Training-Free Motion-Bias Calibration for Precipitation Nowcasting: A Negative Result

Abstract

Deep learning has advanced precipitation nowcasting, yet deterministic models still underperform cascaded diffusion approaches. We hypothesize that deterministic models may exhibit systematic motion bias---under- or over-advecting weather patterns---correctable by post-hoc warping. We propose Motion-Bias Calibration (MBC), a training-free method that estimates motion bias via optical flow and corrects predictions through learned warping. Testing MBC on EarthFormer with SEVIR, we find a negative result: the fitted speed-scale parameter $\alpha = 0.921$ indicates no systematic motion bias ($|\alpha - 1| < 0.1$), and all MBC variants degrade Critical Success Index (CSI) metrics by 0.0014--0.0018 relative to the raw baseline. A random-direction warp control confirms these effects are interpolation artifacts rather than motion correction. The hypothesis that deterministic nowcasters exhibit correctable motion bias is refuted for this model-dataset combination, demonstrating that the performance gap to state-of-the-art cascaded methods requires fundamentally different approaches.