Title: The effect of trimming on algorithms for combining groups of pre-classified observations
Authors: Andrea Cerasa - Joint Research Centre (Italy) [presenting]
Andrea Cerioli - University of Parma (Italy)
Abstract: Three algorithms for merging homogeneous groups of pre-classified observations have been recently proposed. Their application on international trade data allows a synthetic representation of the market and a clear identification of anomalous commercial behaviors. To assess the performance of the algorithms we conducted Monte Carlo experiments. At that stage, methods for mitigating the effect of the possible presence of outliers had been left for forthcoming development of the procedure. Since anomalous declarations are quite usual in international trade data, non-robust procedures may result distorted and unfair trade strategies may be masked and remain undetected. We extend the simulation experiments by introducing three different data contamination structures, in order to quantify the effects of the presence of anomalous observations on operationally relevant results. Moreover, we propose to pre-filter the data using a robust regression approach based on LMS and the Forward Search. The simulation results on filtered data help us measuring the effect of trimming on algorithms accuracy and choosing the best cleaning strategy for international trade data. An application to real data concludes the study.