Essays on empirical asset pricing

Abstract

This thesis examines issues in factor investing, return predictability, and applications of machine learning to empirical asset pricing. The first chapter provides an empirical evaluation of potential explanations for the value premium by extending the decomposition methodology of Hou and Loh (2016). I find that cross-sectional differences in financial leverage are able to explain up to 60 percent of the value premium in individual stocks and book-to-market and industry sorted portfolios. I do not find support for behavioural explanations, or rational explanations based on adjustment costs, human capital, or operating leverage. I show that most of the explanatory power of financial leverage is concentrated in the short-term, and a factor based on financial leverage is able to partially explain the alpha of the Fama-French HML factor. The second chapter provides a comprehensive analysis on the use of supervised machine learning methods for multi-asset class return predictability. I extend the existing literature in three dimensions: time, asset class, and geographical coverage. Using the longest available time series of asset returns and macroeconomic fundamentals across multiple countries I show that most machine learning models fail to beat the historical mean benchmark out-of-sample, often by a large margin. However, combining forecasts across models dramatically improves performance, and leads to sizeable gains for mean-variance investors in trading strategies that aim to exploit return predictability. Additionally, I show that forecasting performance is better in bad times and when market volatility is high. The third chapter documents a new phenomenon I call 'cross-factor momentum.' I show that standard time series momentum strategies applied to long-short equity factors can be enhanced by incorporating predictability from other factors. Exploiting the robust cross-correlation structure between factors can improve Sharpe ratios by up to 1.7 times compared to standard strategies, with lower drawdowns and similar trading costs. This finding holds in international data, cannot be explained by intermediary asset pricing models, and is only partially explained by fund flows.