Airline Loyalty Analysis ✈️

Exploring churn patterns using GLMNET

Background

Used a dummy airline loyalty program dataset to uncover predictors of churn using a logistic regression machine learning model. Check out the dashboard here!

Executive Summary

This analysis identified key drivers of churn in a loyalty program. A few insights: Customers who accumulate points but do not use them have a higher risk of leaving since they are not utilizing their rewards. Relatedly, flight frequency is another big predictor of churn. Occasional flyers are more likely to leave the loyalty program. Finally, the length of someone’s membership makes a difference in their loyalty; people who signed up prior to 2018 are less likely to leave.

Methods

Data cleaning - imputed salaries based on median salary in residing city

Fit a logistic regression model using GLMNET using L1/L2 regularization to predict churners in an airlines loyalty program (dummy data) with 73% recall, 84% precision, and 95% overall accuracy.

Key Model Metrics

• Accuracy: 94.92%
• Sensitivity (Recall): 72.88%
• Specificity: 98.02%
• Precision: 83.84%
• F1 Score: 78% (at threshold 0.75)
• AUC: 0.923

Recommendations

Because of our model, we can predict the likelihood of a user churning based on demographic and behavioral traits. To improve retention, high-risk members can be sent targeted offers that can remind them to use their points and provide additional promotions to help them utilize the benefits of the program.

Visualizations

• the top predictor of churn is enrollment_typeStandard, with a strong negative coefficient, meaning users on this plan are significantly less likely to churn. enrollment_years 2014–2016 have the highest positive coefficients, indicating those cohorts are at greater churn risk compared to the reference year.

• features like total_years, total_flights, and education level show smaller but consistent associations—users with more tenure, more flights, or higher education tend to be slightly more retained. overall, plan type and cohort year are the most influential drivers in the model.