TimeLLM: Time Series Forecasting by Reprogramming Large Language Models

TL;DR

TimeLLM is a methodology that reprograms pre-trained large language models (LLMs) to perform time series forecasting.

Problem

Existing time series forecasting models are specialized for specific domains, resulting in low generalization performance, and the scarcity of large-scale time series datasets hinders the development of multimodal models. Meanwhile, LLMs excel at understanding long token sequences, but directly applying them to time series signals is challenging.

Approach

TimeLLM transforms time series signals into text prototypes (reprogramming) to align with LLM inputs. It also introduces the Prompt-as-Prefix (PaP) technique to enhance contextual information and leverages the reasoning capabilities of LLMs for time series forecasting. The LLM backbone remains frozen to maintain efficiency.

Results & Contribution

On various time series forecasting benchmarks, TimeLLM outperforms existing specialized models, demonstrating the potential of LLMs for time series prediction. As the first large-scale multimodal approach in the time series domain, it suggests future research directions.