Random forest aplicado na análise de churn: comparação do ajuste com dados completos versus ajuste em estratos definidos por variável categórica

Abstract

Customer churn, also known as customer turnover, is the loss of existing customers to competitors. It occurs when a customer decides to switch from one company's product or service to another. In order to thrive, companies must understand what drives their customers away if they want to reduce churn rates and retain their current customers. One way to better understand customer churn, detect patterns and predict customer behavior is by using a machine learning technique known as Random Forest. It is an ensemble method that combines multiple decision trees and outputs the most popular class. This study evaluates the impact of the geographic categorical variable when using the Random Forest model to predict customer churn of a financial institution in Germany, Spain and France, based on a public data repository. To do this, a model was created with the complete database, and models with the database stratified by country. The data from the models were analyzed using quality prediction metrics. The premise that the customer's nationality would be able to impact the statistical model proved to be true in a way for the method and data used. The main differences were observed in the sensitivity and F1 score, both related to the classification quality of the true churns in France, the country that represents the largest part of the database proportionally, and Germany. Although the findings of this study are limited to the data addressed and subject to the conditions of analysis specified here, the results showed that, in the case of binary classification via Random Forest, the exploration to stratify or not of the data can generate interesting conclusions from a practical point of view.