← Buffalo dashboard

What the confidence labels mean

"High / medium / low" on a prediction card is not a feeling — it's the measured historical hit rate of that exact kind of prediction, replayed over 11½ years. ≥65% verified = high, ≥40% = medium, below = low.

Local-rain predictions — measured hit rates

GaugeCategoryVerifiedLabel shownSample
Boxleyslight51%medium2,734 predictions
Boxleymoderate59%medium1,393 predictions
Boxleylarge51%medium124 predictions
Poncaslight50%medium2,047 predictions
Poncamoderate44%medium613 predictions
Poncalarge29%low89 predictions
Pruittslight46%medium2,056 predictions
Pruittmoderate58%medium611 predictions
Pruittlarge36%low110 predictions
St. Joeslight43%medium3,009 predictions
St. Joemoderate43%medium730 predictions
St. Joelarge55%high49 predictions
Harrietslight43%low2,525 predictions
Harrietmoderate48%medium666 predictions
Harrietlarge61%medium44 predictions

Wave-tracker alerts

The wave tiers carry their calibration in their definition: over 11½ replayed years, warnings (band low end ≥ flood) verified as genuine floods 94–100% of the time at Pruitt, St. Joe and Harriet, and over 98% of warnings reached at least 60% of flood — blue-sky warnings essentially don't happen. Watches are the early tier: they catch 91–98% of floods, at the price that roughly half verify below flood stage — that's what an early heads-up costs.

Recession countdowns

Their label reflects actual forecast spread: high when the 25th–75th percentile range is tight and the countdown is under 3 days; low when the range is wide or the horizon is beyond a week (weather becomes the dominant unknown). Thin recent gauge data caps the label regardless.

Every number on this page is generated from the deployed model files and an 11½-year hourly replay (2014-12 → 2026-06, baseline v6). Generated 2026-07-02T22:57:24+00:00. Experimental — never rely on a single number for safety decisions.