Hierarchical modelling for financial data

Thesis

Algorithmic trading has gained great popularity and has been heavily used by institutional investors. Such a trading system involves mathematical and statistical modelling that makes decisions to buy or sell financial securities by following automated trading instructions. In this thesis, we research machine learning based data-driven approaches to construct such trading rules.

We utilise deep learning models to predict stock price movements and develop trading strategies based on the derived signals. Instead of using hand-crafted features, we propose neural network architectures to extract nonlinear features from limit order books data. We expand this work to study Bayesian Deep Learning models and Quantile Regression, obtaining uncertainty estimates on predictive outputs. These uncertainty estimates can be used for position sizing to improve trading performance.

However, a mapping from predictive signals to trade positions is non-trivial to construct. To solve this problem, we research Reinforcement Learning algorithms to optimise a measure of expected utility of final wealth, directly outputting trade positions thus bypassing the explicit forecasting step. Reinforcement Learning algorithms are used to trade individual assets and, subsequently, we extend this end-to-end training framework to portfolio optimisation and directly maximise the Sharpe ratio for portfolios, maximising return per risk.

Authors: Zhang, Z

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Limit order books; Deep learning; Portfolio optimisation; Microstructure data; Reinforcement learning
• Subjects: Machine learning; Finance