Müşteri verisinin daha etkin kullanımı ile müşteri kayıp tahminin iyileştirilmesi: Özel bankacılık sektöründe bir uygulama

Based on the fact that acquiring new customers in any business is much more expensive than trying to keep the existing ones, customer retention is an increasingly pressing issue in today?s ever-competitive commercial arena. This is especially relevant for service related industries. Thus, classification models to detect churning customers have received significant attention in the customer relationship management literature.In this thesis, we focus on the churn prediction problem in the private banking industry under non-contractual settings. Churn models are typically estimated on cross-sectional data pertaining to a particular time period and used for prediction in subsequent periods. This is appropriate under static settings where the sample size of churners is sufficient; however churn is a process and customer behavior is affected by changes in the environment. We show that modeling next-period churn behavior with multiple observations per customer pertaining to different time periods yields better predictive performance when compared to traditional cross-sectional training data, with or without synthetic oversampling techniques, and provides additional managerial insights. Further, reflecting the idea that customers do not always decide on and carry out the churn action overnight; we propose to model churn with multiple models that predict churn several periods ahead, and to use these predictions in an ensemble for improved next-period churn prediction. This approach provides the company with advance notice on customer?s churn propensity, improves out-of-sample next-period churn prediction, and ensures consistency of advance propensity figures with next-period prediction. We evaluate our models with data from the highly dynamic banking industry.