MemAPM: Memory-augmented Large Language Model Agent for Asset Pricing

Anonymous

Abstract: In this study, we propose a hybrid asset pricing model, MemAPM, which utilizes a Large Language Model (LLM) agent to refine information from news and augment it with a memory of past refined news. We perform experiments on a two-year span of news and around 70 years of market data, our method outperforms the state-of-the-art machine learning-based asset pricing baselines in multiple portfolio optimization and asset pricing error evaluations. We also performed an ablation study and evaluated the predictive power of the augmented news features for the price movement of individual stocks and economic indicators. The results show the effectiveness of our proposed memory augmentation technique and hybrid asset pricing network architecture.

Paper Type: long

Research Area: Computational Social Science and Cultural Analytics

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English