Estimating Financial Fraud through Transaction-Level Features and Machine Learning

In today's world, financial institutions (FIs) play a pivotal role in any country's economic growth and are vital for intermediation between the providers of investable funds, such as depositors, investors and users. FIs focus on developing effective policies for financial fraud risk mitigation however, timely prediction of financial fraud risk helps overcome it effectively and efficiently. Thus, herein, we propose a novel approach for predicting financial fraud using machine learning. We have used transaction-level features of 6,362,620 transactions from a synthetic dataset and have fed them to various machine-learning classifiers. The correlation of different features is also analysed. Furthermore, around 5000 more data samples were generated using a Conditional Generative Adversarial Network for Tabular Data (CTGAN). The evaluation of the proposed predictor showed higher accuracies which outperformed the previously existing machine-learning-based approaches. Among all 27 classifiers, XGBoost outperformed all other classifiers in terms of accuracy score with 0.999 accuracies, however, when evaluated through exhaustive repeated 10-fold cross-validation, the XGBoost still gave an average accuracy score of 0.998. The findings are particularly relevant to financial institutions and are important for regulators and policymakers who aim to develop new and effective policies for risk mitigation against financial fraud.