PhD Confirmation Seminar - Hanqing Tian

Topic: When LLMs Read the News: Interpretability, Mispricing, and Asset Prices

Abstract:

This document presents a research agenda on large language models, interpretability, and empirical asset pricing. The first paper, Didisheim et al. [2026], shows that financial news is partly predictable from numerical stock characteristics. After removing this predictable component, the residual “pure news” reveals a monthly return-predictability anomaly with an annualised Sharpe ratio of 3.1, approximately double the next-best anomaly in the Jensen et al. [2022] factor universe. The second paper, Chen et al. [2025], addresses the black-box critique of LLM-based research. Using Sparse Autoencoders (SAEs), we map an LLM’s internal features to economic topics and show that concept-level steering can correct incoherent reasoning, reduce systematic biases, and simulate agents with configurable preferences. A third, preliminary solo-authored project extends the pure-news framework to cross-firm spillovers, showing that the pure news of a firm’s peers predicts its own future returns. Together, these projects show that LLMs can do more than predict asset prices: they can help measure, interpret, and explain how financial information is incorporated into markets.

This seminar will be conducted online via Zoom