Exogenous Dropout: A Simple, Strong Baseline for Corruption-Robust Time Series Forecasting with Covariates
Exogenous Dropout is a model-agnostic training intervention that randomly zeros entire exogenous covariate channels during training to improve robustness against data corruption. The method significantly enhances resilience to Gaussian noise, temporal misalignment, and missing data across electricity, hydrology, and meteorology domains without sacrificing clean accuracy. An unbounded model trained with Exogenous Dropout outperforms specialized bounded architectures like BoundEx, proving that exp
Analysis
TL;DR
- Exogenous Dropout is a model-agnostic training intervention that randomly zeros entire exogenous covariate channels during training to improve robustness against data corruption.
- The method significantly enhances resilience to Gaussian noise, temporal misalignment, and missing data across electricity, hydrology, and meteorology domains without sacrificing clean accuracy.
- An unbounded model trained with Exogenous Dropout outperforms specialized bounded architectures like BoundEx, proving that explicit architectural constraints are unnecessary for robustness.
- The authors release a new corruption-robustness benchmark and recommend Exogenous Dropout as a standard baseline for time series forecasting with covariates.
Why It Matters
This research addresses a critical deployment failure mode in time series forecasting where models relying on external variables collapse under real-world data imperfections. By demonstrating that simple training interventions can replace complex architectural fixes, it offers practitioners a low-cost, high-impact strategy to build more reliable forecasting systems. This shifts the focus from designing intricate fusion mechanisms to improving training protocols for better generalization under distribution shifts.
Technical Details
- Methodology: Exogenous Dropout involves masking whole exogenous feature channels (setting them to zero) randomly during the training phase, forcing the model to learn to rely on endogenous signals or adapt to partial information.
- Comparative Analysis: The approach is compared against BoundEx, a complex architecture featuring learnable gates, residual fallbacks to endogenous backbones, and per-channel FiLM modulation.
- Evaluation Domains: Performance is validated across three distinct sectors: electricity-price forecasting, reservoir hydrology, and meteorology.
- Corruption Types Tested: Robustness is measured against Gaussian noise, temporal misalignment, and scenarios where exogenous channels are entirely missing.
- Theoretical Insight: Representation-level bounds and ablation studies indicate that the robustness gain comes from the training dynamic rather than architectural boundedness, as the unbounded model with dropout surpassed the bounded counterpart in all tested domains.
Industry Insight
- Simplify Model Design: Practitioners should consider replacing complex exogenous fusion modules with simpler architectures combined with robust training techniques like Exogenous Dropout to reduce engineering overhead and maintenance costs.
- Prioritize Training over Architecture: When dealing with noisy or unreliable external data sources, investing in robust training strategies may yield higher returns than building sophisticated, hard-to-tune architectural components.
- Adopt New Benchmarks: Teams deploying time series models should integrate corruption-robustness testing into their validation pipelines, using the newly released benchmark to ensure models perform reliably under real-world data degradation.
Disclaimer: The above content is generated by AI and is for reference only.