Churn prediction using supervised machine learning
29 Jun 2021

Churn prediction using supervised machine learning

Each year, several product and service providers lose between one and twenty percent of their clients. Churn prediction, which predicts the likelihood that a customer will reuse a service/product, has become a top priority, since keeping customers is easier than finding new ones. A company's loyalty and service strategies need to be adjusted effectively. Over the last several years, contract data, customer profiles, the identities of branches and agents providing the service, photographs, and messages were also collected. A company's retention of customers will likely benefit from the insights and facts contained in these elements.

Artificial intelligence, and particularly supervised machine learning, is then a powerful tool for analyzing data from heterogeneous sources (text, images, tabular data, graphs) to make predictions. A natural approach is to predict if a contract might be canceled based on the customer's profile associated with it, its previous contracts, its history, the serving branch, and how it interacts with the customer. The results of the produced predictor then provide important elements to understand customer unsubscribes. 

A model determining the minimum change in contract-related characteristics needed to flip the forecasted result can be used for contracts with the "will be cancelled" prediction to determine what can be done to retain the associated clients. These are data-driven suggestions to keep clients. It's important to provide them to a retention specialist to validate them, detect algorithmic biases, and plan the retention strategy. 

In some cases, a predictor analysis will reveal cancellations have nothing to do with client satisfaction. One example is when a person dies or a contract is transferred. Therefore, the true percentage of cancellations should be determined. In this sense, we are improving the quality of the data. In addition to refining data, there may also be other aspects to address, such as defining a churning client better, eliminating useless attributes, and taking into account branch mentalities. With each iteration, the analysis becomes increasingly related to the business case and passes a cost-benefit analysis. 

Data scientists can refine and process most data. Nevertheless, it is recommended and has been proven that the best way to improve data is to design better data collection processes. One can think about including an overall service rating in the cancellation interface in the context of churn prediction. It will enable us to distinguish between satisfied and dissatisfied customers, to document the successes and failures, and to finally improve the services to retain the unsatisfied customers.