Date of Award
Winter 12-15-2021
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Financial researchers, who often work with large volumes of financial data, need efficient tools to estimate economic relationships and to make predictions. Unfortunately, traditional linear models with conditioning variables that are parsimoniously selected rarely deliver useful information for estimating or predicting financial objects of interest. When the dimension of conditioning variables is very large, the traditional financial models and the statistical methods commonly used in finance fail to reflect the complex relationship behind the enormous datasets and tend to break down in the out-of-sample prediction. However, machine learning can provide an effective tool to handle high dimensional data and apply the right algorithms to dramatically reduce the computational costs.
In this thesis, I investigate several questions in the field of machine learning in finance. In the first chapter, I consider the problem of improving predictability of asset pricing models conditional on high dimensional observable information. In doing so, I introduce a new pseudo-Siamese Network for Asset Pricing (SNAP) model, based on deep learning approaches, for conditional asset pricing. The pseudo-SNAP model outperforms all benchmark models in terms of out-of-sample forecasting and out-of-sample Sharpe ratio. In addition, this model is robust under economic restrictions and has a number of properties in portfolio management.
In the second chapter, co-authored with Manish Jha and Asaf Manela, we consider using natural language processing methods to measure popular sentiment toward finance from millions of books published in eight countries over hundreds of years. We show that our text-based finance sentiment index can be used to predict financial crises and to capture the impulse response of GDP and credit growth.
In the third chapter, once again co-authored with Manish Jha and Asaf Manela, we use a text-based measure of popular sentiment toward finance to study how finance sentiment responds to rare historical disasters and to the ongoing Covid-19 pandemic. We show the heterogeneous effects of different natural disasters on people's views of finance and suggest their sentiments respond differently to the realization of insured versus uninsured risks. In particular, we also investigate the impact of the Covid-19 pandemic on finance sentiment and suggest that finance sentiment seems to depend on the insurance provided by private markets and by public finance.
Language
English (en)
Chair and Committee
Werner Ploberger
Committee Members
Ian Fillmore
Recommended Citation
Liu, Hongyi, "Essays on Machine Learning in Finance" (2021). Arts & Sciences Electronic Theses and Dissertations. 2573.
https://openscholarship.wustl.edu/art_sci_etds/2573