Scoring de fraude en transacciones monetarias con Machine Learning

Abstract

Abstract: Fraud in monetary transactions represents millions in losses for financial institutions and undermines user confidence in digital systems. This work develops a fraud scoring model using Machine Learning algorithms to detect fraudulent transactions in real-time. Four algorithms were implemented and compared: Random Forest, XGBoost, and Neural Networks, using a dataset of 590,540 transactions with extreme class imbalance. The SMOTE technique proved most effective for handling this imbalance. XGBoost emerged as the superior model, significantly outperforming traditional rule-based methods. Interpretability was implemented through SHAP enabling explanation of model decisions a fundamental aspect in regulated financial contexts. The proposed system generates risk scores with differentiated thresholds for automatic approval, additional verification or blocking. Results demonstrate that it is feasible to implement artificial intelligence-based fraud detection systems that balance effectiveness.