Machine Learning in Empirical Asset Pricing

This thesis consists of 5 chapters. Chapter 1 is the introduction of the thesis. Chapter 2, Chapter 3 and Chapter 4 are the main chapters of this thesis, which cover the U.S. bond market, international stock market and U.S. stock market respectively. Chapter 5 is the conclusion of the thesis. Chapter 1 presents the brief introduction to the 3 chapters covered in this thesis. Chapter 2 explores using technical indicators to forecast corporate bond returns with various machine learning (ML) models. We show that technical indicators yield statistically significant and economically meaningful results, consistently outperform ing bond characteristics. Although bond characteristics possess predictive power for bond returns, they do not provide incremental value beyond technical indicators across all bonds. Additionally, machine learning models do not offer substantial improve ments over the benchmark linear model. These results underscore the significance of technical indicators in the corporate bond market. Chapter 3 explores the use of various machine learning models to forecast stock returns in the international market. We calculate 107 technical variables and use them to train an array of machine learning models. Using data from 25 markets, we uncover machine learning’s superior performance over OLS (our benchmark model) in predicting future stock returns, in terms of both prediction accuracy and economic gains. We also show that machine learning generates superior performance among big, liquid firms and adds value during periods of downward market trend. The inter national heterogeinity in machine learning outperformance is associated with market features such as macroeconomic fundamentals, cultural traits, and the information environment. In Chapter 4, we apply genetic programming to systematically generate thousands of non-linear factor strategies and uncover various stock market anomalies that re main robust out-of-sample, even under multiple hypothesis testing. These non-linear strategies outperform raw anomalies by maintaining consistent profitability after controlling for arbitrage constraints (excluding microcaps), trading costs, recent data (post-2020), and risk adjustments; responding to Chen and Velikov (2023)’s critique of limited out-of-sample validity. Compared to Müller and Schmickler (2025)’s double sorted portfolio and single-sorted portfolio using pairwise interactions, our non-linear anomaly portfolio delivers superior performance while remaining transparent, inter pretable, and low-turnover. Moreover, we develop an interpretable non-linear four or five-factor model that consistently outperforms established benchmarks in both right-hand-side and left-hand-side asset pricing tests. Our non-linear factor models consist primarily of accounting-based anomalies, which enhances their viability net of trading costs, owing to the lower turnover and higher investment capacity associated with accounting variables. Chapter 5 summarizes the findings of the three chapters and concludes the thesis.