Essays on Empirical Asset Pricing

Abstract

This thesis investigates the relationship between risk and return in the cross-section of stocks. The thesis consists of three papers that can be read independently. These papers are the result of my Ph.D. studies at the Department of Finance, the Center of Financial Frictions (FRIC), and the Center of Big Data in Finance (BIGFI) at CBS. The first paper, Subjective Risk and Return, investigate the subjective risk-return tradeoff. To do so, I introduce a novel data set of subjective risk and return expectations at the individual stock level. I find that the required compensation for risk is high, but nevertheless, the realized compensation for risk is low. I show that this difference arises because cash flow expectations are systematically too high for risky stocks, which can be explained by investors suffering from optimism bias. As a result of the low realized compensation for risk, I show that risk cannot explain the realized return of most equity factors, and that the best asset pricing models for explaining realized returns are the worst ones for explaining subjective risk compensation. The second paper, Is There a Replication Crisis in Finance (co-authored with Bryan Kelly and Lasse Heje Pedersen), tests whether equity factor research in finance is robust to scientific replication. In particular, we build a global data set of stock returns and characteristics to replicate 153 equity factors in 93 different countries. Further, we develop and estimate a Bayesian model of factor replication, which can handle the multiple testing of many hypotheses, and the issue of publication bias. Our main result is that most equity factors can be replicated. Further, we show that most factors work well out-of-sample, that is, in time periods and countries different from the ones studied in the original paper. We also show that the 153 factors can be grouped into 13 themes, most of which matter for the tangency portfolio. The third paper, Machine Learning and the Implementable Efficient Frontier (co-authored with Bryan Kelly, Semyon Malamud, and Lasse Heje Pedersen), develops a framework that integrates trading-cost-aware portfolio optimization with machine learn-ing (ML). We show theoretically how to solve the optimal portfolio problem for an investor that faces trading costs when returns are predictable by a general function of security characteristics. In addition, we show to implement this solution via a machine learning methodology that learns directly about portfolio weight (rather than returns). Empirically, we find that our method leads to significant out-of-sample gains relative to various sophisticated benchmarks. Finally, our method gives a novel view of which security char-acteristics are economically important.