Our errors, in public.
Forecast Freeway's road-surface forecasts come from pavecast, a pavement energy-balance model built clean-room from the published open science — chiefly Crevier & Delage (2001), the paper behind Environment Canada's METRo, plus the FHWA pavement-thermal literature. Nothing in it is derived from any commercial road-weather product. Where METRo integrates one station at a time, pavecast solves a 12-layer pavement column and a bridge-deck twin at all 5,580 network nodes in one vectorized pass, every hour.
A forecast you can't check is marketing. So this page scores ours against state-DOT RWIS pavement sensors — the thermometers embedded in the road itself — every day, and prints the errors whether they flatter us or not.
The Live Scorecard
ROLLING 30 DAYS · UPDATED DAILYEvery day the current model run's next-30-hour pavement temperatures are banked per sensor, and yesterday's banked forecasts are scored against what the sensors actually read. Bias is model minus sensor; the baseline is the naive call that pavement temperature equals air temperature — the assumption every generic weather app silently makes, and the one a pavement model exists to beat.
The Shakedown Hindcast
20 RWIS STATIONS · FIXED WINDOWBefore the model went live it re-ran a week of history at twenty RWIS stations from Vermont to the Mojave and was scored against their pavement sensors — versus persistence (tomorrow = today, the hardest naive baseline in a calm week) and versus air temperature. This table is frozen: it is the audit that shipped the model.
What This Model Is Not
Read before trusting- Untreated pavement. No salt, no brine, no plows. A treated Interstate will be better than we say — which is the correct direction for a warning product, and v2 territory.
- No traffic heat. Tire friction and engine radiation keep a busy urban freeway a few degrees warmer than our column; a dead-quiet rural bridge at 4 AM matches it closely.
- Parameterized sky. Solar and infrared radiation come through a cloud-cover parameterization (Haurwitz, Kasten–Czeplak, Prata), not a radiative-transfer model. Broken-cloud afternoons are the weak spot, and the scorecard shows it.
- Frost season unvalidated until winter. The physics of frost deposition is in the model; the September shakedown could not exercise it. The live scorecard above is exactly how we — and you — will find out.
- Sensor truth is imperfect. RWIS pavement sensors drift and die; stations whose pavement reading is wildly inconsistent with their own air temperature are excluded from scoring, and the exclusion rule is public (in the builder source).
Cadence
When things updateThe pavement model re-runs the whole network — 48 h URMA spin-up, NDFD forcing out past two days — and republishes the road-temp, surface-state and bridge-frost layers.
The icing nowcast: the latest MRMS radar sweep intersected with the current modeled pavement, tiered watch / likely / ice.
This scorecard: matured forecasts scored against the RWIS archive, rolled over 30 days, published win or lose.