Essays on Predicting and Explaining the Cross Section of Stock Returns

Abstract

My dissertation consists of three chapters that study various aspects of stock return predictability. In the first chapter, I explore the interplay between the aggregation of information about stock returns and p-hacking. P-hacking refers to the practice of trying out various variables and model specifications until the result appears to be statistically significant, that is, the p-value of the test statistic is below a particular threshold. The standard information aggregation techniques exacerbate p-hacking by increasing the probability of the type I error. I propose an aggregation technique, which is a simple modification of 3PRF/PLS, that has an opposite property: the predictability tests applied to the combined predictor become more conservative in the presence of p-hacking. I quantify the advantages of my approach relative to the standard information aggregation techniques by using simulations. As an illustration, I apply the modified 3PRF/PLS to three sets of return predictors proposed in the literature and find that the forecasting ability of combined predictors in two cases cannot be explained by p-hacking. In the second chapter, I explore whether the stochastic discount factors (SDFs) of five characteristic-based asset pricing models can be explained by a large set of macroeconomic shocks. Characteristic-based factor models are linear models whose risk factors are returns on trading strategies based on firm characteristics. Such models are very popular in finance because of their superior ability to explain the cross-section of expected stock returns, but they are also criticized for their lack of interpretability. Each characteristic-based factor model is uniquely characterized by its SDF. To approximate the SDFs by a comprehensive set of 131 macroeconomic shocks without overfitting, I employ the elastic net regression, which is a machine learning technique. I find that the best combination of macroeconomic shocks can explain only a relatively small part of the variation in the SDFs, and the whole set of macroeconomic shocks approximates the SDFs not better than only few shocks. My findings suggest that behavioral factors and sentiment are important determinants of asset prices. The third chapter investigates whether investors efficiently aggregate analysts' earnings forecasts and whether combinations of the forecasts can predict announcement returns. The traditional consensus forecast of earnings used by academics and practitioners is the simple average of all analysts' earnings forecasts (Naive Consensus). However, this measure ignores that there exists a cross-sectional variation in analysts' forecast accuracy and persistence in such accuracy. I propose a consensus that is an accuracy-weighted average of all analysts' earnings forecasts (Smart Consensus). I find that Smart Consensus is a more accurate predictor of firms' earnings per share (EPS) than Naive Consensus. If investors weight forecasts efficiently according to the analysts' forecast accuracy, the market reaction to earnings announcements should be positively related to the difference between firms' reported earnings and Smart Consensus (Smart Surprise) and should be unrelated to the difference between firms' reported earnings and Naive Consensus (Naive Surprise). However, I find that market reaction to earnings announcements is positively related to both measures. Thus, investors do not aggregate forecasts efficiently. In addition, I find that the market reaction to Smart Surprise is stronger in stocks with higher institutional ownership. A trading strategy based on Expectation Gap, which is the difference between Smart and Naive Consensuses, generates positive risk-adjusted returns in the three-day window around earnings announcements.