arXiv:2610.17864v2 Announce Type: replace-cross Abstract: Long-term time series forecasting requires learning temporal dependencies that remain stable and generalizable under temporal distribution shifts. However, real-world time series are often noisy, non-stationary, and heterogeneous, making deep learning models prone to overfitting incidental temporal interactions in the training data.
