Better earthquake forecasts by ditching Poisson's one-size-fits-all assumption

Weekly earthquake forecasting typically assumes earthquakes follow a Poisson distribution with the same variability everywhere—a mathematically convenient but empirically wrong assumption. This work replaces it with neural negative binomial regression that learns different dispersion (variability) for each grid cell in Central Asia. The new model improves the standard benchmark (CRPS) by 12.5% and better calibrates forecasts for rare, large earthquakes—exactly where predictions matter most.