Churn prediction using advanced machine learning models and identification of churn drivers to create a churn management strategy
According to loyalty expert Frederick Reichheld, with customer defection rates in double digits across most industries, firms can lose half their customers in five years, leading to a depression of corporate performance by as much as 25–50%.
It is also a well-known fact that it is multiple times more expensive to acquire a new customer than to retain an existing one, and high attrition rates increase replacement costs of existing customers. A good retention strategy creates competitive advantages, provides a boost to employee morale, produces gains in productivity and growth, and reduces cost. The results can be impressive. It has been observed that a 5-percentage point improvement in customer retention can result in a 75% and upwards increase in customer lifetime value across a range of industries.
Churn in the telecom industry has become a serious concern with rising subscriber churn rates in the backdrop of a global trend of declining ARPU (Average Revenue Per User). Use of machine learning for churn prediction has been an area of focus in the mobile telecom industry; however, the application of the same has been less extensive in the area of broadband internet services.
The return on investment (ROI) on churn models in telecom has also been a challenge due to the dynamic nature of the churn and the high costs of retention offers; such high costs of retention offers tend to increase the cost of false positives and thereby reduce the net ROI of most churn models.
About the Client
Atria Convergence Technologies Limited., branded as ACT, is an Indian telecommunications company headquartered in Bangalore, Karnataka. ACT offers fiber to the home services under the brand name “ACT Fibernet” and digital television services under the “ACT Digital” brand.
Actify further worked closely with the churn management team to put in place a business strategy that would use the output of the model to create various churn management and calling strategies. The model along with the strategy recommendation was effective in reducing churn significantly. Based on an optimized strategy the models were able to save significant amount of churn leading to an impact of ~USD 150,000 per month.
Actify Data Labs had been the data science partner for ACT Fibernet (ACT), one of the largest broadband telecom company in India. ACT operates in a high growth and hyper competitive market where new players are aiming to disrupt the market through extremely aggressive pricing strategies. While on one hand, ACT was acquiring new customers at a very fast pace; on the other hand, the company was losing a large volume of existing customers. Hence, ACT not only needed the ability to predict churn accurately, it also had to understand the potential reasons as to why a given customer might churn. An understanding of the potential reason of churn at an individual customer level would enable ACT to devise customer specific retention strategies.
Actify Data Labs developed a machine learning solution to predict churn and reconnection (from the already churned customer base). Prior to model building, Hive was used to process the large volume of granular transaction data to create a modelling-ready data. Natural Language Processing (NLP) techniques were used to extract features from call centre notes and complaint texts. Multiple machine learning methodologies (ANN, Gradient Boosting, Random Forest and Logistic regression) were attempted and a Gradient Boosting based model was found most optimal for predicting both churn and reconnection.
ACT was also keen to understand the potential reason for churn at an individual customer level. A black box methodology like Gradient Boosting had the challenge of providing intuitive understanding of the drivers of churn. Hence, Local Interpretable Model-Agnostic Explanations (LIME) and Shapley value was used to provide an intuitive understanding of the drivers of churn for each individual customer.
The model was implemented by the Actify team within the technology environment of ACT. Actify also provided a set of utilities which would evaluate the effectiveness of the solution every week (with respect to accuracy, stability and discriminatory power).
Agile able to incorporate new findings fast and collaborative.