A Comparison of the Effectiveness of Techniques for Predicting Binary Dependent Variables

Abstract

The Invoice Disputes Team wants to identify further opportunities to reduce invoice disputes and the team would like to explore whether data analytics can drive further improvements. The purpose of this paper is to compare the effectiveness of the eight approaches to predict binary dependent variables according to the specified data. The techniques examined are Logistic Regression, Probit Regression, CHAID, CART, Neural Networks, Bagging, Random Forests and Boosting. This paper describes the data set, the effectiveness measures used and the approaches, and also shows the results for each of the eight approaches that are examined. The simulation results show that both Bagging and Random forests seem to do better than other approaches.