RAF: A Model Agnostic Framework for Retrieval Augmented Zero Shot Time Series Forecasting

Foundation models have recently advanced zero shot time series forecasting, offering the ability to generalize without task specific training. However, in healthcare settings, where data are highly heterogeneous, exhibit regime shifts, and often contain rare but clinically critical events, these models frequently underperform. We propose Retrieval Augmented Forecasting (RAF), a model agnostic framework that strengthens foundation model predictions. RAF constructs a bank of past trajectories, retrieves nearest neighbor continuations using Euclidean or Dynamic Time Warping similarity, and blends them with foundation model forecasts through a data driven weighting scheme. The method requires no architectural changes, making it readily deployable within existing clinical forecasting pipelines. Across physiological and epidemiological datasets, including vital signs and hospital admission series, RAF consistently improves zero shot accuracy for four state of the art foundation models (Chronos, Lag Llama, MOMENT, and Toto). These gains highlight retrieval augmentation as a lightweight yet effective strategy for enhancing the robustness and clinical utility of time series foundation models in health applications. GitHub repository: https://github.com/anonymous2608878/raf