Lightweight Online Adaption for Time Series Foundation Model Forecasts

Thomas L Lee, William Toner, Martin Asenov, Artjom Joosen, Rajkarn Singh

Keywords: Foundation Models, Time Series, Adaption, Online Forecasting, Online Learning

TL;DR: We propose ELF an efficient method to improve the forecasts of time series foundation models at deployment by using online feedback.

Abstract:

Foundation models (FMs) have emerged as a promising approach for time series forecasting. While effective, FMs typically remain fixed during deployment due to the high computational costs of learning them online.
Consequently, deployed FMs fail to adapt their forecasts to current data characteristics, despite the availability of online feedback from newly arriving data. This raises the question of whether FM performance can be enhanced by the *efficient* usage of this feedback. We propose *ELF* to answer this question.

*ELF* is a lightweight mechanism for the online adaption of FM forecasts in response to online feedback, comprising two parts: **a)** the *ELF-Forecaster*, which learns the current data distribution; and **b)** the *ELF-Weighter*, which combines forecasts from the FM and the ELF-Forecaster. We evaluate ELF with several recent FMs on standard time series datasets and find that it improves performance in *all* cases. This work shows that efficiently leveraging online feedback can enhance FM forecasts.

Submission Number: 10