Are Language Models Actually Useful for Time Series Forecasting?

TL;DR

This paper critically evaluates the effectiveness of large language models in time series forecasting, revealing that simpler models often outperform or match LLMs, questioning their practical utility in this domain.

Contribution

The study systematically compares LLM-based forecasting methods with simpler models, demonstrating that LLMs do not provide significant advantages and are often less efficient.

Findings

Removing LLM components does not harm performance

Pretrained LLMs do not outperform models trained from scratch

Simple attention-based encoders perform similarly to LLMs

Abstract

Large language models (LLMs) are being applied to time series forecasting. But are language models actually useful for time series? In a series of ablation studies on three recent and popular LLM-based time series forecasting methods, we find that removing the LLM component or replacing it with a basic attention layer does not degrade forecasting performance -- in most cases, the results even improve! We also find that despite their significant computational cost, pretrained LLMs do no better than models trained from scratch, do not represent the sequential dependencies in time series, and do not assist in few-shot settings. Additionally, we explore time series encoders and find that patching and attention structures perform similarly to LLM-based forecasters.