Multivariate functional isolation forest for money laundering detection

Money laundering is a complex sequence of money transfers that traditional methods based on aggregated transactions struggle to detect. To improve upon this issue, the detection of such anomalies is framed as identifying unusual (co-) movements across dimensions in time and scale, with interactions between pairs of dimensions quantified using several projection-based methods. Then, this information is embedded in an isolation forest algorithm extended to multivariate functional data by using these interaction scores as splitting criteria, generating an anomaly score. Finally, this anomaly score ranks observations by their deviation from a functional baseline and allows the identification of suspicious transactions. Empirically, this method is applied to a large sample of bank transaction records, where customers' behaviors are represented as multivariate transaction curves across cash, wire, and international credit/debit dimensions, highlighting money laundering cases overlooked by conventional methods. Depth-based global and local feature importance metrics are used to identify the most informative dimensions. Evaluation against functional and tabular baselines confirms the effectiveness of the proposed methodology. In addition, the performance of the methodology is showcased in a series of realistic simulations.