the idea
National forecasts are made on a grid that's miles wide. In the desert, whether it rains on your street can be a completely different story from the grid cell next door. Beacon doesn't try to out-physics NOAA. It learns the local correction between what the global ensemble model (GEFS) predicts and what actually happens in one specific yard.
what i built
- a backyard weather station: a raspberry pi and sensor in a 3D-printed ventilated enclosure, logging every five minutes.
- two decades of training data: historical GEFS ensemble forecasts joined to local and surrounding ring-station observations.
- calibrated correction models per variable and lead time, validated with strict chronological walk-forward splits.
- a two-stage rain model: will it rain (calibrated probability), and if so how much, combined into one expected amount.
- a live daily pipeline that pulls the latest forecasts and observations, scores them, and logs every prediction for later grading.
the result
The rain model beats raw GEFS at every lead time from 1 to 10 days, cutting mean absolute error by about 14% and winning 46–48 of 50 walk-forward tests. Wind speed corrections improve on GEFS by about 25%.