The Great Retail Escape: Catching Runaway Customers with Machine Learning - A Study on Customer Churn Prediction in the Retail Sector

Introduction: This thesis aims to address the challenge of accurately predicting customer churn in a highly competitive and digitized retail environment. As customer loss can lead to significant financial consequences, the aim is to support businesses in identifying at-risk customers early, enabling more effective retention strategies and reducing the risk of customers switching to competitors. Specifically, the thesis explores how interpretable machine learning models can be leveraged to uncover behavioral patterns predicting churn, evaluate the effectiveness of predictive models, and estimate customer lifetime value in order to assist in better resource allocation strategies for companies. Research question: The research questions of this thesis are: 1. What customer behavior patterns are the most significant indicators of churn in a non-contractual retail setting? 2. What machine learning model provides higher predictive accuracy in identifying customer churn in a non-contractual retail setting? 3. How can machine learning models be applied to estimate customer lifetime value in a non-contractual setting? Method: The research strategy employed in this thesis is experimental in nature, as it aims to systematically explore and evaluate the performance of machine learning models. It takes a quantitative approach, using supervised machine learning models to analyze patterns in customer behavior and predict churn. By applying the quantitative methods of logistic regression, decision trees and random forest, the thesis aims to uncover key predictors of churn, determine the best model and estimate customer lifetime value. The data analyzed in this thesis is the synthetic dataset called Online Retail Customer Churn Dataset and was retrieved from the Kaggle platform. Results: The results showed that decision trees outperformed logistic regression and random forest in terms of identifying churners. The analysis of feature importance highlighted transactional patterns and engagement behavior as the most important predictors of churn. To estimate customer lifetime value with machine learning a binary variable was needed to enable classification. The strongest predictors of high or low customer lifetime value across all models were spending behavior and marketing response variables, and the random forest model achieved the best F1-score on the better predicted class. Discussion: The findings suggest that customers’ transactional patterns and engagement behavior are more reliable signals of future churn than who they are demographically. From a business perspective, this insight could be valuable for targeting retention efforts. The decision tree outperformed the other models in detecting churners, which is often the primary focus in churn prediction. Ultimately, the choice of model should be guided by the specific objectives of the churn management strategy and the relative costs of false positives and false negatives. Additionally, the findings suggest that in non-contractual settings, customer lifetime value estimation with the help of machine learning is best approached through modeling with behavioral and demographic features. Some limitations of the study are reproducibility, concerning the code, validity, concerning the historic customer lifetime value approach, and generalizability, due to the relatively small dataset. Further research could explore why transaction amount consistently emerges as a strong predictor in churn, why the number of returns stood out as the most important feature in the decision tree, and study customer lifetime value on a deeper level to predict monetary values and to either corroborate or refute identified predictors.