Empirical Asset Pricing with Large Language Model Agents

Problem: Traditional asset pricing methods rely on manually designed economic indicators or factors whose predictive power tends to diminish over time under the Efficient Market Hypothesis. While qualitative discretionary investing leverages rich linguistic data from news and reports to extract crucial pricing information, quantitative factor models struggle to integrate these insights due to the need for financial reasoning, long-term memory, and complex interactions. Furthermore, existing natural language processing methods in finance typically target trading strategies for individual stocks rather than explaining macroscopic excess asset returns across the entire market.

Method: The proposed hybrid asset pricing framework utilizes an LLM agent to iteratively analyze business news, supported by a continuously updated macro and micro note and an external vector database memory initialized with the SocioDojo knowledge base. The daily embeddings of the generated analysis reports are smoothed over a sliding window of k days using an exponential decay kernel w(j) = e^{-\beta j} to form a daily state s_{t}. This qualitative state is concatenated with manual financial economic factors, downsampled, merged with asset embeddings v_p, and passed to a multi-layer neural network to predict the next-day excess return r_{p, t+1}.

Results: The proposed methodology outperforms traditional machine learning-based baselines, demonstrating a 10.6 enhancement in the Sharpe ratio for portfolio optimization and a 10.0 improvement in the mean pricing error magnitude for anomaly portfolios. Comprehensive ablation studies confirm that components like iterative refinement, long-term notes, and pre-training the hybrid pricing network with historical data all contribute positively to performance. Furthermore, upgrading the base model to the reasoning-enhanced O1-Preview model yields further improvements, raising the Sharpe ratio gains up to 14.1 and reducing average \alpha pricing errors by 13.8 compared to the best baselines.

Limitations: The current model lacks internet access and a broader range of alternative information sources, which limits the depth of the generated discretionary analyses. Additionally, the framework relies on general-purpose foundation models rather than specialized financial LLMs such as FinGPT that could improve analytical reasoning. Finally, the proposed approach is currently unable to process multimodal financial information, such as the diagrams and figures frequently found in financial documents.