Behavioral Modeling for Churn Prediction: Early Indicators and Accurate Predictors of Custom Defection and Loyalty

TL;DR

This paper introduces a comprehensive framework for early churn prediction using large-scale data, combining feature engineering, selection, and supervised learning to achieve high accuracy in identifying customers likely to defect.

Contribution

It presents a unified analytic approach that integrates feature engineering and selection with supervised learning to improve early churn prediction accuracy.

Findings

Achieved 89.4% accuracy in predicting customer churn.

Identified intuitive and surprising early warning signs of churn.

Demonstrated effectiveness on large-scale mobile network data.

Abstract

Churn prediction, or the task of identifying customers who are likely to discontinue use of a service, is an important and lucrative concern of firms in many different industries. As these firms collect an increasing amount of large-scale, heterogeneous data on the characteristics and behaviors of customers, new methods become possible for predicting churn. In this paper, we present a unified analytic framework for detecting the early warning signs of churn, and assigning a "Churn Score" to each customer that indicates the likelihood that the particular individual will churn within a predefined amount of time. This framework employs a brute force approach to feature engineering, then winnows the set of relevant attributes via feature selection, before feeding the final feature-set into a suite of supervised learning algorithms. Using several terabytes of data from a large mobile phone network, our method identifies several intuitive - and a few surprising - early warning signs of churn, and our best model predicts whether a subscriber will churn with 89.4% accuracy.