Taming Language Models for Predicting Stock Price With Causality Guidance

Abstract

Predicting stock price trends not only needs qualitative analysis of textual information but also requires quantitative analysis of historical prices. Although recent advancements in language models have demonstrated their potential across various predictive tasks, current methods mainly focus on using a single modality with limited exploration of time series data in language models. To address both quantitative and qualitative analysis of stock price predicting tasks, we propose a multi-modality framework, the Joint Learning Time Series and Language Models (JLTSL) framework, which aligns time series information to language space. First, we encode time series data into temporal tokens, which can be used in joint training with textual tokens in language models. Second, we identify the presence of Granger Causality between time series sequences and their corresponding language representation. We utilize the Granger Causality test to select the prompt that has the strongest causality with the time series features to improve model performance. Our empirical evaluation across multiple datasets confirms that the JLTSL framework significantly outperforms existing benchmarks, underscoring its efficacy in predicting stock price trend.