Financial Risk Management with Machine Learning

This thesis consists of three applications of machine learning techniques to risk management. The first chapter proposes a deep learning approach to estimate physical forward default intensities of companies. Default probabilities are computed using artificial neural networks to estimate the intensities of the inhomogeneous Poisson processes governing default process. The major contribution to previous literature is to allow the estimation of non-linear forward intensities by using neural networks instead of classical maximum likelihood estimation. The model specification allows an easy replication of previous literature using linear assumption and shows the improvement that can be achieved. The second chapter, titled Causal Networks with Neural Networks is a co-authored work with Damir Filipovic (SFI & EPFL), Negar Kiyavash (EPFL) and Jalal Etesami (EPFL). We develop a data-driven framework to identify the interconnections between firms using an information-theoretic measure. This measure generalizes Granger causality and is capable of detecting nonlinear relationships within a network. Moreover, we develop an algorithm using recurrent neural networks and Granger causality to identify the interconnections of high-dimensional nonlinear systems. The outcome of this algorithm is the causal graph encoding the interconnections among the firms. These causal graphs can be used as preliminary feature selection for another predictive model or for systemic risk management. We evaluate the performance of our algorithm using both synthetic linear and nonlinear experiments and apply it to the daily stock returns of US listed firms and infer their interconnections from 1990 to 2020. The third chapter, titled StockTwits Classified Sentiment and Stock Returns is a co-authored work with Damir Filipovic (SFI & EPFL). We classify the sentiment of a large sample of StockTwits messages as bullish, bearish or neutral, and create a stock-aggregate daily sentiment polarity measure. Polarity is positively associated with contemporaneous stock returns. On average, polarity is not able to predict next-day stock returns. But when we focus on specific events, defined as sudden peaks of message volume, polarity has predictive power on abnormal returns. Polarity-sorted portfolios illustrate the economic relevance of our sentiment measure.