Information systems in financial markets, e-business, banking, accounting, marketing - comparison of decision tree algorithms in identifying bank customers who are likely to buy credit cards

Abstract

In predictive data mining projects, the decision tree algorithms are usually preferred to others since they are easier to interpret business wise. In this study, we describe and test various strategies in cross sell modeling where we undertake four of the most popular decision tree algorithms CART, C5.0, CHAID and QUEST and compare their predictive accuracy in identifying bank customers who are likely to buy credit cards. The comparison of the accuracies of the algorithms is made twofold. First, confusion matrices which reflect the strength the learning process are compared. Secondly, the hit rates of the algorithms in a later period are measured and compared. The relationship between the hit rates and the confusion matrix values is also investigated and some useful hints to developing more accurate models are derived. Experimental results indicate that CHAID algorithm outperforms the others.