Assessing the Efficacy of Machine Learning Analytics in Detecting Financial Frauds to Reduce Overfittings of Traditional Rule-Based Systems

With the Indian financial systems facing a relentless onslaught of data breaches and a sharp rise in financial fraud cases, the demand for effective solutions has reached a critical juncture. Machine learning (ML) has emerged as a proven tool to combat fraud. The current study comprehensively compares select ML-based analytical methods against traditional rule-based systems. The study used a pre-labeled dataset, considering online payments transactional data from an IT-based firm. A synthetic and labeled dataset of digital transactions (based on aggregated metrics and intentional malicious entries) was generated from Kaggle (a simulator PaySim). Transaction type, monetary value of each transaction, customer identification number, recipient’s existing and post-transaction account balance, time step for each transaction, and fraudulent transaction binary indicator are considered variables to build the ML model. The comprehensive data analysis, which includes the effectiveness of each ML analytics, their accuracy, feature reduction, and statistical analysis to understand the dataset’s characteristics and class imbalance and identify critical predictors of financial fraud, instills confidence in the study’s findings. Our work presents the accuracy of such advanced analytics. It addresses the nuances of fraudulent patterns (predictive frauds), wherein it observed that the random forest-based ML approach is the best-performing algorithm that can increase an appropriate level of accuracy and model fit in detecting fraudulent transactions to a scalable fraud detection system. The study shows scalability, deployment, data security, and legal compliance implications and suggests future research in this area for real-time monitoring systems to design reliable fraud detection systems.