Fraud detection in bank transactions using statistical modelling

Fraud detection in bank transactions using statistical modelling

Fraud detection in bank transactions

2026-09-26

Lisa Li

Fraudulent banking transactions create substantial financial and operational risks for financial institutions and their customers. The increasing availability of transaction-level data provides an opportunity to identify potentially fraudulent activity using statistical learning methods. This study aims to develop and compare statistical models for predicting whether a banking transaction is fraudulent. A labelled dataset containing approximately 200,000 transaction records is used, with fraud status represented by the binary response variable Fraud. After the removal of direct personal identifiers and other unsuitable fields, variables describing transaction characteristics, customer information, account balances, merchant categories, locations and transaction devices are considered as potential predictors. The labelled data are divided into a 70% training set and a 30% test set using stratified sampling. Two multiple logistic regression models are developed using different sets or representations of the predictor variables. For each model, backward stepwise selection based on the Akaike Information Criterion is used to reduce the full set of candidate predictors. Five-fold cross-validation is conducted within the training data to examine the predictive performance of the modelling procedures. The two final models are then fitted using the complete training set and applied to the independent test set. The test predictions are compared with the actual fraud outcomes using confusion matrices, and the predictive accuracy of each model is calculated. Model 1 achieved a test accuracy of 96.91%, while Model 2 achieved a slightly higher accuracy of 96.93%. Therefore, Model 2 was selected as the final model based on its higher test accuracy. The findings demonstrate how logistic regression and backward variable selection can be used to support the identification of potentially fraudulent banking transactions. However, the results should be interpreted considering limitations relating to data quality and the possible use of simulated transaction data.

fraud detection, banking transactions, logistic regression, statistical modelling, R