Quantifying Equity Risk Premia: Financial Economic Theory and High-Dimensional Statistical Methods

Abstract

The overarching question of this dissertation is how to quantify the unobservable risk premium of a stock when its return distribution varies over time.

The first chapter, titled “Theory-based versus machine learning-implied stock risk premia”, starts with a comparison of two competing strands of the literature. The approach advocated by Martin and Wagner (2019) relies on financial economic theory to derive a closed-form approximation of conditional risk premia using information embedded in the prices of European options. The other approach, exemplified by the study of Gu et al. (2020), draws on the flexibility of machine learning methods and vast amounts of historical data to determine the unknown functional form. The goal of this study is to determine which of the two approaches produces more accurate measurements of stock risk premia. In addition, we present a novel hybrid approach that employs machine learning to overcome the approximation errors induced by the theory-based approach. We find that our hybrid approach is competitive especially at longer investment horizons.

The second chapter, titled “The uncertainty principle in asset pricing”, introduces a representation of the conditional capital asset pricing model (CAPM) in which the betas and the equity premium are jointly characterized by the information embedded in option prices. A unique feature of our model is that its implied components represent valid measurements of their physical counterparts without the need for any further risk adjustment. Moreover, because the model’s time-varying parameters are directly observable, the model can be tested without any of the complications that typically arise from statistical estimation. One of the main empirical findings is that the well-known flat relationship between average predicted and realized excess returns of beta-sorted portfolios can be explained by the uncertainty governing market excess returns.

In the third chapter, titled “Multi-task learning in cross-sectional regressions”, we challenge the way in which cross-sectional regressions are used to test factor models with time-varying loadings. More specifically, we extend the procedure by Fama and MacBeth (1973) by systematically selecting stock characteristics using a combination of l1- and l2-regularization, known as the multi-task Lasso, and addressing the bias that is induced by selection via repeated sample splitting. In the empirical part of this chapter, we apply our testing procedure to the option-implied CAPM from chapter two, and find that, while variants of the momentum effect lead to a rejection of the model, the implied beta is by far the most important predictor of cross-sectional return variation.