TimeLLM: Time Series Forecasting by Reprogramming Large Language Models

Accurate forecasting of time series is essential to many dynamic real-world systems and has attracted extensive research attention. Unlike natural language processing or computer vision, where one large model can often address multiple tasks, most existing forecasting solutions are highly specialized and confined to the single time series data modality. Advancements in multimodal time series foundation models have significantly lagged behind other domains, mainly because large, high-quality time series corpora remain scarce. At the same time, recent evidence suggests that large language models (LLMs) excel at understanding and reasoning across long token sequences. Exploiting those capabilities for forecasting requires a principled way to bridge numeric time series signals and linguistic tokens. This chapter presents Time-LLM, a model reprogramming framework that repurposes frozen LLM backbones for general time series prediction. As the first multimodal, large-scale approach in the time series community, Time-LLM embeds raw time series (target) signals into text (source) prototypes, aligning the two modalities before the data enters the language model backbone. To strengthen the model's reasoning over time series, Time-LLM incorporates Prompt-as-Prefix (PaP), which augments contextual information and directs the reprogramming of input patches. The modified patches are then projected by the LLM to generate forecasts. Extensive evaluations demonstrate that Time-LLM serves as an effective multimodal forecaster, surpassing specialized baseline models.