Machine learning in portfolio management

Abstract

Financial markets are difficult learning environments. The data generation process is time-varying, returns exhibit heavy tails and signal-to-noise ratio tends to be low. These contribute to the challenge of applying sophisticated, high capacity learning models in financial markets. Driven by recent advances of deep learning in other fields, we focus on applying deep learning in a portfolio management context. This thesis contains three distinct but related contributions to literature. First, we consider the problem of neural network training in a time-varying context. This results in a neural network that can adapt to a data generation process that changes over time. Second, we consider the problem of learning in noisy environments. We propose to regularise the neural network using a supervised autoencoder and show that this improves the generalisation performance of the neural network. Third, we consider the problem of quantifying forecast uncertainty in time-series with volatility clustering. We propose a unified framework for the quantification of forecast uncertainty that results in uncertainty estimates that closely match actual realised forecast errors in cryptocurrencies and U.S. stocks.