Methodological contributions to off-policy evaluation under the sample weighting framework and empirical financial analysis using semiparametric machine learning
Date
2025-08
Authors
Li, Yuyang
Major Professor
Advisor
Yu, Cindy
Kim, Jae-Kwang
Fuller, Wayne
Nordman, Daniel
Hennessy, David
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This dissertation investigates two interrelated research themes situated at the intersection of statistical learning and decision-making under uncertainty. The first theme emphasizes the application of machine learning methodologies to financial statistics, specifically targeting the robust modeling of conditional asset price distributions. Recognizing the limitations of traditional statistical models in capturing the skewness, heavy tails, and rare-event sensitivity inherent in financial markets, I propose a deep quantile regression-based forecasting approach. This methodology significantly improves tail-region predictions and enables the empirical analysis of option-implied risk premia through comparative analysis between predicted physical and risk-neutral densities. Additionally, I develop a novel financial flexibility index by integrating
structured econometric models with qualitative textual analysis of corporate disclosures using advanced large language models.
The second theme advances methodological frameworks in off-policy evaluation for reinforcement learning. Addressing challenges posed by data derived from historical policies, I refine the debiased calibration weighting approach to produce estimators exhibiting semiparametric efficiency and double robustness. This methodology is further generalized to sequential decision-making scenarios, providing multiply robust estimators applicable to dynamic policy evaluations.
Series Number
Journal Issue
Is Version Of
Versions
Series
Academic or Administrative Unit
Type
dissertation