Applying Data Science to Predict Customer Churn
7 steps to get from data to insight in your churn analysis.
Written by Eshwarya Agarwal
Understanding churn behaviour is complex. In our last article we explored how Customer Analytics can help identify customers at risk of churning, looking at the broad approach and data needed.
In this article, we go a level deeper to look at practical steps to undertake the churn analysis itself, sharing some of the methodology we apply when working on a churn analysis project.
1. Define “churn”
What is churn for your business? Defining churn is important and should ideally be kept constant over time. The definition differs for different industries and business models.
For subscription-based businesses, churn is usually defined as customers who have terminated their contract before the renewal date and/or those who have not renewed their contracts.
For non-subscription businesses, the same churn modelling can be applied, but it’s a little trickier. You could define churn as customers who have not made a transaction in the last 30/60/90 days depending on the type of products or services your brand offers. To avoid classifying your occasional buyers as churners, you might first define a subset of your customers as frequent buyers, then define churn as people whose purchase pattern is slowing or has stopped.
2. Define your problem and success criteria
Every company faces customer churn problems and the outcomes/strategies that businesses adopt to tackle it are very different.
- – Why do you want to understand churn?
- – Are you seeing an increasing churn trend?
- – What do you want to gain by doing churn analysis?
- – Are you trying to retain more customers?
- – Are you trying to understand churn behaviour?
Answering the above questions will help you choose the right course of analysis and will get you the right/usable insights from your churn analytics project.
3. Gather relevant data and perform data quality checks
Typical data-sets that you need to start the analysis are customer’s historic transactional data, demographics data, survey/feedback data, contact centre data, etc. to begin the analysis.
No data is perfect and will typically require cleansing and massaging to get it in the right shape. Once the analytical data-set is ready, you can now jump into the analysis!
4. Who is churning?
Before going deep into analysis, it is important to understand who is churning. Are there particular customers cohorts which have a very high churn? Churn rate high in particular demographic, significant difference between new vs old customers, etc. This will help you gauge the customer churn scenarios. Once we have understood who is churning, we would like to understand what drives customer churn.
5. Understand what is driving churn
What factors are causing customers to churn? Are they moving to competitors? Or has there been a change in price for the products/services, the service quality, etc? Also look for season effects and other sources of insight into bigger trends such as social media.
By exploring the trends and patterns in the data we can start to get hints about the reasons for churn. Some patterns are hidden and will come out when you create new data points with a combination or transformation of existing variables.
Prepare data with these factors and other derived variables, before moving on to the next stage.
Since churn is impacted by various factors and the factors differ for each customer, we would go ahead by modelling these complex behaviours with the power of mathematics and statistics!
We use machine learning techniques like Random forest, XGBoost, Logistic regression, etc. to get the best results.
Which technique to use when depends on the business problem. The input to this decision will come from Step 2 where we are defining our objective and the need to do this analysis.
In most scenarios, we have to try more than 1 technique to see which algorithm fits best in that particular business scenario.
Once the model is trained and tested, in return we will get churn probability (or score) for each customer which we can use to identify and target the potential churners. This is normally a measure of the accuracy of the prediction that someone will churn. We can then look for correlated groupings within the set of customers with a high probability of churning, which give us micro-segments for targeting in retention campaigns.
7. Increasing the accuracy
This completes the first round of analysis. If completed as a stand-alone process, you now have an insight into the what, who and why of churn in your business. To improve the accuracy of the model and account for changes in customer behaviour and business dynamics, the model will go through a feedback mechanism where it can be evaluated and improved over time.
This means building churn analytics into your data operations, business intelligence and marketing automation processes – more on this in our next article!