SST-LLM: time series forecasting based on large language models

Time series forecasting is very important for research and applications in real life. Traditional time series prediction methods based on deep learning face challenges in cross-domain, few-shot learning, and zero-shot learning scenarios. Therefore, this paper proposes a time series prediction method based on large language modes (LLMs). Specifically, we freeze the pre-trained LLM, keeping the backbone language model unchanged and not fine-tuning it during the training process, thereby avoiding the need for large amounts of data when training the model. At the same time, we use moving average method and autocorrelation analysis to extract features from the time series and merge them with the original data. This approach not only highlights the trends and seasonality of the time series data but also effectively suppresses data noise. Additionally, we propose a loss function that can enhance the model's ability to extract trends from data. In this way, the LLM can focus more on the actual trends in the data and reduce noise interference. In addition, we employ a time-step patching technique to extract local semantic information and reduce information redundancy. Finally, we input the processed time series data into the frozen LLM, extract the data from the last hidden layer, and obtain the final forecasting results through the output layer. Our research results indicate that the proposed model outperforms other models, achieving superior forecasting results in long-term, short-term, and few-shot forecasting. The code and data are publicly available on GitHub1.