How do large language models bring disruptive change to time series forecasting? A survey and framework

While conventional models have already demonstrated exceptional performance in time series forecasting, several challenges remain to be tackled. The emergence of large language models (LLMs) has introduced disruptive changes to forecasting methodologies, including architectural frameworks and forecasting processes. To conclude existing literature regarding the application of LLMs in time series forecasting, this study, distinct from prior reviews, offers a comprehensive and up-to-date general pipeline block structure that integrates various modeling techniques in time series forecasting with LLMs, with specific emphasis on data scenario usage. To guide practical applications, this study proposes three collaborative patterns for incorporating LLMs in time series forecasting workflows and provides a comparison of LLMs’ performances with conventional models. Overall, this study organizes and evaluates the latest research advancements, delivering key insights and propositions for the potential future development of LLMs in the field of time series forecasting.