Research Projects
Unified Data Protection and Security Provenance in Digital Infrastructure for Fintech Services
Principle Investigator: Associate Professor Liang Zhenkai
Summary: This subproject aims to unify data protection and security provenance monitoring in the fintech infrastructure, where various systems collaborate for transactions. Current security solutions are platform-specific, leading to transaction discrepancies and integration challenges. Our goal is to establish consistent data protection across the stack, including identity authentication, access policy enforcement, data isolation, and privacy in processing. We propose a novel solution addressing challenges such as defining a consistent protection model, developing unified provenance models, and leveraging security mechanisms across platforms.
Real-time Fraud Management based on Graph Analytics and Learning
Principle Investigator: Professor He Bingsheng
Summary: The project focuses on addressing critical challenges in real-time fraud detection within future digital finance applications. Graph analytics can be used to detect money laundering and various types of fraud by examining large volumes of transactions, offering clear insights to investigators. Due to the high velocity of transactions and the expansion of large-scale graphs in pervasive digital finance environments, there is a growing need for more sophisticated analytics tools. This project aims to enhance fraud detection capabilities by researching and developing advanced graph analytics, graph neural networks, and network embedding methods. These methods will leverage emerging parallel hardware, such as graphics processing units (GPUs), to enable real-time processing. The project also plans to explore emerging graph learning and embedding technologies to improve the effectiveness of fraud detection. The project also aims to explore the explainability of the approaches used for fraud detection.
Extracting and Scoring the Reasoning and Prediction of Financial Experts’ comments from Textual Information
Principle Investigator: Associate Professor Huang Ke-wei
Summary: Successful financial decisions require useful information, but investors often face challenges with too much or misleading data. The project's research objective is to create a system along with new algorithms capable of automatically summarizing the financial forecasts of experts, as well as the explanations and reasoning behind these forecasts from a vast amount of textual data. Once the information is cleaned and extracted, the aim is to develop a scoring system that quantifies the quality of both the forecasts and the causal reasoning behind the predictions made by financial experts. This research is valuable for stakeholders in the financial markets, including regulators like MAS, professional investors, and retail investors. It can help verify analysts' recommendations, monitor public firms' forecasts, and protect investors from misleading information.
Credit Recovery Efficiency Analytics
Principle Investigator: Associate Professor Ying Chen
Summary: This upstream research project, "Smart Credit Analytics," aims to develop a new methodology for modeling recovery rates for various debt obligations using modern big-data analytics. Current industry practices often rely on simplified assumptions or discretionary judgment, such as fixed or uniformly distributed recovery rates. This project addresses the industry's recognized deficiencies by focusing on two key challenges: the scarcity of recovery rate observations and the need for non-standard distributional modeling for bimodal recovery rates. The research involves extending modern variable selection techniques to develop a scientifically rooted and robust recovery rate prediction system.
Supply Chain Finance in Agriculture
Principle Investigator: Associate Professor Johan Sulaeman
Summary: The project focuses on addressing barriers in providing digital financing services to smallholder agribusinesses. Firstly, it aims to identify and document these barriers among all relevant actors. Secondly, it seeks to create a repository of scientifically valid surveys to assess the readiness of different stakeholders to engage in digital finance. Thirdly, it aims to conduct extensive data collection across Southeast Asian countries to gauge the readiness of various actors for digital finance. Lastly, to publish and maintain a benchmarking platform online, providing access to survey materials, raw data, and insights for researchers and financial institutions to utilize in their own research or business endeavors.
Portfolio Credit Analytics
Principle Investigator: Associate Professor Huang Ke-Wei
Summary: Current fintech lending primarily focuses on leveraging IT technologies to set up new platforms and distribution channels. Data analytics is traditionally applied in sorting out good from bad credits but credit rating/scoring alone is definitely insufficient to meet the expectations of modern credit analysis in this digital era. The shortcomings are evident in at least two aspects – (1) dynamic credit portfolio perspective of institutional lending, and (2) heterogeneous recovery rates across obligors and/or instrument types. This project addresses the former through developing new big-data deep credit analytics constructed for dynamic credit environments and designed for consistent aggregation of individual credit exposures into a portfolio perspective that can evolve in a point-in-time manner for various application horizons of interest.
Secure and Privacy-preserving Blockchain-agnostic Interoperability with Capital Markets Protocol
Principle Investigator: Professor Ooi Beng Chin
Summary: This project focuses on building an enterprise-grade generic and blockchain-agnostic interoperability protocol in the form of a cross-chain bridge with security and privacy protection as key features. This bridge will implement InterOpera’s [InterOpera] proprietary Capital Market Protocol (CMP) targeting the use case of asset transfers between entities in the capital market and carbon credit market that belong to different blockchains. The project argues that the security and privacy of this protocol and its implementation need to be very high while ensuring compliance with regulations. To achieve privacy, the project study and develop state-of-the-art techniques based on Zero Knowledge Proof (ZKP) [Goldwasser19]. To enhance the security, we plan to develop techniques for program analysis and fuzzing, and machine learning for monitoring illicit activities. While being blockchain agnostic, this project will be evaluated with a few key blockchains, including public blockchains such as Ethereum and Chia Network, and enterprise blockchains such as Hyperledger Fabric and Cosmos. These blockchains are widely used in the capital market and the emerging carbon credit market.
Proven Privacy Preserving Synthetic Data Generation with Generative Adversarial Networks and Differential Privacy
Principle Investigator: Associate Professor Biplab Sikdar
Summary: This project will develop highly representative tabular synthetic data from real data with different kinds of Generative Adversarial Networks (GANs) and Differential Privacy (DP) that is as-good-as real and has no privacy risks, thereby, making it compliant with data protection regulations worldwide. Unlike the one-to-one feature mapping that anonymization is based on, GANs use one-to-infinity feature mapping as they can learn the structure of a dataset. A classic example of this is DeepFakes1 that can be used to generate fake faces of people. However, GANs have only been investigated for unstructured data (image, video, and audio) for the past three years, whereas, this project will explore and study GANs for structured data (rows and columns) and text in unstructured data. Thus, the objective of this project is research and development of the following:
a. Different GAN models to generate synthetic data with different data types (categorical, numerical, and Boolean) for sequential as well as time series data.
b. Advanced features such as ability to model complex datasets that require a combination of relational joins between multiple tables, field constraints and time-series data, and text feature generation prior to generating synthetic data for free-text fields.
c. Data utility and data privacy metrics to create a benchmarking tool to evaluate the utility, fairness and risk of a dataset.
This project will strictly focus on applied research of synthetic data with GANs and privacy-preserving technology to ensure its research outcomes have a direct commercial impact for Financial Institutions (FIs) and the extended research community.
Quantum speed-up in FinTech Algorithms and Optimization
Principle Investigator: Associate Professor Ying Chen
Summary: This project aims to explore the potential of quantum algorithms and optimization in accelerating fintech applications, particularly in portfolio optimization and derivative pricing. The financial industry has experienced significant transformations with the emergence of fintech and the adoption of artificial intelligence and machine learning techniques. However, quantum computing presents new opportunities for more efficient processing.
In recent years, various quantum machine learning algorithms have been proposed, theoretically promising speed-ups over classical counterparts. However, many of these algorithms either require quantum access to data, raising questions about their applicability, or are heuristic in nature with no proven advantage over classical algorithms. Moreover, implementations have been limited to small-scale problems and basic settings where analytical solutions already exist, negating the need for advanced algorithms in practice.
Unveiling the Implications: Investigating the Impact of Digital Private Information Exposure on Fintech Lending and Borrower Behaviors
Principle Investigator: Professor Sumit Agarwal
Summary: The emergence of online intermediaries and Fintech companies has revolutionized lending through technological innovation. The ascent of Fintech has brought forth renewed prospects for enhancing efficiency, but it has also raised fresh concerns about shifting systemic risk to less-regulated and more-fragile financial intermediaries. To gain a comprehensive understanding of the impact of Fintech lending on the financial market, our research intends to investigate the impact of digital private information exposure on Fintech lending and borrower behaviors by analyzing a customer private data regulation shock in Aug 2019.
In August 2019, the Google Developer Policy Center implemented a new privacy regulation, stipulating that apps without default SMS, phone, or assistant handler capability could no longer collect call log and SMS information of customers. Consequently, our researching Fintech lending firm, CASHe has been unable to access this personal information since the regulation came into effect. After the shock, CASHe started to collect borrower information by asking for five contacts and their relationship to call in case of default.
Enhancing Financial Investment Decision-making with Retrieval-augmented Large Language Models
Principle Investigator: Chair Professor Chua Tat Seng
Summary: Existing large language models are mostly trained for the general domain and tend to struggle to correctly understand the various types of financial data due to the unique terminology and domain knowledge in the financial area. The application of large language models in vertical domains like finance is still underexplored. The project proposes the development of a retrieval-augmented financial large language model (FLLM) to improve decision-making in finance. The project aims to adopt a novel approach to develop domain-specific large language models (LLMs) that can effectively integrate and interpret diverse data from multiple channels timely and precisely by combining LLM and information retrieval (IR) models. Additionally, the reliability and accuracy of the FLLM on solving complex financial tasks will be enhanced by teaching it to perform automatic problem decomposition and leverage reliable external tools to derive more trustable results.
Predicting Debt Crisis with Artificial Intelligence and Real-Time Big Data
Principle Investigator: Assistant Professor Zheng Huanhuan
Summary: The project represents a new approach to improve the prediction of sovereign debt crises by addressing the limitations of traditional early warning systems. The project aims to enhance early warning systems for predicting debt crises using AI and real-time big data and, in the process, empower policymakers to proactively prepare for and navigate macroeconomic crises. Leveraging artificial intelligence (AI) and real-time big data extracted through text analysis of global news, commentary, and trending searches, the project promises substantial advancements over existing methods. In particular, it will pursue the following objectives:
- Constructing real-time data feeds to nowcast and forecast economic activities.
- Incorporating international spillover effects to improve crisis prediction.
- Innovating new methodologies to identify early warning signals.