[Paper Review] Prediction of Customer Churn in Banking Industry

Sina Esmaeilpour Charandabi|arXiv (Cornell University)|Jan 30, 2023

Customer churn and segmentation6 citations

TL;DR

This study evaluates six supervised learning models to predict customer churn in the banking sector using 10,000 European bank customers' demographic and personal data. Artificial Neural Networks (ANN) with a single hidden layer of five neurons outperform other models, demonstrating robustness to class imbalance and outliers, and showing no significant overfitting—making it the optimal choice for churn prediction in this context.

ABSTRACT

With the growing competition in banking industry, banks are required to follow customer retention strategies while they are trying to increase their market share by acquiring new customers. This study compares the performance of six supervised classification techniques to suggest an efficient model to predict customer churn in banking industry, given 10 demographic and personal attributes from 10000 customers of European banks. The effect of feature selection, class imbalance, and outliers will be discussed for ANN and random forest as the two competing models. As shown, unlike random forest, ANN does not reveal any serious concern regarding overfitting and is also robust to noise. Therefore, ANN structure with five nodes in a single hidden layer is recognized as the best performing classifier.

Motivation & Objective

• To identify the most effective machine learning model for predicting customer churn in the banking industry.
• To evaluate the impact of feature selection, class imbalance, and outliers on model performance.
• To compare the robustness and generalization ability of artificial neural networks (ANN) and random forest in churn prediction.
• To determine whether ANN or random forest better handles data noise and class imbalance in a real-world banking dataset.
• To provide a data-driven recommendation for banks seeking to implement effective customer retention strategies.

Proposed method

• Six supervised classification techniques were applied to a dataset of 10,000 European bank customers with 10 demographic and personal attributes.
• Feature selection was applied to reduce dimensionality and improve model efficiency.
• Class imbalance was addressed using techniques such as resampling or class weighting, particularly for models sensitive to imbalanced data.
• Outlier detection and treatment were performed to assess their impact on model generalization.
• Artificial Neural Network (ANN) models were trained with a single hidden layer containing five neurons, optimized for performance.
• Random forest was used as a baseline model for comparison, leveraging ensemble learning to reduce overfitting.

Experimental results

Research questions

• RQ1 Which supervised learning model performs best in predicting customer churn in the banking industry?
• RQ2 How does class imbalance affect the performance of ANN and random forest in churn prediction?
• RQ3 To what extent do outliers impact the generalization of ANN and random forest models?
• RQ4 Does feature selection improve the predictive accuracy of churn prediction models?
• RQ5 Is ANN more robust to data noise and class imbalance than random forest in this context?

Key findings

• The artificial neural network (ANN) with five neurons in a single hidden layer achieved the highest predictive performance among all models tested.
• Unlike random forest, ANN showed no significant overfitting, indicating strong generalization on the validation set.
• ANN demonstrated robustness to data noise and outliers, maintaining stable performance across different preprocessing conditions.
• Class imbalance had a measurable but manageable impact on model performance, with ANN handling it more effectively than random forest.
• Feature selection improved model efficiency and contributed to better performance, particularly in reducing overfitting in complex models.
• The study concludes that ANN is the most suitable model for customer churn prediction in the banking sector due to its resilience and high accuracy.