Dissertation
From return predictability to fund performance evaluation: a machine learning perspective
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2023
DOI: 10.25820/etd.007019
Abstract
This dissertation focuses on solving problems in the area of empirical asset pricing and investment management using machine learning methods. Previous literature has identified numerous factors and anomalies, but it remains unclear which ones are genuinely important in asset pricing. Additionally, evaluating fund performance has been challenging due to the dynamic nature of strategies and flexibility in risk exposures. The aim of this dissertation is to develop machine learning methods to offer new perspectives on these problems.
In the first chapter, I develop a customized machine learning model to identify important stock characteristics for mutual fund performance and stock return predictions. Existing studies have emphasized the dominant role of stock technical signals in return prediction. However, my model demonstrates that fundamental characteristics become more important for funds when accounting for the high correlations among these characteristics. The model successfully predicts both stock returns and fund performance, with the nonlinear model delivering superior performance. Using the model’s results, I further explore how active mutual funds utilize anomalies and their impact on fund performance. The results suggest that funds exploit certain anomalies (e.g., profitability) in the right direction but misapply others (e.g., trading frictions). While fund performance varies widely based on exposure to anomalies, the positive and negative exposures offset each other in aggregate fund performance. Additionally, I investigate the long-standing intrigue in the fund literature regarding the negative correlation between stock selection and market timing. I find that the negative relationship is not caused by fund pursuit of anomaly-based investment strategies.
The second chapter (co-authored with Ashish Tiwari) addresses the challenge of identifying factors that are genuinely related to the cross-sectional variation in expected stock returns. We employ a Bayesian ensemble-of-trees approach that combines the strengths of machine learning with a Bayesian inference framework. By applying this methodology to the U.S. stock market, the study reveals that a firm’s market value is consistently the most relevant characteristic for stock return predictions. Furthermore, we confirm that a stochastic discount factor based on a sparse set of factors successfully explains a significant portion of the expected returns variation.
In the third chapter (co-authored with Ashish Tiwari), we explore the machine learning-based approaches for benchmarking and evaluating hedge fund performance. For a substantial proportion of funds (referred to as zero-R2 funds), the explanatory power of the linear multi-factor benchmark models typically used for hedge fund performance evaluation is effectively zero. As a response, we examine alternative benchmark models based on various machine learning techniques to evaluate individual hedge fund performance. Compared to traditional linear benchmark models, machine learning methods significantly enhance the ability to track fund performance, particularly for zero-R2 funds. We find that a Bayesian ensemble-of-trees approach is particularly valuable in this context. The improved tracking performance enables more precise estimates of fund alphas, leading to more accurate identification of superior funds. Consequently, the proposed machine learning-based methodologies outperform traditional multi-factor models in various contexts, including real-time fund selection and fund failure prediction. These results provide compelling evidence that machine learning-based benchmark models effectively capture nonlinearities and interaction effects among risk factors, which are crucial for accurately characterizing the risks associated with hedge fund strategies. To further underscore the benefits of the machine learning methods, we reexamine the well-established positive relationship between strategy distinctiveness and hedge fund performance. The findings suggest that the observed relationship is a consequence of benchmark model error that contaminate the inference based on conventional factor models.
Details
- Title: Subtitle
- From return predictability to fund performance evaluation: a machine learning perspective
- Creators
- Tengjia Shu
- Contributors
- Ashish Tiwari (Advisor)Tong Yao (Advisor)Wei Li (Committee Member)Nick Street (Committee Member)Suyong Song (Committee Member)Qihang Lin (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Business Administration (Finance)
- Date degree season
- Autumn 2023
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007019
- Number of pages
- xiv, 233 pages
- Copyright
- Copyright 2023 Tengjia Shu
- Language
- English
- Date submitted
- 08/02/2023
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 214-233).
- Public Abstract (ETD)
This dissertation focuses on addressing several unresolved issues in empirical asset pricing including (a) identification of the factors that are genuinely related to the cross-sectional differences in asset returns, and (b) evaluating the performance of actively managed portfolios which employ dynamic portfolio strategies with considerable flexibility in their risk exposures. The aim of this dissertation is to develop machine learning methods to offer new perspectives on these problems.
In the first chapter, I develop a customized machine learning model to identify stock characteristics that are important for mutual fund performance and stock return predictions. Existing studies have emphasized the dominant role of technical signals in return prediction. However, my model demonstrates that fundamental characteristics become more important for funds when accounting for the high correlations among these characteristics. The model successfully predicts both stock returns and fund performance, with the nonlinear model delivering superior performance.
The second chapter (co-authored with Ashish Tiwari) focuses on determining the genuine importance and sparsity of factors in the equity market. The analysis relies on a flexible Bayesian ensemble-of-trees methodology that helps to model the complex interactions among the various factors. We confirm that a sparse set of factors-based stochastic discount factor effectively explains a significant portion of the variation in expected returns.
The third chapter (co-authored with Ashish Tiwari) highlights the value of machine learning methods for benchmarking and evaluating hedge fund performance. We show that machine learning methods outperform traditional multi-factor models in various contexts, including the identification of superior funds in real time, and in predicting fund failures.
- Academic Unit
- Tippie College of Business
- Record Identifier
- 9984546944102771
Metrics
3 File views/ downloads
24 Record Views