Abstract

This dissertation studies biases on both sides of modern financial markets---those of human investors and those of the machine-learning tools increasingly used to study them. The first chapter identifies a textual property of news---narrative unpredictability---through which investors can overcome the bias-prone, heuristic (System~1) default of news processing, finding that the cognitive strain it induces can pull investors into slower, more deliberate (System~2) analysis and thereby attenuate market underreaction. The second chapter, which provides the methodological infrastructure that makes such a study credible, develops chronologically consistent large language models that are free of lookahead bias---a measurement-side bias whose mitigation supports more credible backtests and predictions across finance and other social-science domains. The third chapter turns to a different human bias---racial bias in capital allocation by mutual fund managers---documenting a co-racial tilt in portfolio choices consistent with inaccurate statistical discrimination, together with an asymmetric investor response in which minority-dominant funds are penalized similarly to White-dominant funds for poor performance but not rewarded as much for superior performance. Taken together, the three chapters trace bias in financial markets across three distinct layers: the bias-prone System~1 default through which investors process news, the measurement-side distortions of the language models used to study them, and the race-related tilts shaping how professional capital allocators direct capital across firms.

Committee Chair

Asaf Manela

Committee Members

Andreas Neuhierl; Brett Green; Guofu Zhou; Xing Huang

Degree

Doctor of Philosophy (PhD)

Author's Department

Finance

Author's School

Olin Business School

Document Type

Dissertation

Date of Award

5-5-2026

Language

English (en)

Download

Available for download on Thursday, May 04, 2028

Included in

Finance Commons

Share

COinS