This paper explores the potential of large language models (LLMs) in advancing time series analysis. Time series analysis is crucial for understanding complex systems and supporting decision-making. While traditional time series models rely on domain knowledge and extensive tuning, LLMs offer new possibilities for efficient and universal time series analysis. The paper argues that LLMs can revolutionize time series analysis by enhancing data and models, acting as effective predictors, and serving as next-generation agents. It emphasizes the need for trust in LLM-based approaches and highlights the integration of LLMs with existing time series technologies. The paper discusses three key roles of LLMs in time series analysis: (1) as data and model enhancers, improving data understanding and model performance; (2) as effective predictors, leveraging internal knowledge and reasoning for various prediction tasks; and (3) as next-generation agents, actively engaging in and transforming time series analysis. It also explores the challenges and opportunities in integrating LLMs with time series data, including data sparsity, noise, and the need for efficient training. The paper presents a roadmap of time series analytical models, spanning four generations: statistical models, deep neural networks, pre-trained models, and LLM-centric models. It discusses various approaches for LLM-based time series prediction, including tuning-based and non-tuning-based methods, and highlights the potential of LLMs in enhancing time series analysis through in-context learning and prompt engineering. The paper also addresses the challenges of using LLMs as time series agents, including their limitations in understanding complex patterns and the risk of hallucination. It proposes strategies for improving LLM-based time series agents, such as aligning time series features with language model representations, fusing text embeddings and time series features, and teaching LLMs to utilize external tools. The paper concludes that LLMs hold significant promise for time series analysis, but their reliability and effectiveness require further research and development. It emphasizes the importance of accountability, transparency, privacy, and ethical considerations in the application of LLMs to time series analysis. The paper calls for continued exploration of LLM-based approaches to advance time series analysis and develop robust, reliable systems for general-purpose use.This paper explores the potential of large language models (LLMs) in advancing time series analysis. Time series analysis is crucial for understanding complex systems and supporting decision-making. While traditional time series models rely on domain knowledge and extensive tuning, LLMs offer new possibilities for efficient and universal time series analysis. The paper argues that LLMs can revolutionize time series analysis by enhancing data and models, acting as effective predictors, and serving as next-generation agents. It emphasizes the need for trust in LLM-based approaches and highlights the integration of LLMs with existing time series technologies. The paper discusses three key roles of LLMs in time series analysis: (1) as data and model enhancers, improving data understanding and model performance; (2) as effective predictors, leveraging internal knowledge and reasoning for various prediction tasks; and (3) as next-generation agents, actively engaging in and transforming time series analysis. It also explores the challenges and opportunities in integrating LLMs with time series data, including data sparsity, noise, and the need for efficient training. The paper presents a roadmap of time series analytical models, spanning four generations: statistical models, deep neural networks, pre-trained models, and LLM-centric models. It discusses various approaches for LLM-based time series prediction, including tuning-based and non-tuning-based methods, and highlights the potential of LLMs in enhancing time series analysis through in-context learning and prompt engineering. The paper also addresses the challenges of using LLMs as time series agents, including their limitations in understanding complex patterns and the risk of hallucination. It proposes strategies for improving LLM-based time series agents, such as aligning time series features with language model representations, fusing text embeddings and time series features, and teaching LLMs to utilize external tools. The paper concludes that LLMs hold significant promise for time series analysis, but their reliability and effectiveness require further research and development. It emphasizes the importance of accountability, transparency, privacy, and ethical considerations in the application of LLMs to time series analysis. The paper calls for continued exploration of LLM-based approaches to advance time series analysis and develop robust, reliable systems for general-purpose use.