Explainable Machine Learning for Customer Churn Prediction - A study of customer characteristics and the effect of external and macroeconomic factors

Aug 5, 2026

Upload by Riley

This master’s thesis studies customer churn using explainable machine learning on bank loan data combined with external and macroeconomic variables. Logistic Regression and XGBoost are used for churn prediction, while SHAP and LASSO support variable analysis to interpret model drivers. Results show that loan customer age and credit risk (PD) are key explanatory factors, along with LTV, repayment plan, customer duration, and boligkreditt loan balance. Housing price index and policy rate also matter, especially for higher-valued customers. Predictive performance is weak, suggesting the need for more data and better class balance, plus improved data quality and richer qualitative inputs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages75

Reading Time3 h 9 min

AccessFree to download

Table of Contents

List of Figures

List of Tables

1 Introduction

1.1 Research questions

1.2 Structure of the thesis

2 Literature

2.1 Customer churn and factors

2.2 External and macroeconomic factors

3 Data

3.1 Data description

Master’s thesis NT NU Norwegian University of Science and Technology Faculty of Economics and Management NT NU Business School Jesper Ødegård Marcus André Glover Meek Explainable Machine Learning for Customer Churn Prediction A study of customer characteristics and the effect of external and macroeconomic factors Master’s thesis in Economics and Business Administration Supervisor: Ranik Raaen Wahlstrøm May 2023 Jesper Ødegård Marcus André Glover Meek Explainable Machine Learning for Customer Churn Prediction A study of customer characteristics and the effect of external and macroeconomic factors Master’s thesis in Economics and Business Administration Supervisor: Ranik Raaen Wahlstrøm May 2023 Norwegian University of Science and Technology Faculty of Economics and Management NTNU Business School Preface This thesis is created as a part of our 2-year master’s at NTNU Handelshøyskolen. There have been many challenges throughout, and we look forward to even more challenges ahead. We thank our collaborative bank for access to the data and the opportunity to work on and create this exciting thesis. We would also like to extend a great thanks to our supervisor Ranik Raaen Wahlstrøm at the Faculty of Economics and Management at NTNU Business School, for his efficiency, valuable insights and great suggestions along the way. The content of this assignment is the responsibility of the authors. Abstract This thesis is delimited to the financial data of customers in a bank with the help of a collaborative bank and external and macroeconomic data. The thesis investigates what characteristics in loan customers can influence the likelihood of their churn, with the help of prediction through the machine learning (ML) methods; Logistic Regression and XGBoost. The variable analysis of SHAP and LASSO are used to better understand the ML models and the importance of variables. The findings of this thesis indicate that the age of the loan customer and credit risk, as indicated by PD, are key factors in explaining customer churn. Additionally, loan-to-value (LTV), repayment plan, customer duration, and the loan balance of”boligkreditt”, significantly influenced the likelihood of customer churn alongside other moderately impactful variables. However, variables that might be expected to have an impact, such as area of residence, size of the households and others showed no significant influence. The characteristics influencing churn in valuable customers differed from those in other customers. For valuable customers, the age of the loan customer had a lesser impact, while income, DTI, and repayment loan balance played a more significant role. The variable analysis confirms that external and macroeconomic factors influence customer churn. Factors such as the housing price index and policy rate held significance, especially for higher-valued customers. Regarding the ability to predict customer churn using ML methods, both XGBoost and LR showed weak predictive results. However, there were indications that churn predictions within banks could be made with an increased amount of data and reduced imbalances. In conclusion, this thesis contributes to understanding the factors influencing loan customers’ churn and the predictive abilities of ML methods on financial banking data. Improvements in data quality, model performance, and incorporation of qualitative data are suggested for future studies to achieve more robust and actionable results. Sammendrag Denne avhandlingen er avgrenset til finansielle data om kunder fra en samarbeidende bank, og eksterne og makroøkonomiske data. Avhandlingen undersøker hvilkeegenskaper hos l˚anekunder som kan p˚avirke sannsynligheten for kundefrafall vedhjelp av følgende maskinlæringsmetoder (ML) for prediksjon; Logistisk Regresjon og XGBoost. Variabel-analysene av SHAP og LASSO brukes for˚a forst˚a resultatene fra ML-modellene og variablenes betydning. Resultatene fra de

FAQs

Which machine learning methods are used to predict customer churn?

The thesis uses Logistic Regression and XGBoost to predict customer churn probabilities.

What factors are identified as key drivers of churn?

Loan customer age and credit risk (PD) are central factors, and churn is also significantly influenced by LTV, repayment plan, customer duration, and boligkreditt loan balance.

How do external and macroeconomic variables affect churn and prediction?

External and macroeconomic factors influence churn, with housing price index and policy rate showing significance, especially for higher-valued customers. However, churn prediction performance with ML methods remains weak, indicating benefits from more data and reduced imbalances.

Aug 5, 2026

Upload by Riley