Business Problem:
Churn is when customers cut ties with the firm either because of poor customer service or competitive offered a better deal. It is usually measured on a yearly basis. However, this could be done more frequently if subscriptions renew every month, which is the case for popular mobile service plans.
It is important to figure out what the churn rate is to retain them because the cost of acquiring new customers is far more expensive (5-25x more expensive).
Determining those that are churning can be determined by app sessions, engagements, conversions, chats to customer service, number or canceled orders to name a few. When certain behavior is linked to customer attrition, churn prediction can be implemented for a business to intervene earlier.
And this would help decrease churn rate, increase total active users YoY and profitability.
Metric to Lift:
Total Active Users QoQ/YoY
Hypothesis:
The sample data used is from a business that provides internet/phone service. The best guess for churn for this type of business is that fee is too high for a given package or speed given competitor’s rates that may change month-to-month.
Analysis Approach:
- Did some data cleaning since there were some NULL/missing values to get a more probable picture of sample population.
- Split population between those who churned and those who did not churn.
- Looked at all given columns about customers versus churn and non-churned population and able to see some strong and not-so-strong correlations with churn and non-churn behavior.
- Cleaned up data for machine learning models’ training and test data. Decided to use 30% of data for testing and therefore 70% of the data for training.
- Used multiple classifiers to predict churn and measured accuracy as well as f1-score.
- Best performing model was Voting Classifier that was an aggregate of other models.
Recommendation to Business Stakeholders:
Have an effective retention strategy that focuses on price optimization for customers.
Conclusion:
Highest correlator to churn was month-to-month prices. And best model for churn prediction that could use improvement in churn prediction was Voting Classifier.