Essays on High-Frequency Factors

Abstract

This dissertation provides an empirical analysis of asset pricing factors in a continuous-time framework. The value of such a framework is that it affords a distinction between continuous returns, generally modelled by Brownian motion, and discontinuous returns, often referred to as "jumps.'' The focus of the dissertation is on how the asset pricing implications of these two types of risk differ. To measure these types of risk, it is crucial to use high-frequency return data to justify the use of infill asymptotic arguments that enable the identification of price jumps and volatility functionals. Correspondingly, in the first essay, I develop a dataset of high-frequency factor portfolio returns that allows for exactly this and, armed with this dataset, I further engage in a continuous-time econometric study of factor jump risk as well as factor continuous and (semi)jump risk premia. To begin, I study the jumps embedded in the factor portfolio returns, finding significant evidence thereof and suggesting that non-market systematic jump risk is non-trivial. These findings motivate a deeper analysis of said risk and, in particular, its pricing implications. To this end, I estimate the risk premia of each factor in my dataset, first estimating continuous/jump betas using high-frequency regressions and second estimating the risk premia using cross-sectional regressions. My results show that these two categories of risk draw different premia with jump risk being far more important for explaining cross-sectional return variation. Having illustrated the existence and importance of non-market jump risk, I then go on to study exactly what economic events drive said risk in the second essay of my dissertation, co-authored with Tim Bollerslev. In this latter essay, we construct an estimate of the tangency portfolio using the high-frequency factor returns previously defined; this portfolio serves as a univariate representation of all systematic risk and thus facilitates a straightforward analysis of both systematic market and non-market jumps. We then connect the jumps in this portfolio, again with identification leaning on infill asymptotics, with a large dataset of newswire articles to understand what news topics drive the jumps. We further price each of these topics, estimating their risk premia to understand the economic significance of each. We find that monetary policy and finance news is the most important followed by news about international affairs and macroeconomic data. We also decompose the jumps into more primitive economic shocks, finding that these systematic equity jumps primarily correspond to growth and short-rate shocks and that there exists substantial heterogeneity in what news topics drive the different primitive shocks. The third essay of this dissertation, co-authored with Tim Bollerslev and Mathias Siggaard, takes a different angle, instead trying to understand whether the market portfolio can be predicted using the lagged high-frequency factor returns. Our analysis in this essay similarly exploits the precision of jump detection afforded by the high-frequency data but uses the separated continuous/discontinuous returns to explore heterogeneity in the predictive content therein. Moreover, we consider using the continuous returns as regressands in our in-sample regressions, leaning on the well-established theoretical no-arbitrage argument that any predictability should lie solely in the continuous component of returns. Our forecasts, constructed using Lasso regressions to handle the high-dimensionality of our features, corroborate this idea, evincing market return predictability with strong statistical and economic significance. In addition, much of the predictive content of our forecasts can be traced back to a few factors related to turnover, connecting our results to past work on volume and information absorption.