Customers’ churn analysis in a SaaS company focused on customer relationship marketing

Abstract

Customer Relationship Management has become increasingly vital for business success in the digital transformation era. In particular, customer churn prediction has emerged as a crucial task for subscription-based services like Software-as-a-Service (SaaS). This study investigates churn prediction at E-goi, a B2B SaaS company focused on customer relationship marketing, by applying various machine learning techniques within the CRISP-DM framework. The primary objective was to identify the most effective predictive model and evaluate its practical implications for customer retention strategies. The methodology included testing models using three different sampling techniques, undersampling, SMOTE, and original class distribution. Random Forest, XGBoost, Decision Tree, Logistic Regression, Feedforward Neural Network (FNN), and k-Nearest Neighbours (KNN) were chosen to be tested and assessed using F1-score as the primary metric due to the imbalanced nature of the dataset. Grid Search hyperparameters tuning was applied to find the best parameters before final model training. Time-sliced cross validation (TSCV) approach was chosen to validate the results to maintain the chronological integrity. The RF and XGBoost using the original class distribution emerged as the most effective models, achieving F1-scores of 0,6752 and 0,6703, respectively. While RF demonstrated superior recall, making it better at detecting actual churners, XGBoost showed higher precision, minimising false positives. Sampling techniques, including SMOTE and undersampling, did not consistently improve model performance, with the original data distribution yielding the most reliable results. The study also employed SHAP to analyse feature importance, revealing that variables related to payment activity and customer engagement were the most influential predictors of churn. These insights underline the importance of proactive monitoring of key customer behaviours to mitigate churn risks.

Date of Award	23 Jul 2025
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Vera Lúcia Miguéis Oliveira e Silva (Supervisor)