Statistical Analysis And Forecasting Of Customer Churn

In the stage of high homogeneity of the product marketing, the mainperformance of competition among these enterprises depends on the satisfaction ofthe customers. This makes every business pay attention to customer churn problem,which is having more and more effect on the development and income of thebusiness.An early warning mechanism of the customer churn can solve the problem ofreduction of market share, increase in marketing costs, and decrease of income, whichoccurs due to the customer churn. With these results, companies win back customersdirectly, so as to maintain the level of corporate profits.Therefore, in recent years,churn research has become “Hot Issues” in many industries. Based on traditionalstatistical methods and data mining methods, the research on churn problem hasachieved fruitful results. Despite all the valuable results, there are still some problemsthat need further research.In this paper,Generalized Linear Mixed Model (GLMM) is introduced to theresearch of customer churn problem.By setting an appropriate random effect, thismethod can separate the defecting clients and customers of normal status effectivelyand achieve the goal to determine the model of client status prediction. This couldalso deepen the understanding of the costumer churn problem from a new point ofview. With this framework of the prediction model, this paper uses the customer dataof Kunming Telecom to empirically analyze the customer churn prediction model.Then compares the prediction results between the GLMM and other methods. Themain conclusions are as follows:1. Introduce the Generalized Linear Mixed Models to the study of the customerchurn prediction, which provide a new research path for this field. With the researchof the customer consumption data, we found that, it is a typical longitudinal data. Andthe traditionalcustomer churn prediction research methods have a very similar place with the structure of generalized linear mixed models. GLMM not only can face thenon-normal data problem, but also can detect the heterogeneity and correlation byadding random effects. So we consider introduce the Generalized Linear MixedModel to analysis the customer churn problem.2.It is reasonable to set the regional factor as a random effects to build customerchurn prediction model. Based on the newly constructed analytical framework,wefound the major factors about the customer churn by using the Kunming Telecomcustomer data, then set acustomer churn prediction model.Found that set the regionalfactor in to the GLMM model can effectively improve the fitting effect and theprediction accuracy.This provides a theoretical basis for the telecomenterprise torecoup their customers, also corroborated the applicability of GLMM for the customerchurn prediction problem.3. Use the misclassification costs as the weight to build the evaluation index formodel compare is more scientific. In the analysis of the customer churn problem, theproportions of the two types of customers are unbalanced.It is easy to produce theillusion of the evaluation results by using the traditional evaluation index to comparemodels. So we add the misclassification costs into the evaluation index, which canmake the results much closer to the real. Then, by calculating the index, we found thatGLMM get an excellent performance, which demonstrate the effectiveness, reliabilityand practicality of the Generalized Linear Mixed Models.