The client’s risk model had been developed and deployed by dozens of expert statisticians over many years. The client was interested in whether any new insights would be produced by using modern machine learning techniques. To be adopted, the new model needed to have significantly better predictive performance and be suitable to run in the client’s production scoring environment.
Analyzing wireless carrier churn is a complex analytical challenge. Many variables are involved and some are difficult to measure accurately. Because of the complexity of the assignment, Elder Research implemented a multi-step approach. The initial process included correcting software deficiencies and improving the accuracy of input data. Elder Research built predictive models connecting churn to key variables and predict 90 days in advance which customer were most likely to churn. The call center could then target high-risk customers with appropriate incentives for a positive intervention. To improve predictive probabilities, Elder Research used text analytics to extract valuable information from customer comments recorded by the call center, increasing the churn targeting effectiveness by 3.1 percent. Survival analysis was used to identify larger-scale trends. Using this technology, Elder Research was able to show nTelos that customers who prepaid for services by an automatic debit arrangement were less likely to churn than customers who prepaid by credit card, and far less likely to churn than customers who paid with cash. Based on this information, management modified its prepayment plans to encourage automatic debits, and churn decreased for this category.
Measurements conducted using statistically significant control groups for multiple categories of customers showed that Elder Research’s solution decreased churn from 3.5 percent to 2.9 percent — the lowest level in years. The new models prioritized likely churners so that intervention was twice as effective as before and boosted annual profits by more than $1 million.