ASSESSMENT OF THE PROBABILITY OF FRAUD IN THE PROCESS OF LENDING TO THE BANK'S CUSTOMERS

Abstract

The article is devoted to the current topic of assessing the likelihood of credit fraud in banks. This issue is related to the growth of economic processes digitalization and the transfer of payment transactions to the digital space. Its solution is carried out in eight scientific areas, confirmed by the construction and analysis of a map of scientometric bibliography of research on the problem of fraud in lending to bank customers. The article highlights clusters of scientific papers related to processes of protection of online transactions, machine, ensemble and incremental training to solve the problems of credit fraud, probabilistic approaches, techniques of detecting anomalies in operations related to money laundering in banks, the process of finding fraud in the financial sector, risk assessments, Data Mining. The data set from 122 variables and 307,511 records of the bank's customers were used to conduct a study to assess the likelihood of credit fraud in banks. The construction of the conceptual model made it possible to outline the stages of modelling, which was carried out using the modern Python programming language. The data was cleared of missing information and checked for compliance with the normal distribution law. As a result of the obtained data set, three models were built - logistic regression, decision tree and neural network. It turned out that the share of correct predictions in the training sample for logistic regression was 93.09%, for the decision tree and neural network - 100.00%, and in the test sample, respectively - 93.60%, 99.15%, 86, 67%. It indicates the adequacy of the data of both pieces and the high accuracy of forecasting. The constructed models were also tested for accuracy and quality. As a result, it turned out that all models are pretty accurate and high quality, but the decision tree is the most accurate, high quality and adequate model. Built-in models are universal tools for detecting fraudulent transactions, but they require constant monitoring and updating of information in connection with the emergence of new signs of criminal activity in the process of lending to customers.