Effectiveness of LLMs for Financial Forecasting

Determine whether large language models, including transformer-based systems such as GPT-4 and LLaMA-family models, provide reliable and superior performance for financial time series forecasting, and characterize the market regimes, data modalities, and task setups under which these models succeed or fail relative to traditional machine learning and deep learning baselines.

Background

The survey reviews recent studies applying LLMs to financial forecasting and highlights mixed evidence. Some works report strong performance of models like GPT-4 and LLaMA2-7B on stock-return prediction using news headlines, while others find that ChatGPT underperforms traditional and state-of-the-art approaches in zero-shot multimodal stock movement prediction.

Based on these conflicting results across datasets, modalities, and evaluation designs, the authors explicitly state that the debate remains unresolved, motivating the need for systematic assessment of when and how LLMs are effective for forecasting financial time series.