Customer Churn Analysis And Prediction Of A Certain Mobilcom Based On Boosting Algorithm

Resources are very limited for many enterprises. How to make good use of limited resources is one of the key problems needing to be considered in the process of enterprises development. The problem is especially worse in Mobile Industry because of the specialty in this area. At the moment, many Mobile Telecom carriers take great efforts to develop new clients instead of focusing on the effective maintenance of their existing customers, resulting in the so-called Revolving Door Effect. Namely, more customers churned while the company is devoted to developing new clients. The above mentioned situation leads to the waste of internal resources, suffering Mobile Telecom carriers a lot.In order to study the above problem, this essay chooses a certain mobilcom as research object. On the better understanding of internal circumstances and the careful analysis of the problems in the company, it is concluded that the problem lies in its great number of customers and the uncertainty of many customers'behavior. Therefore, it is very hard to make effective maintenance and many resources are needed in the operation. So a prediction mechanism is of great necessity to forecast customers'future behavior. In consideration of these, this essay sums up six basic causes of the customer churn in the company and adopt stepwise selection method to make corresponding data processing from the respect of whether identification card information is integrity, defining the data time range etc. separately by using general information and consumption information provided by the company, and then generalizes fifteen variables which may affect customer churning. After all these operations, the author imposes the bar graph and Chi-square test to verify the variables'predictive power and chooses thirteen variables as the model's variable set. On these bases, the author builds up a corresponding Boosting algorithm model by programmed in Matlab, and uses this model to study the actual data of the company. Meanwhile, the outcomes of the prediction on 999 pieces of data, 504 pieces of churn data and 495 pieces of non-churn data, in the testing set are 509 pieces of churn data and 490 pieces of non-churn data. Based on these results, the index values are calculated. It is found that the Accuracy comes to 93.69%, the Sensitivity amounts to 94.25%, the Precision reaches 93.32% and Specificity achieves 93.13%.The diagrams and the index values acquired from the experiments prove that the model possesses fairly good precision and excellent performance at the churn prediction problem introduced in this essay.