This is the authors' abstract. We don't add key points for this paper.
Long-term Time Series Forecasting (LTSF) plays a crucial role in various fields, including energy consumption and weather prediction. Recent advancements in deep learning, particularly transformer models, have significantly improved LTSF performance. However, these models often face criticism for their lack of transparency, limited interpretability, and high computational requirements, which can hinder their practical application. Simple linear models have emerged as a promising alternative, demonstrating excellent performance in LTSF while offering greater simplicity. Yet, as these models become more complex, they tend to lose their advantages. To address these challenges, we introduce SRLinear, a lightweight and interpretable LTSF framework based on symbolic regression. Our approach first breaks down time series data into seasonal and trend components, then divides these into smaller segments. We use symbolic regression to extract meaningful and interpretable features from these components, which are then combined and processed through a single linear layer for future predictions. This lightweight design ensures computational efficiency, while symbolic regression generates symbolic expressions, providing clear insights into the feature extraction process and enhancing model interpretability. We tested SRLinear on 8 real-world datasets, achieving state-of-the-art results on multiple benchmarks. Importantly, our framework produces highly interpretable features that can be easily visualized, enhancing the overall transparency of the prediction process. Our code is publicly available at https://github.com/fantasy-99/SRLinear.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: