Beyond MSE: Improving Precipitation Nowcasting with Multi-Quantile Regression

Researchers evaluated SmaAt-UNet for radar-based precipitation nowcasting over the Netherlands.

Training with multi-quantile pinball loss cut test-set MSE by 8.6% versus standard MSE training.

The quantile approach also produced useful upper-tail outputs, improving heavy-rainfall risk signaling.

The study shows quantile regression can boost forecast accuracy without changing the model architecture.