Loading...
Please wait, while we are loading the content...
Similar Documents
Defection Detection : Improving Predictive Accuracy of Customer Churn Models
| Content Provider | Semantic Scholar |
|---|---|
| Author | Neslin, Scott A. |
| Copyright Year | 2004 |
| Abstract | and to an anonymous wireless telephone carrier that provided the data for this study. We also thank participants in the Tuck School of Business, Dartmouth College, Marketing Workshop for comments. Abstract This paper investigates how methodological factors contribute to the accuracy of customer churn predictive models. The study is based on a tournament in which researchers from business and academia downloaded data from a publicly accessible website, estimated a churn prediction model on that data, and made predictions on two validation databases. These predictions were merged with the actual churn records for the validation data and scored on two criteria, top-decile lift and the Gini coefficient. A total of 33 participants submitted 45 entries. The results suggest several important findings. First, methods do matter: The differences observed in predictive accuracy across submissions could change profit contribution by $100,000's. Second, models have staying power: The churn models in our data typically have very little decrease in performance if used to predict churn for a database compiled three months after the calibration data. Third, researchers use a variety of " approaches " to develop churn models, described by a combination of estimation technique, variable selection procedure, time allocations to various steps in the model-building process, and number of variables included in the model. These approaches can be labeled " Logit, " " Trees, " " Novice, " " Discriminant, " and " Explain. " We find that the Logit and Tree approaches are associated with relatively higher predictive performance, the Novice approach is associated with middle-of-the road predictive performance, while the Discriminant and Explain approaches are associated with lower predictive performance. We discuss implications of these results for both researchers and practitioners. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www.crmlandmark.com/library/customerchurn.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
