The application of machine learning in developing next-generation fintech solutions

In recent years, machine learning has played a significant role in facilitating the advancement of financial technology, enabling financial institutions to leverage the power of data and artificial intelligence to enhance their services. While machine learning is already implemented in various areas including loan underwriting, credit scoring, and algorithmic trading, other aspects of the financial domain have yet to fully embrace the potential of contemporary machine learning. Through collaboration with an industry partner, we investigate three distinct, yet interrelated problems that are currently underrepresented in the literature: lead generation for loan products, account code suggestion, and bank reconciliation. Our research provides insights into the shared challenges and nuances involved in developing customer-centric and business-level solutions, both of which are essential for sustaining a competitive advantage in the current financial services market. A fundamental component to managing a direct marketing campaign is identifying prospects and selection of leads. This can be challenging for financial products, particularly loans, as there are many factors to consider in modelling a customer's intention, such as risk, utility, and financial maturity. This thesis presents a bi-level approach modelling loan intention and eligibility to ensure that leads are both interested in and qualified for a loan. By using convex combination, we manage the dual influence of models and demonstrate that soft classifiers, like deep learning techniques, yield superior lead quality and prospect ranking. Notably, our proposed approach also offers valuable managerial insights, facilitating resource management decisions. Cloud-based bookkeeping platforms have opened up new possibilities for automated support in invoice processing tasks, such as account code suggestion and bank reconciliation. Precise account code suggestion is critical for accurate financial reporting, but maintaining consistency with a continually evolving list of account codes can be challenging for bookkeepers. This thesis proposes a hierarchical single-label classifier for account code suggestion instead of the traditional multi-class approach. By incorporating hierarchical information from induced taxonomies, we show significant improvements in classification and recommendation performance, with further enhancement through the use of our proposed post-processing strategy, Top-K Parent Boosting. Bank reconciliation is a laborious task for bookkeepers, requiring meticulous comparison and matching of financial records from two distinct datasets. Automating this process is vital for large businesses with high volumes of daily transactions. In this thesis, we propose an alternative methodology that employs representation learning and downstream link prediction to predict matches for bank reconciliation. Our proposed models surpass industry benchmarks and offer a more robust solution, improving the identification of group-based reconciliations; a feature often disregarded by traditional approaches. Additionally, we introduce a novel post-processing technique, Top Boundary Ranking, which enhances the detection of grouped reconciliations.