Title: Stationarity-Aware Retrieval-Augmented Time Series Forecasting

Abstract:

While traditional time series forecasting leverages historical patterns, real-world data frequently undergoes non-stationarity and regime shifts that pose significant challenges to fully parametric models. Drawing inspiration from Retrieval-Augmented Generation (RAG), recent approaches enhance forecasters by retrieving pertinent historical segments to serve as external evidence during inference. However, the inherent non-stationarity of time series data means that a highly similar past segment does not guarantee a similar future outcome. Consequently, retrieval strategies relying solely on similarity are fragile and often result in redundant information.

To address these limitations, we introduce Stationarity-Aware Retrieval-Augmented Time Series Forecasting (SARAF). This framework adaptively manages the trade-off between relevance and diversity during retrieval. SARAF initially constructs a candidate pool based on temporal similarity, enhanced by time alignment. It then employs a diversity-aware selection mechanism to encompass heterogeneous historical regimes, with the intensity of diversification automatically adjusted according to dataset-level stationarity metrics. Furthermore, the model utilizes stationarity-aware aggregation to effectively fuse the retrieved future segments.

Extensive evaluations across eight real-world datasets demonstrate that SARAF delivers competitive forecasting performance. It significantly enhances both average accuracy and robustness compared to strong baselines, offering particularly pronounced advantages in challenging, non-stationary environments.

Code: https://github.com/ShiqiaoZhou/SARAF

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC