Essays in empirical asset pricing with machine learning

This thesis consists of four papers on topics in empirical asset pricing with a particular focus on applications of machine learning. The first paper investigates the interplay of predictable trading behaviour and asset prices. We show that predictable order ow is associated with improved liquidity and market efficiency. In addition, we find evidence for a priced factor constructed from order ow predictability, contributing to the literature that connects market microstructure features and asset prices. The second paper evaluates the efficacy of machine learning based forecasts of bond excess returns and contributes to a better understanding of the formation of bond risk premia. We show that machine learning techniques outperform the principal components benchmarks used in extant literature and deliver substantial economic gains to investors. The third paper investigates the risk-reward trade-off in index options through the lens of a factor modelling approach. We show that a factor model with five factors and time-varying loadings instrumented with option characteristics, explains the vast majority of variation in delta-hedged option returns. The recovered factors point to jump, volatility and term structure spread risks. Finally, the fourth paper studies the systematic drivers of asset holdings in a novel factor modelling approach. I document the existence of a factor structure in holdings changes that points to distinct, well-understood economic channels as drivers of asset holdings. Using investor-specific factor loadings I find evidence for pro-cyclical trading of banks and mutual funds as well as counter-cyclical trading of investment advisors and pension funds. Furthermore, I document that changes to institutional investor holdings driven by systematic factors are negatively associated with future returns, suggesting a price pressure channel as a driver for return reversals.