This paper explores the use of Large Language Models (LLMs) for generating equity stock ratings, a task traditionally performed by financial analysts. The authors employ GPT-4-32k (v0613) with a training cutoff before the evaluation period to prevent information leakage. They design five prompting methods that incorporate varying data modalities: historical returns (Vanilla), news summaries, news sentiment scores, financial fundamentals, and a combination of fundamentals with sentiment. Ratings are evaluated by comparing them to forward return quintiles over 1, 3, 6, 12, and 18-month horizons, using Mean Absolute Error (MAE) as the metric.

The results show that all LLM-based methods outperform traditional analyst ratings in terms of average MAE, with the Fundamentals and Fundamentals + Sentiment methods achieving the best performance (MAE ~1.42 vs. analyst MAE of 1.57). Interestingly, adding news data improves short-term (1-month) predictions but harms longer-term accuracy, suggesting that news can introduce noise or bias. Sentiment scores provide a token-efficient alternative to full news summaries without significant performance loss. The study also highlights that analysts exhibit a strong bias toward buy ratings, while LLM-generated ratings are more evenly distributed.

The paper provides a reproducible framework for leveraging LLMs in investment analysis, demonstrating that even without fine-tuning, LLMs can generate more accurate and less biased stock ratings than human analysts. Future work is suggested to extend the analysis to longer horizons, incorporate additional data sources, and test newer models.