Abstract
Foundation models (FMs) have emerged as a promising approach for time seriesforecasting. While effective, FMs typically remain fixed during deployment dueto the high computational costs of learning them online. Consequently, deployedFMs fail to adapt their forecasts to current data characteristics, despite theavailability of online feedback from newly arriving data. This raises thequestion of whether FM performance can be enhanced by the efficient usage ofthis feedback. We propose ELF to answer this question. ELF is a lightweightmechanism for the online adaption of FM forecasts in response to onlinefeedback. ELF consists of two parts: a) the ELF-Forecaster which is used tolearn the current data distribution; and b) the ELF-Weighter which is used tocombine the forecasts of the FM and the ELF-Forecaster. We evaluate theperformance of ELF in conjunction with several recent FMs across a suite ofstandard time series datasets. In all of our experiments we find that using ELFimproves performance. This work demonstrates how efficient usage of onlinefeedback can be used to improve FM forecasts.