Time series forecasting (TSF) is gaining significance in various applications. In recent years, many pre-trained large language models (LLMs) have been proposed, and some of them have been adapted for use in TSF. When applying LLMs to TSF, existing strategies with complex adapters and data preceding modules can increase training time. We introduce StreamTS, a highly streamlined time series forecasting framework built upon LLMs and decomposition-based learning. First, time series are decomposed into a trend component and a seasonal component after instance normalization. Then, a pre-trained LLM facilitated by the proposed BC-Prompt is used for future long-term trend prediction. Concurrently, a linear model simplifies the fitting of future short-term seasonal term. The predicted trend and seasonal series are finally added to generate the forecasting results. In our efforts to achieve zero-shot forecasting, we replace the linear prediction part with a statistical learning method. Extensive experiments demonstrate that our proposed framework outperforms many TSF-specific models across various datasets and achieves significant improvements over LLM-based TSF methods.
Song et al. (Fri,) studied this question.