Hunter's Blend of Models
Blend of some models to help with WxChallenge forecasts! Made by Hunter Hayden
HBM downloads raw GRIB2 files for the GFS, NAM, and HRRR models directly from NOAA/AWS, and scrapes the USL model. Instead of using nearest-neighbor approximations, it mathematically interpolates the complex 1D and 2D model grids to extract the exact values for the target station's specific latitude and longitude.
Before making a prediction, HBM looks back at the last 10 days. It compares what each model predicted against the actual verified observations (fetched via Meteostat) to calculate each model's Mean Absolute Error (MAE) and Mean Bias Error (MBE) for that specific station.
HBM calculates an inverse-MAE weight for each model. If a model has been highly accurate over the last 10 days (low MAE), it gets a larger percentage of the final vote. If a model has been struggling, its influence on the final blend is mathematically reduced.
Models often have consistent flaws (like the 3km NAM consistently over-predicting wind gusts). In Bias Corrected mode, HBM takes the historical Mean Bias Error (MBE) and subtracts it from the raw model output before applying the weights, to attempt to reduce model flaws.