Recent Time-Series Research Summaries (Oct 25‑31 2025)

FreIE: Low‑Frequency Spectral Bias in Neural Networks for Time‑Series Tasks

Paper: http://arxiv.org/pdf/2510.25800v1

Code: https://github.com/Chenxing-Xuan/FreLE

Problem: Multivariate time‑series exhibit strong autocorrelation, which makes long‑term forecasting challenging.

Evidence: Recent studies incorporate frequency‑domain information to assist long‑term forecasting. Independent observations reported a spectral bias in neural networks, where models first fit low‑frequency components before high‑frequency ones.

Analysis: The authors conducted extensive empirical experiments measuring spectral bias across a variety of mainstream time‑series models (including Transformer‑based, convolutional, and recurrent architectures). Results showed that almost all evaluated models exhibited the low‑frequency bias, indicating that the bias is a cross‑model phenomenon rather than specific to any particular architecture.

Proposed method: FreLE (Frequency Loss Enhancement) is introduced as a plug‑in loss module. It adds explicit frequency regularization (directly penalizing frequency‑domain errors) and implicit frequency regularization (encouraging the model to capture high‑frequency components during training). The module can be attached to existing loss functions without altering the underlying model architecture.

Result: Large‑scale experiments on multiple benchmark time‑series datasets demonstrate that models equipped with FreLE achieve consistently lower forecasting error than the same models without the module, confirming that mitigating spectral bias improves generalization.