Detection of abnormal behavior in trade data using Wavelets, Kalman Filter and Forward Search

In this paper we address the issue of the automatic detection of abnormal behavior in time series extracted from international trade data. We motivate, review and use three specific methods, based on solid frameworks: Wavelets, Kalman Filter and Forward Search. These methods have been successfully applied to an important EU policy issue: the analysis of trade data for antifraud and antimoney-laundering, fields in which specialists are often confronted with massive datasets. Our contribution consists in an in-depth study of these approaches to assess their performance, qualitatively and quantitatively. On the one hand, we present these three approaches, underline their specific aspects and detail the used algorithms. On the other hand, we put forward a rigorous assessment methodology. We use this methodology to evaluate each method and also to compare them, on simulated time series and also on real datasets. Results show each method has its specific advantages. Their joint use could be of a high operational impact for our applications, to deal with the variety of patterns occurring in trade data