This thesis addresses practical, real-world problems in the financial services industry using Deep Learning architectures. The main focus is on advancing current approaches in the areas of Risk Management and Quantitative Trading. The former is concerned with the risk of investment, whereas the latter refers to identifying profitable investment opportunities. Both areas are crucial in assessing the profitability and risk levels of investments. This research has three important findings which contribute to the academic literature.
First, the impact of news headlines on forecasting short-term volatility is explored. A novel neural network architecture, the Multimodal Hierarchical Attention Network (MHAN) is introduced to learn the joint representation of multiple data modalities (prices and news). The architecture addresses two aspects which are essential for news representation: news relevance and news novelty. The importance of the components of MHAN is investigated using the ablation method, with different sentence encoders being used to represent text. The experimental results confirm that adding news headlines consistently improves volatility forecasting across different market sectors and that the MHAN model outperforms the widely used GARCH model.
Second, a novel graph neural network architecture, the Graph Transformer Network (GTN) is proposed to better deal with the Range trading strategy which profits from short-term distortions in stock prices. The relationship between stocks is represented by a graph. The effect of other stocks on the target stock range prediction is then investigated by injecting prior knowledge via a graph into the learning process. The proposed approach is evaluated on a large number of stocks. Experiments confirm the profitability of the strategy and demonstrate that the GTN outperforms current state-of-the-art graph networks.
Third, Deep Reinforcement Learning is successfully applied to the pair trading strategy. This thesis presents the Investment Strategy with the Investors' Preferences (ISIP) framework to integrate optimal portfolio allocations with a risk management component. This component targets a constant level of risk pre-set by the investor. The experimental results confirm that the proposed framework improves the performance of existing approaches and is effective at restricting portfolio volatility.
