Abstract
In this thesis, I embark on an insightful journey into the realm of financial technology through the lens of deep learning, presenting novel methodologies to tackle intricate problems within this domain. Across its three pivotal essays, the thesis seamlessly weaves together diverse yet interconnected challenges, showcasing the versatility and depth of deep learning applications in FinTech.The first essay addresses the challenges of analyzing multivariate sequential data characterized by temporal irregularities. It proposes an end-to-end solution that accentuates the information content inherent in nonuniform time intervals and component misalignment, utilizing unique hidden states and a conditional flow representation. This approach non-parametrically captures the data distribution and its temporal dynamics, offering significant improvements over traditional probabilistic forecasting methods. The effectiveness of this model is demonstrated through comprehensive comparisons and real-world dataset applications, highlighting its ability to account for statistical heterogeneities and temporal dependencies.
The second essay introduces a novel framework for causal inference in distributional outcomes with continuous treatments. It establishes two pivotal causal quantities, the causal map and the causal effect map, to address the intricacies of distributional data. Through the introduction of three estimators—Distributional Direct Regression (Dist-DR), Distributional Inverse Propensity Weighting (Dist-IPW), and Distributional Double Machine Learning (Dist-DML)—and the development of a deep learning model combining the Neural Functional Regression Net (NFR Net) and the Conditional Normalizing Flow Net (CNF Net), this study provides a robust methodology for accurate causal inference. The practical application of these methodologies to E-commerce platform data underscores the nuanced consumer behavior in response to credit limit adjustments, revealing a preference for purchasing more expensive items over a proportional increase in overall spending.
The third essay presents NeuCredit, a cutting-edge deep learning model that utilizes Neural Ordinary Differential Equations and online shopping behavior data to significantly enhance consumer creditworthiness assessments. By capturing sporadic, time-varying features and their nonlinear relationships, NeuCredit outperforms conventional time series models, offering a more inclusive financial landscape for consumers with limited credit history. The model's innovative approach to assessing expected loss and its remarkable predictive accuracy, even in the absence of financing-related data, underscores the potential of integrating consumer behavior data with traditional credit management information. Additionally, the essay examines the fairness of NeuCredit's predictions and introduces an advanced adversarial training module to balance predictive accuracy with fairness considerations.
Together, these essays advance the field of financial technology by harnessing deep learning to solve complex problems, from causal inference and probabilistic forecasting to creditworthiness assessment. This thesis not only showcases the potential of deep learning applications in FinTech but also sets the groundwork for future research that bridges the gap between advanced computational techniques and practical financial applications.
| Date of Award | 3 Aug 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Qi WU (Supervisor) |