This paper explores the integration of quantitative factors with newsflow representations derived from large language models (LLMs) for stock return prediction. The authors propose a hybrid framework that combines traditional financial indicators (e.g., value, momentum, quality) with embeddings from LLMs fine-tuned on financial news. The key innovation is the use of fine-tuned LLMs to capture nuanced market sentiment and event-driven signals that are not fully reflected in quantitative factors. The model is evaluated on the North American stock universe, with portfolio backtesting comparing standalone factor models, standalone news models, and their combination.

Results demonstrate that fine-tuning LLMs on domain-specific news significantly enhances prediction accuracy compared to using static pre-trained embeddings. The combined model outperforms both individual components, achieving higher Sharpe ratios, lower volatility, and better risk-adjusted returns. The synergy is particularly strong for small-cap and high-volatility stocks, where news sentiment plays a larger role. The paper provides empirical evidence that LLM representations can complement traditional quantitative factors, offering a more comprehensive view of market dynamics.

The study's significance lies in bridging the gap between quantitative finance and natural language processing, showing that LLMs can be effectively adapted for financial prediction tasks. The findings suggest that investors can improve portfolio performance by incorporating news-driven signals alongside traditional factors, especially in environments where market sentiment is rapidly evolving. The work also highlights the importance of fine-tuning LLMs on financial data to capture domain-specific language and context.