Exogenous Dropout Boosts Robustness in Time Series Forecasting.

Time series forecasting models that rely on exogenous covariates often struggle in real-world deployments when these covariates are noisy, temporally misaligned, or entirely absent. This fragility can cause strong exogenous-fusion models to perform worse than simpler, endogenous-only approaches. Researchers investigated whether specialized architectures are necessary for robustness or if a simpler training intervention could suffice. They propose "exogenous dropout," a straightforward, model-agnostic technique where entire exogenous channels are randomly zeroed out during the training process. This method aims to make models less dependent on perfect covariate inputs, forcing them to learn more robust representations. Extensive experiments across diverse domains like electricity-price forecasting, reservoir hydrology, and meteorology demonstrated that exogenous dropout substantially improves robustness against Gaussian noise, temporal misalignment, and completely missing channels. Crucially, it achieves these gains without sacrificing accuracy on clean data and even outperforms more complex, architecturally bounded solutions. The findings suggest that explicit architectural boundedness is not essential for achieving significant robustness in covariate-dependent time series forecasting.