Domain: Data Science
Client: Telecom sector in the Netherlands
Subject: Predicting customer churn
A Dutch-based provider of telecommunications and information technology services has a high percentage of customers cancelling their subscription on monthly basis.
Customer retention is one of the important elements for growth, especially in subscription-based services. In general, it can cost five times more to attract new customers than to retain existing customers. So, churn is an important KPI for businesses with customers as subscribers with recurring payments which makes churn rate important to measure and act accordingly.
"Predicting customer churn"
The project
The main goals of the project was to decrease the churn rate using all the data available about the customers and their history.
How we handled this
We set up a pilot to identify customers ‘at-risk’ to churn using predictive modelling and launched an email campaign to these customers with a sales offer. What we did briefly:
- Create a customer view with all available historical data: clicks on website, subscription details (GB, mins, fee), customer demographics, customer service tickets, CRM, etc.
- Train- and test different classification models on the data to identify customers ‘at-risk’. ML algorithms we used are decision tree, random forest, logistic regression, XGBoost and SVM
- Model evaluation and - selection
- Use ‘winning’ model to score existing customers resulting in a churn probability for each individual customer (satisfied customers vs customers at risk)
- Set up a campaign for customers at risk of churn
- Campaign evaluation
Step 1 to 3: customer view, training- and testing & model evaluation and - selection
Results
- A random forest model with an AUC of .91 and accuracy of 86%
- Important predictors of churn are: number of months till end of contract, subscription price, subscription with or - without discount, customer and provider relation length in months, history of non-payments
- Existing customer base with churn probability for each customer using the random forest model. Initial retention strategy: top 10% of customers with highest churn probability received an offer to renew their subscription with a discount
- Average churn-rate decreased from 4% a month to 3,25% a month after the campaign
- At the end of the campaign the ROI was calculated. Investments included updating and running the predictive model periodically and initial campaign costs (setup and lost revenue). Returns included incremental revenue due to lower churn rate
Conclusion
We demonstrated briefly how to set up and execute a data driven retention strategy step-by-step using a churn model. Although the predictive model performed as expected and the churn rate decreased, overall the ROI was negative for this specific business case. The company and the customer base was too small at that given time to make the pilot a solid investment. The results of this pilot were presented within the organization to encourage more data driven solutions for specific business challenges.
Do you have any questions?
Your contact person
Arjan Schoe
Director Data & Analytics
Want to get more value out of your data? Whether you are at the start, middle or end of the journey to become a data-driven organisation, we help you to understand and prioritize the keys to success in each stage. Get in touch with us!
