Dynamic Knowledge Graph Asset Pricing

Abstract

The macro-finance system stands as a complex network. We model this system in a Large-Language-Model-based (LLM) dynamic knowledge graph using the entire corpus of the Wall Street Journal from 1999 to 2023. We identify a constellation of systematic themes, such as economic growth. A novel network-based measure named polar centrality is introduced to quantify the systematic importance of corresponding theme factors to the whole economy. We apply an asset pricing model to validate our structural theme factors as reliable proxies for state variables in the unconditional ICAPM framework. Our theme indexes inherently account for the time-varying importance of different macro themes. They demonstrate forecasting competency for future economic activities, unconditional pricing power in the cross-section, and incremental contribution to the stock market factor zoo. The LLM-powered graph AI system unveils a new thematic factor lens that represents the financial system with high resolution.

Keywords: Knowledge graph, Large language model, Cross-sectional asset-pricing, ICAPM, Factor model, Graph theory, Generative AI, Factor zoo

JEL Classification: C38, C52, E00, G12