Predicting Customer Churn in Telecom Using Logistic Regression and Decision Trees
0 viewsAbstract
Customer churn is costly for telecom operators, and managers need models they can understand and act on. This study compares two interpretable classifiers, logistic regression and a decision tree, for predicting churn, and examines how the classification threshold affects recall and precision. Using the public IBM Telco Customer Churn dataset (7,032 customers after cleaning, of whom 26.6% churned), the models were trained on 80% of the data and evaluated on a held-out 20%, with bootstrap confidence intervals and an exact McNemar test used to assess differences between them. Churn was most strongly associated with month-to-month contracts (42.7% churn, versus 2.9% for two-year contracts), fiber optic internet service (41.9%), electronic-check payment (45.3%) and short tenure (median 10 versus 38 months). Logistic regression achieved slightly higher accuracy (79.7% versus 77.9%) and AUC (0.835 versus 0.822) than the decision tree, but the difference was not statistically significant (McNemar p = .061). Lowering the threshold from 0.5 to 0.3 raised the logistic regression's recall from 56.4% to 75.9%, while precision fell from 63.4% to 50.7%. The findings suggest that a simple, interpretable model is sufficient for this problem, that the threshold should reflect the cost of retention offers, and that contract length and early-tenure customers are natural targets for retention efforts. The data come from a fictional single-company sample, and the results show associations, not causes.
Keywords
Classification Threshold, Customer Churn, Customer Retention, Decision Tree, Logistic Regression, Telecommunications