1 “Assessing classification methods for churn prediction by composite indicators” M. Clemente*, V. Giner-Bosch, S. San Matías Department of Applied Statistics, Operations Research and Quality, Universitat Politècnica de València, Camino de Vera s/n, 46022 Spain. *Corresponding author. Tel.: +34 96 387 74 90; Fax: +34 96 387 74 99. Email addresses: mclement@eio.upv.es (M. Clemente), ssanmat@eio.upv.es (S.San Matías), vigibos@eio.upv.es (V. Giner-Bosch) Keywords: Churn prediction, Composite indicators, Classification, Data mining. Abstract Customer churn prediction is one of the problems that most concern to businesses today. Predictive models can be developed for identifying future churners. As the number of suitable classification methods increases, it has become more difficult to assess which one is the most effective for our application and which parameters to use for its validation. To select the most appropriate method, other aspects apart from accuracy—which is the most common parameter— can and should be considered as for example: robustness, speed, interpretability and ease of use. In this paper we propose a methodology for evaluating statistical models for classification with the use of a composite indicator. This composite indicator measures multidimensional concepts which cannot be captured by a single parameter and help decision makers to solve this complex problem. Three alternatives are proposed giving different weights to the involved parameters considering the final user priorities. Our methodology finds which the best classifier is by maximizing the value of the composite indicator. We test our proposal on a set of five churn classification models drawn from a real experience, three of them being based on individual classifiers (logistic regression, decision trees and neural