Modelling Multimodal Financial Data with Inductive Logic Programming and Ontologies

This thesis describes an attempt to implement automated, machine learning-based algorithms for processing financial reports with the aim of extracting relevant, actionable information. The purpose of the research is to apply hybrid machine learning to the financial domain using both structured and unstructured data.

The ontology of Security Exchange Commission (SEC) financial reports provides a suitable model for the representation of structural data, such as the company management hierarchy, the main shareholders and the relationships between individuals in these roles and the companies in which they play a role. Using SEC financial reports, I built an ontology by extracting and transforming the filings into a structured graph of companies, people and their relations.

Using the ontology that I built, I show how a hybrid learner can be used to find strategies and rules using Inductive Logic Programming (ILP) and traditional machine learning algorithms. This research shows that it is possible to increase the probability of selecting cointegrated pairs of stocks compared to random selection. This has the potential to enhance the decision-making process for choosing the right pairs of stocks while saving on computation time and resources.

To address scalability when learning from large ontologies, I developed P-CONNER, a concurrent Description Logic learner that accelerates the learning process using data-parallel algorithms on both the CPU and GPU. I conducted experiments on benchmark tests and demonstrated that P-CONNER can learn from large ontologies much faster than traditional learners.

Overall, the techniques I introduced in this thesis help automate the construction of a financial ontology from company reports. Combining symbolic and sub-symbolic machine learning for a finance pair trading strategy demonstrates statistically significant improvement by utilising an ontology structure with interpretable rules. This is facilitated by the hardware-independent concurrent Description Logic learner, which enables scalable learning from ontologies.