Development of a Scoring Model to Counter Money Laundering Based on Transactional Data

The main purpose of this work is to check the effectiveness of using a scoring model based mainly only on transactional data to automatically identify customers who use their own Bank accounts for the purpose of legalizing funds obtained by criminal means. One of the fundamental questions in this study was the level of accuracy of dividing customers into "bad" and "good", which can be achieved by using aggregated information about customer transactions. Thus, the task was formalized as a binary classification problem, where one class corresponded to a group of law-abiding clients, and the other - a group of clients who committed economic crimes. Ranking systems, or scoring systems, are widely used in the banking sector to assess the credit risks of clients at the applicational stage. However, in this study, we investigated the effectiveness of using scoring systems to solve a completely new problem: firstly, we assessed the risk of involving the client in money laundering schemes instead of credit risks; secondly, we used data describing the client's behavior in the Bank, including its transactional activity, instead of standard applicational data. During the work, the main issues related to the laundering of illegally obtained funds were considered: the global nature of the problem; the main stages of process of legalization of income, such as placement, layering, and integration; impact of money laundering on the economy and national security of countries; existing methods of combating this economic crime. At the next stage, the following theoretical aspects of modeling were analyzed and studied in detail: the basic principles of building and implementing models in the banking sector; the interpretability of models; as well as methods that can be used to assess the predictive ability of trained models. At the practical stage of the research, the programs in the R language were developed to provide execution of algorithms for generating, converting, and selecting variables necessary for model development. Microsoft Office Excel was also used to create tables with data and graphs which illustrate such characteristics of variables as risk stability and flow stability. Due to the clearness of visual demonstration of these properties, it was concluded that it is impossible to include some variables in the final list of variables suitable for modeling. As a result, a mathematical model (linear logistic regression) was developed on a final set of variables, predicting the probability of engagement in money laundering for every client. The quality of the results obtained from the model was evaluated using the AUC ROC metric. An application was also developed that to estimate the expected effect of implementing the model in the bank. Based on the tests conducted, it was concluded that it is advisable to implement the developed solution, since it will significantly reduce the share of manual labor used to detect cases of money laundering. This work is presented on 43 A4 sheets, and also includes 5 illustrations and 4 tables. Information from 32 different sources was used in the preparation of the work.