2025 Theses Doctoral

Essays in Financial Economics and Machine Learning

Buchta, Matthias Nikolaus

This dissertation comprises three essays in financial economics and machine learning. It leverages expansive datasets and modern empirical techniques to analyze the dynamics and origins of macroeconomic announcement risk premia, time variation in high-frequency return predictability, and stochastic components of factor risk prices.

In Chapter 1, I examine the dynamics and economic drivers of a persistent market phenomenon: stock returns are disproportionately high on days of scheduled macroeconomic announcements. Leveraging recent methodological advances and the availability of index options with frequent expiration dates, I estimate the conditional distributions of daily market returns, which enable dynamic analyses of daily risk premia and their origins in return space. I demonstrate that within my sample from 2016 to 2023, announcement risk premia were relatively muted in the initial years but began to spike sharply following pandemic-related surges in inflation, regularly exposing short-horizon investors to severe tail risks. Drawing on macroeconomic news data and interest rate forecasts, I propose a mechanism whereby heightened inflation triggers increased uncertainty about future monetary policy paths; this uncertainty, in turn, amplifies market sensitivity to various macroeconomic releases and ultimately leads to elevated announcement risk premia. Extrapolating this mechanism beyond my sample, my estimations suggest that monetary policy uncertainty has accounted for a substantial excess in announcement-day returns between 1983 and 2024.

In Chapter 2, I explore the dynamic properties of high-frequency stock return predictability across machine learning models of varying complexity. The results show that complex non-linear machine learning models based on gradient boosted trees or neural networks substantially and consistently outperform simpler linear models like ridge or partial least squares. Among non-linear models, both the general level of predictability and the predictive benefits of increased model complexity vary systematically across time and are strongly associated with a small set of aggregate market variables. In particular, retail trade volume emerges as a key driver of increased predictability, while return volatility is strongly associated with diminishing benefits to model complexity. Moreover, I demonstrate that gaps of as little as one day between estimation and prediction samples lead to significant losses in predictive accuracy, illustrating the substantial structural dynamics in high-frequency financial markets and evidencing the need for frequent model re-estimation as commonly practiced in academic applications. Finally, analysis of predictor importance using Shapley values reveals that buy-sell trade imbalances and bid-ask spreads are consistently the most potent high-frequency return predictors within my feature set.

In Chapter 3, I develop a strategy for the identification and robust estimation of continuous, predictably discontinuous (overnight), and unpredictably discontinuous (jump) factor risk premia from the cross section of stock returns. Recovering the continuous and respective discontinuous factor spaces via principal component analysis, I obtain the corresponding risk premia of observable factors through a sequence of regressions allowing for latent factors of different stochastic types. Empirically, I employ my novel risk premium estimator to various potentially priced equity risk factors at intraday frequency in a sample spanning from 2004 to 2022. My analysis corroborates that the market risk premium can primarily be attributed to jumps and further extends this finding to a large number of other well-known risk factors.