(Deep) Learning Analyst Memory [Link to Paper] 

This paper links 290,000 equity analyst reports to I/B/E/S forecasts and introduces a new contextual recall measure, showing that transformer models are equivalent to standard psychological models of contextual retrieval, enabling a model that explains 43% of recall behavior, far exceeding traditional belief formation models.

• Awards: Second-Year Paper Award (Liew Fama-Miller Fellowship)

• Presentations: Machine Learning in Economics Summer Conference 2025 (MLESC25) 

Shock-Restricted Markov Switching Structural VARs [Link to Paper] 

This paper introduces a novel identification procedure using a Markov-switching VAR with external shock restrictions to examine the bidirectional relationship between uncertainty and economic activity, finding that uncertainty shocks have significant long-run adverse effects, while macroeconomic uncertainty can also stimulate industrial production, supporting the growth options theory.

• Awards: CRSP Summer Paper Award

Machine Learning Attention: Understanding Selective Information Processing with Embeddings

I model investor-specific attention to publicly available information using a transformer-based architecture. Although investors have access to the same disclosures, their interpretations vary, leading to differences in forecast accuracy and investment decisions. I represent investor information processing as sequences of text embeddings and estimate attention weights that capture how individuals selectively focus on different aspects of the available information. These attention profiles predict forecast errors and reveal systematic heterogeneity in the use of public information. This framework offers a new approach to studying attention-driven anomalies in asset prices.

• Awards: Third-Year Paper Award (Yiran Fan Memorial Fellowship)