This paper proposes ELATE (Evolutionary Language Model for Automated Time-series Engineering), a novel method for automating feature engineering in time-series forecasting using machine learning models. ELATE automates feature engineering for time-series data by leveraging language models within an evolutionary framework. It automates the traditionally manual and time-consuming feature engineering process, generating features using time-series statistics and feature importance measures, and removing redundant features. Experimental results show an average 8.4% improvement in prediction accuracy across various domains.
