Modelling for the optimal product to offer a financial services customer
Mukomberanwa, John Shingirai
This study, illustrates how various statistical classification models can be compared and utilised to resolve cross-selling problems encountered in a financial services environment. Various statistical classification algorithms were deployed to model for the appropriate product to sell to a financial services customer under a multi-classifier setting. Four models were used, namely: multinomial logistic regression, multinomial bagging with logistic regression, multinomial random forests with decision trees and error correcting output coding. The models were compared in terms of predictive accuracy, generalisation, interpretability, ability to handle rare instances and ease of use. A weighted score for each model was obtained based on the evaluation criteria stated above and an overall model ranking thereof. In terms of the data, banked customers who only had a transactional account at the start of the observation period were used for the modelling process. Varying samples of the customers were obtained from different time points with the preceding six to twelve months information being used to derive the predictor variables and the following six months used to monitor product take-up. Error correcting output coding performed the best in terms of predictive accuracy but did not perform as well on other metrics. Overall, multinomial bagging with logistic regression proved to be the best model. All the models struggled with modelling for the rare classes. Weighted classification was deployed to improve the rare-class prediction accuracy. Classification accuracy showed significant limitation under the multi-classifier setting as it tended to be biased towards the majority class. The measure of area under the receiver operating characteristic curve (AUC) as proposed by Hand and Till (2001) proved to be a powerful metric for model evaluation.
A research report submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in fulfilment of the requirements for the degree of Master of Science. Johannesburg, 2014.