Abstract: The suspicious transaction reporting system is the principle mechanism to fight against money laundering, and it is a technical problem to detect suspicious transaction for financial institutions and the financial intelligence unit. To help anti-money laundering analysts screen customers' unusual transactions and behaviors in massive financial transaction information, a new method, composition of predictive error and statistic treatment(CPEST) is presented, which can be used to detect unusual behaviors from the inconsistency of customer behaviors. CPEST models a customer's behavior, tests a customer's behavior at a particular time using estimated errors, and uses a window test to improve the ability to identify suspected of money laundering. Applying the method based on support vector regression and kernel density estimation to real data examples and simulations, the experiment results suggest that the method,which is feasible and effective, has high value in popularization and application.

Key words: anti-money laundering, anomaly detection, time series, support vector regression, kernel density estimation