CardSim：支付卡欺诈检测研究的贝叶斯模拟器（英）2025

Investment Rating - The report does not provide a specific investment rating for the industry. Core Insights - Payment card fraud has significantly increased in recent years, with 11.5% of credit card owners and 9.4% of debit card owners experiencing fraud in 2023, more than double the rates before the COVID-19 pandemic [7] - The report introduces CardSim, a Bayesian simulator designed to generate synthetic payment card transaction data for fraud detection research, addressing the lack of publicly available transaction data [4][12] - The simulator is modular and can be easily updated to reflect new payment trends and fraud patterns, facilitating the testing of machine learning models for fraud detection [4][12] Summary by Sections Introduction - Payment card fraud is a major concern, with significant growth observed since the COVID-19 pandemic, as evidenced by increased fraud reports and survey data [7] - The development of generative AI tools poses both risks and opportunities for fraud detection and prevention [7] Related Work - The lack of publicly available payment transaction data hampers research in fraud detection, leading to the development of simulation methodologies to fill this gap [10][11] - Existing simulators have limitations, and there is a need for more sophisticated methods that reflect the dynamics of fraud in payment systems [11][19] Simulation Methodology - The simulator generates synthetic transaction data and fraud patterns for consumer-to-business non-prepaid debit and credit card payments [23] - It operates in three phases: developing payer and payee characteristics, running the transaction simulator, and generating fraud labels using Bayes' theorem [23][44] Results - An example dataset produced by the simulator included 3.16 million records, with a fraud ranking threshold set at one percent, which is higher than many industry estimates [61] - The distributions of card and location types across payment classes were consistent with marginal probabilities, while fraudulent transactions skewed towards features with higher conditional probabilities [62][63]