Comparative Analysis of Predictive Models Using Synthetic Data for Financial Fraud Detection

Abstract

This paper conducts a comparative analysis of predictive models in financial fraud detection using synthetic data. Financial fraud is an adversarial problem that is becoming increasingly common with technological advancement; fraudsters use techniques such as social engineering to lure victims into releasing confidential information for nefarious purposes. One of the types of fraud this research aims to tackle is detecting fraudulent transactions processed on customers' accounts as customers are usually responsible for bearing the costs associated with fraud. The study examines machine learning techniques applied to mitigate fraud in financial transactions under controlled five scenarios which include unusual transaction type, account behaviour profiling, and temporal correlation of transactions, time-based velocity checks and 1 hour window analysis. A synthetic dataset comprising 2,549,085 transactions categorized into fraud and non-fraud classes (849,695 fraudulent transactions and 1,699,390 non-fraudulent transactions) was utilized for analysis. Transaction-level features were employed to evaluate the performance of various machine learning classifiers and Deep Learning Algorithms. The classifiers examined include SVM, Decision Trees, Logistic Regression, Gradient Boosting, LSTM, Deep Neural Network, CNN, and Random Forest. Among these, the Random Forest Classifier demonstrates superior performance, achieving an accuracy rate of 99.18%. The evaluation primarily focuses on comparing the performance of classifiers using the confusion matrix as the primary evaluation metric. The findings indicate that the Random Forest Classifier, Decision Trees and Gradient Boosting are suitable classifiers for detecting financial fraud.