Paper: arXiv 2410.22519

Authors: Hugo E. Caceres, Ben Moews

Abstract

Financial regulators such as central banks collect vast amounts of data, but access to the resulting fine-grained banking microdata is severely restricted by banking secrecy laws. Recent developments have resulted in mechanisms that generate faithful synthetic data, but current evaluation frameworks lack a focus on the specific challenges of banking institutions and microdata. We develop a framework that considers the utility and privacy requirements of regulators, and apply this to financial usage indices, term deposit yield curves, and credit card transition matrices. Using the Central Bank of Paraguay’s data, we provide the first implementation of synthetic banking microdata using a central bank’s collected information, with the resulting synthetic datasets for all three domain applications being publicly available and featuring information not yet released in statistical disclosure. We find that applications less susceptible to post-processing information loss, which are based on frequency tables, are particularly suited for this approach, and that marginal-based inference mechanisms to outperform generative adversarial network models for these applications. Our results demonstrate that synthetic data generation is a promising privacy-enhancing technology for financial regulators seeking to complement their statistical disclosure, while highlighting the crucial role of evaluating such endeavors in terms of utility and privacy requirements.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.5: no out-of-sample test; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 8.5)

Why this score: The paper applies established synthetic data generation methods (marginal-based inference vs. GANs) to real-world banking microdata from the Central Bank of Paraguay, featuring public dataset release and specific domain applications like yield curves and transition matrices, indicating strong empirical implementation, while the math remains accessible without heavy theoretical derivations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Assess utility of synthetic banking microdata<br>for regulatory applications"] --> B["Methodology Framework"]
  
  B --> C["Data Input<br>Central Bank of Paraguay<br>Banking Microdata"]
  
  C --> D["Computational Processes<br>Marginal-based Inference vs GAN Models<br>Utility & Privacy Evaluation"]
  
  D --> E["Application Scenarios<br>Financial Usage Indices<br>Term Deposit Yield Curves<br>Credit Card Transition Matrices"]
  
  E --> F["Key Findings/Outcomes"]
  
  F --> G["Frequency-based applications<br>show highest utility"]
  F --> H["Marginal inference<br>outperforms GAN models"]
  F --> I["PETs complement<br>statistical disclosure<br>with public datasets available"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a utility‑and‑privacy framework for synthetic banking microdata and provides the first synthetic datasets for financial usage indices, term deposit yield curves, and credit card transition matrices.

Source sentence
We develop a framework that considers the utility and privacy requirements of regulators, and apply this to financial usage indices, term deposit yield curves, and credit card transition matrices.

Assumes 5

  • Inputs to the marginal‑based inference mechanisms must be categorical. data
    Source sentence
    As they use discrete marginals, inputs are required to be categorical.
  • Differential privacy noise is added according to the specified DP parameters (σ, ε, δ). statistical
    Source sentence
    The mechanism then calculates the amount of noise that should be added to the sensitive dataset to satisfy the criteria, using the differential privacy parameters as σ = …
  • Numeric features are discretised via k‑means clustering before synthetic generation. model
    Source sentence
    In the data-driven PS, we discretise numeric features through k-means transformations, as credit management actions such as selling non‑performing loans result in non‑normal and multi‑modal distributions.
  • Synthetic information products require fewer resources than maintaining integrated data portals. other
    Source sentence
    resources required to implement synthetic information products can be lower than maintaining integrated data portals or processing individual information requests.
Show 1 more
  • Pre‑processing strategies (CBP vs data‑driven) affect utility and error propagation. model
    Source sentence
    We also assess two separate pre‑processing strategies, one following the CBP’s current data definitions and one data‑driven.

Data

  • Central Bank of Paraguay banking microdata (synthetic) · Central Bank of Paraguay synthetic public
    approximately 3.7 million individuals, 13,000 term deposits, 700,000 credit cards
    Source sentence
    Our financial inclusion, term yield curve, and credit risk domain applications cover approximately 3.7 million unique individuals, 13,000 term deposits, and 700,000 credit cards, respectively.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Applications based on frequency tables are especially suitable for synthetic data, and marginal‑based inference mechanisms outperform GAN models for these tabular tasks. performance
    Source sentence
    We find that applications less susceptible to post‑processing information loss, which are based on frequency tables, are particularly suited for this approach, and that marginal‑based inference mechanisms to outperform generative adversarial network models for these applications.
  • Synthetic microdata yields relatively high error on weighted average interest and total capital (1%‑10% error and large absolute deviations). statistical
    Source sentence
    Here, all mechanisms average relatively high error with regard to the original WAI and TC domain (1% - 10% and 47 billion PYG, respectively).
  • Utility metrics should be evaluated at the intended granularity because errors are unevenly distributed across granularity levels. theoretical
    Source sentence
    We find that utility metrics for domain applications should be evaluated at their intended granularity, as errors may be distributed unevenly between granularity levels.
  • Synthetic data can provide new granular information products, such as credit‑card loss data not currently disclosed. other
    Source sentence
    For example, the loss data for credit cards is information not currently featured and valuable to financial institutions, researchers, and practitioners.
  • Regulators with limited experience and resources may encounter additional challenges when adopting synthetic data generation. other
    Source sentence
    Our data extraction, validation, and processing indicate that regulators with less experience and resources are likely to face additional challenges.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (68)

  1. Democracy and productivity: The Glass-Steagall Act and the shifting discourse of financial regulation (2012)
  2. Central bank transparency and independence: Updates and new measures (2014) — link
  3. Market discipline and banking system transparency: Do we need more information? (2012)
  4. Tragedy of the data commons (2011) — link
  5. Non-performing loans: What matters in addition to the economic cycle? (2013) — link
  6. Tracking the COVID-19 crisis with high-resolution transaction data (2021) — link
  7. Central bankers as supervisors: Do crises matter? (2018) — link
  8. The boundary between privacy and utility in data publishing (2007)
and 60 more
  1. Statistical disclosure control for microdata (2017) — link
  2. Broken promises of privacy: Responding to the surprising failure of anonymization (2010) — link
  3. Statistical disclosure limitation: New directions and challenges (2018) — link
  4. Privacy and synthetic datasets (2019) — link
  5. INEXDA – The granular data network (2019) — link
  6. \textit k -anonymity: A model for protecting privacy (2002) — link
  7. Practical data-oriented microaggregation for statistical disclosure control (2002) — link
  8. The algorithmic foundations of differential privacy (2014) — link
  9. Calibrating noise to sensitivity in private data analysis (2006) — link
  10. Confidentiality protection in the 2020 US Census of population and housing (2023) — link
  11. ABS perturbation methodology through the lens of differential privacy (2019)
  12. A critical review on the use (and misuse) of differential privacy in machine learning (2022) — link
  13. The limits of differential privacy (and its misuse in data release and machine learning) (2021) — link
  14. Differential privacy for government agencies – are we there yet? (2023) — link
  15. Statistical disclosure limitation (1993) — link
  16. Generating synthetic data in finance: Opportunities, challenges and pitfalls (2021) — link
  17. Stochastic gradient descent with differentially private updates (2013) — link
  18. Modeling Tabular data using Conditional GAN (2019) — link
  19. PATE-GAN: Generating synthetic data with differential privacy guarantees (2018) — link
  20. Winning the NIST Contest: A scalable and general approach to differentially private synthetic data (2021) — link
  21. Benchmarking differentially private synthetic data generation algorithms (2021) — link
  22. Comparative Study of Differentially Private Synthetic Data Algorithms from the NIST PSCR Differential Privacy Synthetic Data Challenge (2019) — link
  23. Differentially Private n-gram Extraction (2021) — link
  24. AIM: An adaptive and iterative mechanism for differentially private synthetic data (2022) — link
  25. Generative adversarial nets (2014) — link
  26. Deep learning with differential privacy (2016) — link
  27. Differentially private GANs by adding noise to discriminator’s loss (2021) — link
  28. Semi-supervised knowledge transfer for deep learning from private training data (2017) — link
  29. Does financial inclusion reduce poverty and income inequality in developing countries? A panel data analysis (2020) — link
  30. The little data book on financial inclusion (2022) — link
  31. Measuring financial inclusion: a multidimensional index (2018) — link
  32. The deposit base – multibanking and bank stability (2019) — link
  33. The lending-deposit rate relationship in eastern european countries: Evidence from the rank test for non-linear cointegration (2010) — link
  34. The U.S. Treasury yield curve: 1961 to the present (2007) — link
  35. Revolving versus convenience use of credit cards: Evidence from U.S. Credit Bureau data (2020) — link
  36. Private numbers in public policy: Census, differential privacy, and redistricting (2022) — link
  37. Skewed distributions in finance and actuarial science: A review (2015) — link
  38. LOWESS: A program for smoothing scatterplots by robust locally weighted regression (1981) — link
  39. Convex total least squares (2014) — link
  40. Determinants of interest rates on time deposits and savings accounts: Macro factors, bank risk, and account features (2018) — link
  41. Mergers and acquisitions of financial institutions: A review of the post-2000 literature (2009) — link
  42. Intensity models and transition probabilities for credit card loan delinquencies (2014) — link
  43. Recent advances in optimization techniques for statistical tabular data protection (2012) — link
  44. An explainable federated learning and blockchain-based secure credit modeling method (2024) — link
  45. Comparison of non-linear optimization algorithms for yield curve estimation (2009) — link
  46. Modelling credit card exposure at default using vine copula quantile regression (2023) — link
  47. Models and forecasts of credit card balance (2016) — link
  48. A new mixture model for the estimation of credit card exposure at default (2016) — link
  49. An empirical comparison of conventional techniques, neural networks and the three stage hybrid Adaptive Neuro Fuzzy Inference System (ANFIS) model for credit scoring analysis: The case of Turkish credit card data (2012) — link
  50. Bank-sourced credit transition matrices: Estimation and characteristics (2021) — link
  51. Data envelopment analysis, truncated regression and double-bootstrap for panel data with application to Chinese banking (2018) — link
  52. Decomposing banking performance into economic and credit risk efficiencies (2019) — link
  53. A class of categorization methods for credit scoring models (2022) — link
  54. Optimizing credit limit adjustments under adversarial goals using reinforcement learning — → 2306.15585
  55. Dynamic survival models with varying coefficients for credit risks (2019) — link
  56. Parsimonious modeling of yield curves (1987) — link
  57. Estimating and interpreting forward interest rates: Sweden 1992–2994 (1994) — link
  58. Anonymization: The imperfect science of using data while preserving privacy (2024) — link
  59. Differential privacy for statistics: What we know and what we want to learn (2010) — link
  60. Synthetic Data – What, why and how? (2022) — link

1 of 68 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.