Money Laundering Transactions Chronology Analysis using Artificial Intelligence

Money laundering is one of the most significant threats for the financial systems of the world, as it promotes other crimes and destabilizes the economy. This thesis presents an enhanced analysis of AI and ML on financial transaction data for the identification of patterns that relate to money laundering. It uses techniques like the Isolation Forest for the identification of outliers and clustering techniques like K-means and DBSCAN for identification of unusual transaction patterns across the four high risk countries that include United States, United Kingdom, Japan and Italy. The preprocessed dataset included different types of transactions like cross-border transactions, deposit, credit card transactions and so on, handling with missing values, duplicate values and categorical values. Feature selection was done followed by normalization of the continuous variables; feature extraction which involves the use of the domain knowledge for the construction of the new features that were believed to enhance the performance of the model. Therefore, several models are proposed in the present study using Random Forest, GBM, and neural networks to classify the transactions and to predict the probability of illicit activity for better anti-money laundering. The best result was obtained with the Gradient Boosting Classifier where accuracy and precision were 73 % and 0.79 respectively, which guarantees a reasonable level of effectiveness of the system in identifying suspicious transactions while minimizing the number of false positives. Network graph analysis has been conducted to show the relationship between suspicious transactions and the clusters and the key accounts that are involved in money laundering. The findings are presented in the form of networks and heat maps to illustrate the structure of laundering networks and transactional relationships. The paper ends with the discussion on recommendations for enhancing the AML systems, feature engineering for the frequency of transactions and network-based features. The real-time fraud detection systems must also be developed. Finally, it would be useful to discuss some models within the deep learning framework, for sequential data, for example, Recurrent Neural Networks. Moreover, there is a need to find techniques that could be effective for altering the strategies of money laundering. The findings from this study have been useful in the development of AI based AML systems that are compliant with international policy and legal frameworks, and insights for financial institutions and policy makers while designing systems to combat financial crimes.