For many business leaders, the words “Customer Churn” or “Customer Attrition” send a shiver down the spine (and not in the same way as hearing your favourite song at a concert). It is widely understood that new customers are far more expensive to obtain than simply retaining an existing customer, and I use the word “simply” loosely here. Why then, is customer churn still such a problem for many services businesses? What makes Jane change her mobile carrier at the end of her contract? Is it the same reason that John cancels his streaming service in favour of another?
Probably not.
Customer churn costs businesses an exorbitant amount in lost revenue every year, and so a reduction in customer churn is a logical move for a business looking to increase revenue in an ever-competitive business environment. In this blog post I will detail several Artificial Intelligence (AI) methods available that can be used to predict customer churn including: Gradient Boosted Classifier, Support Vector Machine, Random Forest Classifier, K-Nearest Neighbour, and Logistic Regression.
The Bigger Picture:
Measuring customer churn can be as simple or as complicated as you like. At its most basic level, customer churn is simply the number of customers lost divided by the total number of customers in a given period e.g.
Customers lost / total customers for the year 201X
The numerator and the denominator can be defined in all sorts of ways depending on your business. Elements such as seasonality, customer segments, and sample size may be taken into account. Before considering the approach to use, it is essential to understand your data and the variables that may be important to the analysis. As with most AI methods, the more data the better. The models may be able to find links between disparate variables that help to predict customer churn. Figure 1 below (kindly provided by Bain and Company) is a fantastic visual representation of the circumstances leading up to a churn event. A key activity in predicting customer churn is feature/variable importance, that is, using AI methods to assign a weighting indicating how important each feature/variable is to predicting customer churn. Gathering data from each stage in the churn funnel is an important activity in ensuring a robust churn dataset.
Figure 2 shows the features analysed for a client in the telecommunications industry using a Gradient Boosted Classifier. Reading this figure can require digging into the drivers of each feature to find insight. The third most important feature shown is the number of international calls which might seem like an odd predictor of customer churn, though it might indicate that many customers receive large phone bills after calling their overseas home and change services as a result. This may mean that the fees for international calls are deemed to be unreasonably expensive or opaque and require review - an unexpected action item from a churn analysis. The Total Day Minutes and Customer Service Calls features go hand-in-hand as the 1st and 2nd most important features respectively. A rule extrapolated from this ranking can be “Customers making more customer service calls and spending more time on said calls are more likely to churn”.
Once the feature importance has been analysed, rules can be written that look for specific combination of features or features values. These are then classified using a machine learning approach into the desired buckets e.g. “churned” or “not churned”.
The Methods:
The machine learning part of this analysis is a classification task, one that can be performed by several methods. The five shown below in Table 1 are some of my preferred classifiers. Using the same telecommunications client example from above, I have shown the accuracy score produced by each method using a Python package called SKLearn. This score references the accuracy with which the model predicts the values of the test samples, that is, a customer that has churned or not. A score of 0.95 indicates that the model correctly classifies a customer in the test set 95% of the time.
Method Accuracy Score
Gradient Boosted Classifier 0.95
Support Vector Machine 0.92
Random Forest Classifier 0.95
K Nearest Neighbour Classifier 0.89
Logistic Regression 0.86
Table 1: A summary of the accuracy scores by classification method
Machine learning methods perform differently under different scenarios i.e. different datasets and use cases. For this reason I generally produce results using several methods and assess the accuracy of each before pushing a model into production. The 5 shown in table 1 above are but a few from the many that are available.
Next steps:
The key to successfully combating customer churn is a nimble and agile approach, and there are 2 areas where business leaders can give themselves every chance of success in my mind:
1. Follow up structure:
If an AI model produces an insight such as a classification of a valuable customer as having a high risk or changing services, the company must have process in place to follow up with said customer to see whether there is anything to be done to stop them leaving. This process must be trackable such that any attempt made to keep the customer can have its success measured.
2. Humility to make a change: