Abstract:

For years, fraudsters have been using debit and credit card numbers that they have printed them on a blank plastic card so that they can use them in brick and mortar stores. For the past five years, banks had introduced Europay MasterCard and Visa (EMV) chip card technology, which made it possible for merchants to twitch demanding a PIN (here reference as personal identity number) for each transaction made. As frequency of transactions is increasing, number of fraudulent transactions is rapidly increasing. Hence the objective of this study to determine the effectiveness of logistic model tree (LMT) and gradient boosting decision tree (GBDT) to detect the amount of fraudulent transactions in South Africa (SA). A daily data for the period of 01 Mach 2020 to 10 July 2020, was obtained by, web-scrapping credit and debit card fraud from SA. The LMT and GBDT were constructed using training data on the 17 financial conditioning factors identified by the LogitBoost algorithm. We applied these models and validated them using receiver operating characteristics (ROC), and predictive accuracy (ACC). Overall, the two models exhibit reasonably a good performance; the LMT exhibits the highest predictive capability as compared to GBDT model. Our results indicated a success rate of 99.95% and a prediction rate of 99.9% when using the logistic model tree. This is a promising technique for online fraud susceptibility detection. The results of this study are useful for decision makers and SA financial sector for future use and planning in credit and debit prone areas.