📄 中文摘要
利用大语言模型（LLMs）及其输出，已成为各种机器学习任务的趋势，包括在语言模型出现之前就已解决的预测任务。该研究探讨了如何将LLM嵌入应用于时间序列预测，通过特征工程的方法，提升预测模型的性能。通过实证分析，研究表明，LLM嵌入能够有效捕捉时间序列数据中的潜在模式，从而改善预测精度。这为传统的时间序列分析方法提供了新的视角，展示了LLM在非语言领域的应用潜力。研究结果为机器学习实践者在特征选择和模型构建方面提供了有价值的参考。
大语言模型嵌入能否改善时间序列预测？一种实用的特征工程方法
📄 English Summary
Can LLM Embeddings Improve Time Series Forecasting? A Practical Feature Engineering Approach
The use of large language models (LLMs) and their outputs has become a trend in various machine learning tasks, including predictive tasks that were addressed long before the emergence of language models. This study investigates the application of LLM embeddings in time series forecasting through a practical feature engineering approach to enhance the performance of predictive models. Empirical analysis demonstrates that LLM embeddings effectively capture underlying patterns in time series data, thereby improving forecast accuracy. This provides a new perspective on traditional time series analysis methods and showcases the potential applications of LLMs in non-language domains. The findings offer valuable insights for machine learning practitioners regarding feature selection and model building.