Exploratory analysis and implementation of machine learning techniques for predictive assessment of fraud in banking systems

Abstract

Regarding combating financial fraud in banking systems, this study analyzes machine learning algorithms used in predictive assessments for fraud detection. For a given methodological approach, a database with more than six million records referring to the financial transactions of a given banking organization is used. First, an exploratory data analysis clarifies the main variables influencing the evaluation process, with binary and financial percentages related to fraud loss. Concerning the unbalance between the expected records and those classified as fraud, Random Under Sampling, SMOTE, and ADASYN techniques are used to balance and train these bases by Logistic Regression, Naive Bayes, KNN, and Perceptron techniques. With the implementation of machine learning algorithms, we present the main feasibilities of each model in each scenario. At the end of the study, we expose the final considerations and proposals for future work.