SERAF boosts time series forecasting with dual retrieval of text and data

Time series forecasting models have long leveraged historical patterns to predict future values. Inspired by Retrieval-Augmented Generation (RAG), recent approaches retrieve relevant past segments to boost accuracy. However, under non-stationary conditions—where data distributions shift over time—simple numerical similarity often fails. To solve this, a team including Shiqiao Zhou, Zipeng Wu, Holger Schöner, Edouard Fouché, IAG Wilson, and Shuo Wang propose SERAF (Semantics-Enhanced Retrieval-Augmented Time Series Forecasting). Their key innovation: dual retrieval over both raw time series and their self-generated textual descriptions. The model retrieves two complementary sets of historical patterns and corresponding futures, then selectively combines them to guide predictions. This multimodal approach bridges numerical and semantic views, making forecasting more robust to regime changes and trend shifts.

Experiments across seven real-world datasets demonstrate that SERAF consistently beats state-of-the-art baselines, confirming the value of adding semantic context to numeric retrieval. The work has been accepted to the ICML 2026 Workshop on Forecasting as a New Frontier of Intelligence. While still an early-stage research paper, the framework opens a promising path for integrating language and time series data—potentially useful for financial forecasting, energy demand prediction, and IoT anomaly detection. As foundation models become better at generating descriptive text from time series, SERAF-like architectures could become a standard component in production forecasting pipelines.