Customer attrition in retailing: An application of Multivariate Adaptive Regression Splines

• We propose a model for partial churn prediction in retailing.
• We use Logistic regression and Multivariate Adaptive Regression Splines as classifiers.
• We compare the performance of MARS with Logistic regression.
• We consider Logistic regression combined with stepwise feature selection.
• Stepwise feature selection approach outperforms MARS.

The profit resulting from customer relationship is essential to ensure companies viability, so an improvement in customer retention is crucial for competitiveness. As such, companies have recognized the importance of customer centered strategies and consequently customer relationship management (CRM) is often at the core of their strategic plans. In this context, a priori knowledge about the risk of a given customer to mitigate or even end the relationship with the provider is valuable information that allows companies to take preventive measures to avoid defection. This paper proposes a model to predict partial defection, using two classification techniques: Logistic regression and Multivariate Adaptive Regression Splines (MARS). The main objective is to compare the performance of MARS with Logistic regression in modeling customer attrition. This paper considers the general form of Logistic regression and Logistic regression combined with a wrapper feature selection approach, such as stepwise approach. The empirical results showed that MARS performs better than Logistic regression when variable selection procedures are not used. However, MARS loses its superiority when Logistic regression is conducted with stepwise feature selection.