Exogenous Dropout: Simple training trick makes time series models robust to corrupted data
A single-line fix outperforms complex architectures for handling noisy, missing, or misaligned covariates.
Researchers Hao Hu and Xue-shan Ai propose Exogenous Dropout, a model-agnostic training method that randomly zeroes entire exogenous channels during training. Tested on electricity-price forecasting, reservoir hydrology, and meteorology, it substantially improves robustness under Gaussian noise, temporal misalignment, and fully missing channels—while preserving clean accuracy. Applied to a dual-correlation network, it outperforms a deliberately strong bounded architectural foil called BoundEx, which uses a learnable gate, a fallback residual, and per-channel FiLM modulation. The authors release a corruption-robustness benchmark and recommend exogenous dropout as a simple, strong baseline.
- Exogenous Dropout randomizes entire covariate channels during training, making models robust to noise, misalignment, and missing data without specialized architecture.
- Outperforms BoundEx—a purpose-built bounded model with gating, residuals, and FiLM modulation—on electricity, hydrology, and meteorological benchmarks.
- Works across model types (LSTM, transformer, dual-correlation) and preserves clean accuracy; authors release a corruption-robustness benchmark.
Why It Matters
Pragmatic robustness for production time series forecasting where covariate corruption is common, with no architectural overhead.