Exogenous Dropout — Simple defense for fragile time-series forecasters
Time series models that fuse external data (prices, weather, flows) fail hard when those covariates arrive noisy, delayed, or missing — but specialized architectures aren't the only fix.
Researchers propose exogenous dropout: randomly zero out external channels during training. Tested on electricity, hydrology, and weather forecasting, the method holds clean accuracy while gaining robustness to noise, temporal shifts, and full missing-data events.
Model-agnostic, no architectural change required — a training-only intervention that scales across domains.