Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction
创建于 更新于
摘要
This paper studies multimodal stock return prediction by fusing structured quantitative factors with LLM-derived news representations, compares three fusion architectures (combination, summation, attention), proposes a two-component mixture model with a decoupled training scheme to stabilize learning, and shows via multi-universe backtests that (i) simple representation combination often yields the best fusion results, (ii) the decoupled mixture model improves robustness across universes and portfolio types, and (iii) LLM fine-tuning has inconsistent, universe-dependent effects [page::0][page::10]
