Telecommunications Churn Solution Overview
The Challenge
A major U.S. telecommunications provider with tens of thousands of employees and annual revenue of tens of billions requested assistance regarding prediction of which of their existing customers were likely to churn.
High voluntary customer deactivation rates, referred to as churn rates, are a problem faced by all mobile communications providers. Modeling and predicting the propensity of customers to churn is essential for
determining the cause of these deactivations, implementing interventions and ultimately lowering churn rates. The telecommunications provider’s churn model did not have sufficient predictive power and suffered from long execution times, limiting its utility. In addition, the provider’s model did not take into account social networking data, such as which customers in a subscriber’s calling network have churned and when, which can have a substantial impact on customer churn rates. To address these limitations, the telecommunications provider partnered with Dataspora to create a model with more predictive power.
The Solution
Dataspora analyzed data from approximately 500,000 customers that signed up for a specific rate plan. A churn model was created in R using a Generalized Additive Modeling (GAM) technique. The Dataspora model was shown to have the same amount of predictive power as the client’s model, before taking into account any social networking data. In order to include social networking variables, call histories were extracted for given subscribers for three consecutive months prior to the final analysis of churn rate. A high-risk category of consumers called “churn-chatters” was identified. Churn-chatters are defined to be subscribers who receive more than 10% of incoming calls from churners. Before incorporating social networking data, the model predicted the churn rates of customers not considered to be churn-chatters fairly well. For churn-chatters, however, the churn rate increased quite dramatically and the model no longer accurately predicted the churn propensity.
By adding social-networking variables (see left), Dataspora built a model that describes customers as a whole, while also describing high-risk groups such as churn-chatters. Dataspora was also able to quantify the influence of certain factors on churn-chatters versus all others and demonstrated that variables such as length of remaining contract, activity level and tenure affect churn-chatters more strongly than the other subscribers. The model also revealed that subscribers with a higher number of calls to customer care agents have a higher propensity to churn.
In order to further improve the model, Dataspora recommended that the telecommunications provider use data from a more comprehensive sample of subscribers. The specific rate plan used for this project skews the data because these customers are all bound by contracts, which likely plays a large role in churn rate. Furthermore, the model was designed to have execution time on ~1 million rows in the minutes range. Taking advantage of this, the provider could run the model scoring on a daily basis and more closely monitor customer activity. Dataspora suggested implementing an early detection warning to identify high-risk subscribers early, thereby allowing time for the provider to make the right interventions at the right time.
The Benefit
Detecting high-risk subscribers early is crucial for the detection and prevention of churn. Connecting with these customers promptly and providing them with the right resources, offers for free or discounted products, or services at the right time can potentially prevent churn. The model developed by Dataspora can be run on a daily basis within the provider’s operational workflow. By leveraging the model’s output to identify early warning signs of churn, the telecommunications provider now has a valuable tool to aid in lowering churn rates. Additionally, Dataspora can go beyond identifying which customers will churn to determining the best course of action that will prevent the customer from churning before it happens.
