Using machine learning for financial fraud detection in the accounts of companies investigated for money laundering

Benford’s Law is a well-known system use in accountancy for the analysis and detection of anomalies relating to money laundering and fraud. On that basis, and using real data from transactions undertaken by more than 600 companies from a particular sector, behavioral patterns can be analyzed using the latest machine learning procedures. The dataset is clearly unbalanced, for this reason we will apply cost matrix and SMOTE to different detecting patters methodologies: logistic regression, decision trees, neural networks and random forests. The objective of the cost matrix and SMOTE is to improve the forecasting capabilities of the models to easily identify those companies committing some kind of fraud. The results obtained show that the SMOTE algorithm gets better true positive results, outperforming the cost matrix implementation. However, the general accuracy of the model is very similar, so the amount of a false positive result will increase with SMOTE methodology. The aim is to detect the largest number of fraudulent companies, reducing, as far as possible, the number of false positives on companies operating correctly. The results obtained are quite revealing: Random forest gets better results with SMOTE transformation. It obtains 96.15% of true negative results and 94,98% of true positive results. Without any doubt, the listing ability of this methodology is very high. This study has been developed from the investigation of a real Spanish money laundering case in which this expert team have been collaborating. This study is the first step to use machine learning to detect financial crime in Spanish judicial process cases.