Why global models lose at your doorstep
ICON, GFS and ECMWF are extraordinary machines — and they are solving a different problem. A grid cell averages over hills, valleys, lakes and heat islands. Your garden has a microclimate the cell cannot represent: the frost pocket, the afternoon shading, the wind funnel between buildings. The model isn't wrong about the box; the box is wrong about you.
The bet: one station, taken seriously
Our approach is station-level post-processing: take the global models' guidance, then correct it with what nine years of one sensor's real history says about how this exact spot deviates from its grid box. The entire pipeline runs on a Raspberry Pi. No cluster, no cloud dependency — the intelligence is in the method, not the hardware bill.
- Measured, not simulated: every headline number is evaluated on held-out 2026 data the model never saw during fitting.
- Beats all three operational references — ICON, GFS and ECMWF — on the target variables at the station.
- Same lab discipline: where the signal doesn't support a confident forecast, the system abstains rather than inventing precision.
The part most forecasts skip: honesty about uncertainty
A forecast that is sometimes brilliant and sometimes silently wrong is worse than useless — you can't tell which day you're getting. KODON Weather ships with an explicit uncertainty gate: it separates what it can predict from what it can't, so a confident forecast actually means something. No fake precision on a day the sky won't commit.
See today's forecast, and what the system refuses to claim, at weather.koscak.ai.