0.082
Log loss0.31
Calibration error (avg)±2.1%
Forecast horizon1–42d
By confidence bucket
Each row links to the individual forecasts and their outcomes.
| Predicted | Realized | Delta | n |
|---|---|---|---|
| 50–59% | 53% | −0.5 | 44 |
| 60–69% | 64% | −0.5 | 89 |
| 70–79% | 72% | −2.5 | 110 |
| 80–89% | 87% | +2.5 | 98 |
| 90–99% | 94% | −0.5 | 71 |
Per category
Same scoring, sliced.
Weather n = 284
0.061 Brier
Cal. err ±1.6% · best-calibrated category
Econ n = 84
0.118 Brier
Cal. err ±3.2% · noisier; thin sample at 80–89
Politics n = 44
0.142 Brier
Cal. err ±4.4% · small n; treat with caution
How these numbers are computed
Forecasts are written to validation_cf6 at publish time with a forecast_id, the predicted probability, and the resolving market. Outcomes attach when the market settles. Nothing is retroactively edited.