940. Domain applications of privacy-preserving mechanisms in synthetic data disclosure of financial regulators
Invited abstract in session WD-9: Quantitative methods in finance, stream OR in Finance and Insurance .
Wednesday, 14:30-16:00Room: Clarendon SR 2.01
Authors (first author is the speaker)
| 1. | Ben Moews
|
| Institute for Astronomy, University of Edinburgh |
Abstract
In today's information age, the vast amounts of data collected by financial regulators such as central banks stand in contrast to sensible legislative frameworks like banking secrecy laws. These constraints, however, limit the ability of the research community to make use of the available banking microdata in areas ranging from financial inclusion to credit risk. Central banks also have a mandate to promote stability and public trust, which external data access and validation enable. For these reasons, we adapt recent developments in privacy-preserving mechanisms, machine learning, and graphical models for synthetic data generation. The resulting models feature built-in privacy guarantees and are trained on the Central Bank of Paraguay's microdata as the first implementation of synthetic banking microdata of this kind. We apply these models to three different use cases; financial inclusion indices, term deposit yield curves, and credit card transition matrices. Our results show that marginal-based inference mechanisms outperform generative adversarial approaches, with models less susceptible to post-processing information less being particularly suitable to these domain applications. The findings of our work demonstrate the utility of synthetic data generation to enhance statistical data disclosure by financial regulators, while showcasing the importance of stringent evaluation to maintain performance and privacy.
Keywords
- Finance and Banking
- Machine Learning
- Risk Analysis and Management
Status: accepted
Back to the list of papers