Predicting Bank Customer Churn Using Machine Learning Classification

and Stacked Ensemble Techniques

Pragyan Sharma (21BAI1486)

I. Abstract

In the banking industry, where keeping clients has a direct impact on profitability and competitive
advantage, customer attrition is a serious problem. This study uses a dataset from Kaggle to
provide a thorough machine learning method for forecasting customer churn. The ability of several
classification models, such as Random Forest, Support Vector Classifier (SVC), Logistic
Regression, and XGBoost, to forecast a customer's likelihood of leaving the bank is assessed. To
increase the level of performance and to experiment with an approach that can combine the best
of all worlds, a stacked ensemble model is also used by combining predictions from several
separate classifiers. Important customer characteristics that affect churn are highlighted using
visualization tools, and model performance is compared using accuracy, precision, and recall
metrics. To improve interpretability, individual predictions are explained using the Local
Interpretable Model-agnostic Explanations (LIME) technique, which explains why some
consumers are more vulnerable than others and how some features like age and number of products
are more important for decision making than others. The outcomes show that the stacked ensemble
method achieves higher prediction accuracy than individual classifiers. This study helps banks
better identify and retain at-risk clients by providing insightful information about the effectiveness
of machine learning techniques in customer churn prediction.

Keywords: customer churn, classification, roc-auc, ensemble model, features

II. Introduction improve loyalty, and identify risk factors that

could lead to customer exit. With
In today’s highly competitive financial advancements in data science and machine
sector, customer retention is crucial for learning (ML), predictive analytics offers a
sustained growth and profitability. Customer powerful means of understanding customer
churn, which refers to the loss of existing behavior and proactively addressing issues
clients or subscribers, is especially costly for that may prompt clients to leave.
banks, as attracting new customers often
incurs higher expenses than retaining current This paper focuses on the application of
ones. Churn prediction has therefore become machine learning techniques to predict
a vital area of interest for banking institutions customer churn in a bank using a dataset
that aim to strengthen customer relationships, sourced from Kaggle. The dataset comprises
multiple features, including customer improvement by mitigating the weaknesses
demographics, financial history, and inherent in each individual model. The
behavioral patterns. By developing a ensemble approach should provide higher
predictive model, the bank can better accuracy and a superior ROC-AUC score,
understand which customers are most likely demonstrating greater efficacy in predicting
to churn, thereby allowing targeted retention churn.
strategies to be implemented more
effectively. Traditional statistical methods A distinguishing feature of this project is the
for churn analysis, while valuable, often lack incorporation of explainable AI techniques,
the capacity to capture complex patterns in specifically the Local Interpretable Model-
data. Machine learning methods such as agnostic Explanations (LIME). While the
Random Forest, Support Vector Classifier black-box nature of ML models often limits
(SVC), Logistic Regression, and XGBoost transparency, LIME offers a way to interpret
offer a robust approach to detecting subtle specific predictions. This technique provides
trends and correlations that might otherwise localized insights by explaining the
be missed. Additionally, the ensemble predictions made for individual customers,
approach—specifically a stacked identifying the specific features or behaviors
ensemble—leverages the strengths of each that contribute to a high churn risk. Such
model type to boost predictive performance, interpretability is invaluable in the banking
combining insights from multiple algorithms context, as it enables customer relationship
into a single, optimized prediction. teams to devise targeted interventions based
This study incorporates various supervised on a clear understanding of why certain
ML algorithms and compares their customers are likely to leave.
performance in terms of accuracy and the The novelty of this project lies in its practical,
Area Under the Curve (AUC) of the Receiver user-friendly implementation. The code
Operating Characteristic (ROC) curve. developed allows bank personnel or end-
Accuracy provides a straightforward metric users to input values for key customer
for assessing the model's correctness, while features directly, simulating the conditions
the ROC-AUC metric gives a comprehensive under which the model was trained. This
view of the model’s sensitivity and input-based system then outputs a prediction
specificity, particularly in imbalanced indicating whether the customer is at risk of
datasets often seen in churn prediction churning. Such interactivity is particularly
scenarios. These performance metrics beneficial for non-technical users, as it
provide clear indicators of how well each facilitates quick risk assessments and
model identifies at-risk customers, allowing supports immediate decision-making without
for a thorough comparison among the the need for complex software or statistical
individual models and the stacked ensemble. expertise. Moreover, the proposed system
The stacked ensemble model, which bridges the gap between complex ML
combines predictions from several base algorithms and accessible, actionable
classifiers, is expected to yield a performance insights.
The expected results from this study include III. Related Work
achieving high accuracy and ROC-AUC
scores, which will highlight the effectiveness Ahmad et al [1] addressed customer churn
of the stacked ensemble model over prediction as a critical concern, particularly
individual classifiers. The ensemble for telecom companies due to its direct
approach is anticipated to outperform single impact on revenue. Research focused on
models by capitalizing on their combined developing models for predicting potential
predictive power, thus providing a robust and customer churn by using machine learning on
reliable prediction system for customer large-scale data, with performance measured
churn. Additionally, the insights derived by the Area Under Curve (AUC) metric,
from LIME will underscore the relevance of achieving an AUC of 93.3%. Social Network
specific customer attributes in the churn Analysis (SNA) features were also
prediction process, offering a foundation for incorporated, which significantly enhanced
further strategic planning and targeted model performance. The study used Spark for
customer engagement initiatives. By the end big data processing, working on a
of the project, this churn prediction model comprehensive dataset from SyriaTel, and
aims to provide an advanced yet user-friendly tested models including Decision Trees,
solution that equips banks with actionable Random Forests, GBM, and XGBoost, with
insights, enabling them to intervene XGBoost providing the highest predictive
proactively and retain valuable customers. accuracy.
In summary, this paper examines the utility
of machine learning and ensemble methods Singh et al [2] addressed customer attrition in
for predicting customer churn in a bank banking due to its significant impact on
setting. The work leverages multiple customer retention and growth in a
classification models and a stacked ensemble competitive market. This research aimed to
for performance comparison, focusing on forecast potential customer churn using
both predictive accuracy and model machine learning algorithms and developed a
interpretability through LIME and feature Data Visualization RShiny app for analyzing
importance visualization. With the customer churn trends. Despite achieving
integration of user inputs for model testing, relative success, the study identified areas for
the solution aims to offer an accessible, enhancement, including access to larger and
efficient, and interpretable tool for bank more granular datasets for more precise
managers and customer relationship teams to predictions and recency metrics to detect
identify at-risk customers and take behavioral shifts preceding churn. Future
preemptive action. work could involve automating model
training, adding new data points, and
building specialized churn prediction models
for regional customer segments to increase
accuracy and generalizability.
 the potential of DF models as robust solutions
Banu et al [3] developed an AI-based for CCP in the telecom industry.
Customer Churn Prediction model (AICCP-
TBM) specifically for the Gurung et al [5] analyzed consumer churn in
telecommunications sector to improve early the USA using ML models, specifically
churn detection. The AICCP-TBM model Random Forests and Decision Trees, to
employed a Chaotic Salp Swarm predict churn and identify influential factors.
Optimization-based Feature Selection The study employed a dataset of 3333 clients
(CSSO-FS) for effective feature selection and from the UCI repository, with model
a Fuzzy Rule-based Classifier (FRC) performance evaluated through metrics like
optimized by Quantum Behaved Particle accuracy, precision, recall, F-measure, and
Swarm Optimization (QPSO) to enhance Phi coefficient. Random Forest achieved
churn classification accuracy. Experimental superior results, attaining an accuracy of
results on benchmark datasets demonstrated 96.25% and outperforming Decision Tree in
that their model outperformed other state-of- predictive capability. The study's findings
the-art models, achieving accuracies of suggested that government organizations
97.25%, 97.7%, and 94.33%. The study could leverage such models to understand
suggested future expansions to big data churn drivers better, facilitating targeted
platforms to enhance predictive performance retention strategies that support stable
on large-scale datasets. revenue and growth.

Usman-Hamza et al [4] addressed the Wagh et al [6] focused on telecom customer

challenge of customer churn prediction churn, where they applied machine learning
(CCP) in the telecommunications sector by techniques, specifically Random Forest and
developing intelligent decision forest (DF) Decision Tree classifiers, to predict churn
models, including Logistic Model Tree and analyze the reasons behind customer
(LMT), Random Forest (RF), and Functional attrition. The system leveraged classification
Trees (FT). These models were further models to accurately predict churn, achieving
enhanced using weighted soft voting and high precision, recall, and accuracy (around
weighted stacking methods to improve 99% using Random Forest) on an upsampled
performance in the face of class imbalance. dataset. A survival analysis approach,
Extensive testing on publicly available including the Cox Proportional Hazard model
telecom datasets showed that these DF and survival curves, was also employed to
models, particularly their ensemble variants, provide targeted retention strategies. Future
outperformed traditional machine learning research suggested incorporating deep
classifiers, such as Naive Bayes (NB), k- learning models, such as recurrent neural
Nearest Neighbors (kNN), and Decision networks, to capture complex patterns for
Stumps (DS), in both imbalanced and more precise churn predictions in the telecom
balanced datasets. The results demonstrated sector.
superior predictive accuracy and established
De & Prabu [7] presented a systematic review 85.16%. The study identified key factors like
of machine learning techniques in churn age, account balance, and salary as critical
prediction, analyzing 38 primary studies churn indicators, and underscored the value
from 2018-2021. It highlighted gaps in model of data augmentation in improving predictive
generalizability across industries, accuracy for imbalanced datasets, ultimately
emphasized the need for integrating richer aiding in more effective customer retention
customer interaction data, and explored the strategies.
effectiveness of hybrid sampling strategies
for imbalanced datasets. The paper also Murugesan & Trivedi [10] introduced a Tri-
identified a lack of standardized evaluation Strategy Remora Optimization Algorithm
metrics for churn models, offering key (TS-ROA) combined with Support Vector
insights for advancing research in this Machine (SVM) for improved customer
domain. churn prediction in telecommunications.
Tested on the IBM Telco and CCP 2020
A study by Chang et al [8] investigated datasets, the TS-ROA-based SVM method
ensemble learning models for customer churn achieved high accuracy and outperformed
prediction in the telecommunications sector, models like ChurnNet and Grey Wolf
which faces a high churn rate of 30%. By Optimizer with Ensemble Neural Network
comparing models like Decision Trees, (GWO-ENN). The approach demonstrated
Boosted Trees, and Random Forests, the effectiveness in selecting relevant features
study identified Random Forest as the top and enhancing predictive performance.
performer with 91.66% accuracy. The
integration of explainable AI tools, LIME
and SHAP, enhanced model transparency, IV. Dataset Description
providing actionable insights for targeted
retention strategies. The findings emphasized The dataset for this study, obtained from
the strategic value of ensemble models in Kaggle, is derived from a U.S. bank and
improving consumer loyalty and profitability includes 10,000 customer records. Each
through precise churn forecasting and record represents a unique customer profile
intervention. with attributes relevant to their banking
history, demographic information, product
Arefin et al [9] presented a detailed analysis usage, and activity with the bank. This rich
of customer churn prediction in the banking set of features enables detailed analysis to
sector, comparing traditional machine predict customer churn, defined in this
learning models with deep learning context as the likelihood of a customer
approaches on datasets processed with and exiting or leaving the bank. The dataset
without SMOTE. Findings revealed that comprises 18 features in total, with a mixture
SMOTE-enhanced data significantly of numerical, categorical, and binary
improves model performance, with variables. Below is an outline of the dataset’s
LightGBM achieving the highest accuracy at
key columns and their significance in as either male or female. Gender may
predicting churn. contribute to understanding churn
patterns, as previous studies suggest
that customer engagement and
IV (A) Column Description: preferences may vary across genders.
7. Age: Age is a critical demographic
1. RowNumber: This column serves feature that indicates the customer’s
as an index for each row in the age in years. Younger customers
dataset, from 1 to 10,000, and has no might exhibit different loyalty
bearing on the prediction model. patterns compared to older clients, as
2. CustomerId: Each customer is age can influence banking product
assigned a unique identifier to preferences and customer
distinguish them from others. Like relationship duration.
RowNumber, this feature is not used 8. Tenure: This numerical feature
for prediction but is necessary for represents the number of years a
customer reference. customer has been with the bank. A
3. Surname: This column records the longer tenure could indicate customer
customer’s last name, which is loyalty, potentially reducing the
similarly not relevant to the likelihood of churn.
prediction model and is therefore 9. Balance: The customer’s average
excluded during model training. account balance is provided in this
4. CreditScore: This numerical column. High or low account
feature represents each customer’s balances may influence a customer's
credit score, a crucial factor in relationship with the bank, as
assessing creditworthiness and customers with lower balances may
financial behavior. A higher or lower have fewer financial commitments to
credit score can correlate with a the bank.
customer’s financial stability, 10. NumOfProducts: This feature
influencing their likelihood of staying indicates the number of bank products
with the bank. each customer utilizes. Higher
5. Geography: This categorical product usage could correlate with
feature specifies the customer’s greater customer investment in the
location, with three possible values: bank and potentially lower churn risk.
Spain, France, and Germany. 11. HasCrCard: This binary feature
Geographical location could reflect signifies whether a customer has a
differences in market behavior, local credit card with the bank. Owning a
banking regulations, or cultural credit card could be linked to higher
factors that impact customer loyalty. engagement with the bank’s services.
6. Gender: This binary feature 12. IsActiveMember: This binary
indicates the gender of each customer feature indicates whether a customer
has been active recently. Inactivity of the study, using the other features
may signal a higher churn risk, as to model customer churn behavior.
customers who use fewer services
may be more likely to leave. By utilizing these features, this dataset
13. EstimatedSalary: This numerical provides a robust foundation for
feature represents the estimated understanding customer behavior and
salary of the customer, which may predicting churn in a banking context. The
indirectly affect their financial needs inclusion of demographic, behavioral, and
and banking habits. transactional features enables comprehensive
14. Complain: This binary feature modeling of churn, making this dataset well-
records whether the customer has suited for machine learning applications
filed a complaint recently. Customers aimed at customer retention strategies.
with complaints may be more likely
to churn if their issues remain
unresolved. V. Methodologies Used
15. Satisfaction Score: The
customer’s satisfaction score, rated In this project, we developed a
out of 5, reflects their satisfaction comprehensive machine learning pipeline to
with the bank’s services. Lower predict customer churn using the
scores can be a direct indicator of `bank_churn.csv` dataset. The pipeline
churn risk, as dissatisfied customers follows a structured approach that includes
are more likely to leave. data loading, preprocessing, model training,
16. Card Type: This categorical evaluation, and explainability to ensure both
feature indicates the type of credit high accuracy and interpretability of the
card a customer holds, ranging from results.
Silver to Diamond. Card type might
reflect the customer’s engagement The first step involves loading and exploring
level with the bank’s premium the dataset to understand its structure and
services. identify any missing values. This is followed
17. Point Earned: The cumulative by visualizing key data distributions,
reward points earned by the customer including histograms for numerical features,
over their tenure with the bank. High pie charts for the churn distribution, and
points might reflect higher count plots for categorical variables such as
engagement, potentially lowering ‘Geography’ and ‘Gender’. These
churn risk. exploratory analyses provide insights into the
18. Exited: This is the target variable, data and help guide the preprocessing
indicating whether a customer has left decisions.
the bank (1) or is still with the bank
(0). Predicting this variable is the goal
 prediction. Notably, we excluded the
‘Complain’ feature due to its heavy influence
on the predictions, which led to significant
overfitting. By removing ‘Complain’, we
focused on other features to better assess their
impact on customer churn, allowing the
model to make more generalizable
predictions using machine learning
techniques.

Fig. 1: Histogram depicting distribution of values in

numerical columns

Fig. 3: Correlation between numerical features of

data

Fig. 2: Distribution of target variable in dataset

Various machine learning models are then

initialized, including Random Forest,
For data preprocessing, categorical variables
Logistic Regression, Support Vector
such as 'Geography' and 'Gender' are encoded
Classifier (SVC), and XGBoost. Each model
using `LabelEncoder` to convert them into
is trained and evaluated using metrics like
numerical values. Additionally, the features
ROC-AUC and accuracy. Detailed
are scaled using `StandardScaler` to ensure
performance reports, including classification
that all input features are on the same scale,
reports and confusion matrices, are generated
improving the performance of several
for each model to assess their strengths and
machine learning models. The model training
weaknesses. This initial evaluation helps
utilizes CreditScore, Geography, Gender,
identify the most promising models for
Age, Tenure, Balance, NumOfProducts,
further optimization.
HasCrCard, IsActiveMember,
To enhance model performance, a stacked
EstimatedSalary, Card Type, and Satisfaction
ensemble classifier is created, incorporating
Score as features that are relevant to the
Logistic Regression, SVC, XGBoost, and
Random Forest as base classifiers, with features and the target variable.
Logistic Regression serving as the final Despite being less complex than tree-
estimator. This ensemble method combines based models, it is interpretable and
the predictions of multiple models to produce efficient.
a more robust and accurate prediction. The 3. Support Vector Classifier (SVC):
performance of the stacked classifier is then SVC is a powerful classifier that finds
compared to the individual models to the hyperplane that best separates the
evaluate the benefits of ensemble learning. data into different classes. It is
effective in high-dimensional spaces
and works well with both linear and
non-linear data using kernel methods.
SVC is known for its ability to
generalize well, especially when the
number of features is large.
4. XGBoost: XGBoost is an advanced
boosting algorithm that builds an
ensemble of decision trees in a
sequential manner. It focuses on
correcting the errors made by
Fig. 4: Methodology Flowchart previous trees and is highly effective
for structured/tabular data. XGBoost
is known for its speed, accuracy, and
V (A). Models: ability to handle large datasets with
complex patterns.
1. Random Forest: This ensemble 5. Stacked Classifier: This ensemble
model uses multiple decision trees to model combines the predictions of
improve classification accuracy by multiple base classifiers (Logistic
averaging the predictions of each tree. Regression, SVC, XGBoost, and
It is robust to overfitting and well- Random Forest). The stacked
suited for handling complex datasets classifier leverages the strengths of
with both numerical and categorical individual models, allowing for a
features. Random Forest is more accurate and robust prediction
particularly effective when there are by using a final estimator to aggregate
interactions between features. the base model predictions. Logistic
2. Logistic Regression: A simple yet Regression is used as the final
powerful model for binary estimator in the stacked model to
classification tasks, Logistic optimally combine the predictions of
Regression estimates the probability the four base models. Its simplicity,
of a customer churning based on a linear decision boundary, and ability
linear relationship between input to minimize overfitting make it
 effective in improving overall model to create a robust, accurate, and interpretable
performance and interpretability. customer churn prediction model. Additional
improvements include hyperparameter
tuning, cross-validation for more robust
performance estimation, and further feature
engineering to optimize model accuracy.

VI. Results & Discussion

Fig. 5: Stacked Ensemble Model The goal of this project was to predict
customer churn for a banking institution,
The models' performance is visualized by
identifying customers who are likely to leave.
comparing accuracy and misclassification
We utilized multiple machine learning
rates using bar plots, providing a clear
models to predict churn, including Random
overview of how each model performs.
Forest, Logistic Regression, Support Vector
Additionally, ROC curves for each model are
Classifier (SVC), XGBoost Classifier, and a
plotted, showcasing their ability to
Stacked Ensemble model. Additionally, we
distinguish between churned and non-
applied LIME (Local Interpretable Model-
churned customers. The AUC values of the
Agnostic Explanations) to explain individual
ROC curves are also displayed to highlight
predictions and provide insights into the
the effectiveness of each model in capturing
factors influencing the models' decisions.
true positives. For model explainability,
LIME (Local Interpretable Model-agnostic
The performance metrics for each model and
Explanations) is employed to offer local
the stacked ensemble are summarized below:
explanations for individual predictions from
the test set, enhancing transparency and trust
Random Forest Accuracy: 86.55%
in the model's decisions. Feature importance
Logistic Regression Accuracy: 81.65%
for selected models helps us understand
Support Vector Classifier Accuracy: 86.00%
which column has the most influence on the
XGBoost Classifier Accuracy: 85.60%
model’s prediction. Lastly, an interactive
input interface is provided for users to input
The Stacked Ensemble model, which
their own customer data for churn prediction.
combines the predictions of all the models,
The model predicts whether a customer will
achieved the highest ROC-AUC score of
churn, and partial dependence plots can be
0.8723 and an accuracy of 86.85%.
generated to visualize the impact of specific
Stacked Ensemble ROC-AUC: 0.8723
features on the prediction.
Stacked Ensemble Accuracy: 86.85%

Overall, this methodology combines various

machine learning techniques, including
ensemble learning and model explainability,
The Confusion Matrix for the Stacked important patterns in the data. The top three
Ensemble model shows the following: performing models had very similar
True Positives (TP) = 198 (Predicted as performance as observed.
churn, actually churned) The Support Vector Classifier had an
False Positives (FP) = 68 (Predicted as churn, accuracy of 86.00%, very close to Random
but stayed) Forest and XGBoost, but the ROC-AUC
True Negatives (TN) = 1539 (Predicted as score of 0.8197 was lower than the others,
staying, actually stayed) indicating it struggled more with
False Negatives (FN) = 195 (Predicted as distinguishing between the churn and non-
staying, but churned) churn classes.
Logistic Regression performed the worst,
with an accuracy of 81.65% and ROC-AUC
of 0.7636. This suggests that the linear
assumptions of logistic regression were not
well suited for the complex, non-linear
relationships in the dataset.

Fig. 6: ROC Curves for models used

VI (A). Model Comparison:

Fig. 7: Comparison of model accuracies
The Stacked Ensemble outperformed all
individual models based on the ROC-AUC
score and accuracy. The ensemble approach
likely benefited from the diversity of the base
models, combining their strengths. The
second-best model was the XGBoost
Classifier, which had a very similar ROC-
AUC score and accuracy to the stacked
ensemble, indicating its robustness in
handling imbalanced datasets and capturing
complex relationships. The Random Forest
model was slightly behind in performance
but still performed well with an accuracy of Fig. 8: Comparison of number of records in test set
86.25%, suggesting that it also captured misclassified by the models
 bank, such as identifying at-risk customers
and improving customer retention strategies.
Model ROC- Accuracy No. of
AUC (in %) correct Analyzing feature importance helps us
classific understand how each feature contributes to
ations model predictions. For Random Forest, the
most influential feature is Age with an
Stacked 0.8723 86.65 1737
importance score of 0.313, followed by
Ensemble
NumOfProducts (0.227) and Balance
XGBoost 0.8719 86.75 1735 (0.094). This suggests that age is a significant
Classifier factor in predicting customer churn,
indicating that younger or older customers
Random 0.8648 86.25 1725
Forest might have distinct patterns associated with
Classifier churn likelihood. The number of products a
customer holds also impacts the prediction,
Support 0.8197 86 1720 possibly because customers with multiple
Vector
products may have a stronger relationship
Classifier
with the bank, affecting their retention.
Logistic 0.7636 81.65 1633 In contrast, XGBoost identifies
Regression NumOfProducts as the most important
feature (0.243), followed closely by
Table 1: Model Comparison IsActiveMember (0.236) and Age (0.208).
This shift in feature priority suggests that
XGBoost places greater weight on active
VI (B). Explainability using Local engagement and the variety of products held
Interpretable Model-Agnostic Explanations by a customer as churn indicators. This
(LIME) and Feature Importance: model’s emphasis on engagement through
IsActiveMember indicates that actively
LIME was used to explain the individual engaged customers are less likely to churn,
predictions made by the models, especially underscoring the relevance of customer
the stacked ensemble, providing transparency relationship management.
into how each model arrived at its decision
for specific customers. We performed
visualization for the XGBoost model as well
as the stacked model. By using LIME, we
were able to identify which features
contributed the most to each individual
prediction. This interpretability is valuable
for both understanding the model's behavior
and for providing actionable insights to the
Fig. 9: Interpreting XGBoost using LIME for first
prediction in test set
Fig. 10: Interpreting Stacked model using LIME for
first prediction in test set

The differences in feature importance across

the models reveal subtle differences in how Fig.11: Output for given User Input and Feature
they interpret the data and make predictions. Importance plot
Random Forest relies more heavily on
demographic factors like age, while
VII. Conclusion
XGBoost focuses more on behavioral factors
such as engagement and the range of
This paper demonstrates the development
products. This distinction may explain why
and implementation of a machine learning
the Stacked Ensemble model, which
model for predicting customer churn in the
combines both approaches, achieves higher
banking sector using a comprehensive
accuracy and a balanced interpretation of
dataset. Through a systematic comparison of
demographic and behavioral factors. These
algorithms, including Random Forest,
insights can guide the bank's retention
Logistic Regression, Support Vector
strategies by highlighting the importance of
Classifier, and XGBoost, we identified
maintaining product diversity and fostering
distinct patterns in customer behavior and
customer engagement, particularly for age
demographic factors that influence churn.
demographics more prone to churn.
The final Stacked Ensemble model,
LIME also explained that for specific
combining these algorithms, achieved the
instances, the ‘number of products’ was one
highest accuracy (86.85%) and ROC-AUC
of the most significant features influencing
score (0.8723), outperforming individual
predictions of churn. Similarly, ‘age’,
models by leveraging their diverse strengths.
‘balance’, and ‘active user’ also emerged as
Feature importance analysis revealed that
important features for the decision-making
demographic features like age, as
process, which was in line with domain
emphasized by Random Forest, and
knowledge regarding customer engagement.
behavioral features such as the number of
products held and active membership status,
as prioritized by XGBoost, are influential in
churn prediction. These insights underscore
the importance of customer engagement and Eklas Hossain, Investigating customer churn
product diversification in reducing churn in banking: a machine learning approach and
risk. The combination of demographic and visualization app for data science and
behavioral predictors in the ensemble management, Data Science and
approach offers a more holistic view, Management,
balancing the unique insights from each Volume 7, Issue 1, 2024, Pages 7-16, ISSN
model to improve prediction accuracy. 2666-7649,
[Link]

VIII. Future Scope [3] Faritha Banu, J., Neelakandan, S., Geetha,
B.T, Selvalakshmi, V., Umadevi, A.,
The study's results suggest actionable Martinson, Eric Ofori, Artificial Intelligence
strategies for churn mitigation, emphasizing Based Customer Churn Prediction Model for
the value of customer relationship Business Markets, Computational
management. In future work, the model can Intelligence and Neuroscience, 2022,
be refined by exploring deeper feature 1703696, 14 pages, 2022.
engineering, advanced hyperparameter [Link]
tuning, and real-time data for more
responsive predictions. Additionally, [4] Usman-Hamza FE, Balogun AO, Capretz
incorporating Explainable AI methods, such LF, Mojeed HA, Mahamad S, Salihu SA,
as SHAP and LIME, in real-world Akintola AG, Basri S, Amosa RT, Salahdeen
applications can further enhance trust in NK. Intelligent Decision Forest Models for
model predictions, providing actionable Customer Churn Prediction. Applied
insights for decision-makers. This study Sciences. 2022; 12(16):8270.
highlights the potential of ensemble learning [Link]
in financial analytics and offers a foundation
for future research to develop adaptive, data- [5] Nisha Gurung, Md Rokibul Hasan, Md
driven strategies for customer retention. Sumon Gazi, & Faiaz Rahat Chowdhury.
(2024). AI-Based Customer Churn Prediction
Model for Business Markets in the USA:
IX. References Exploring the Use of AI and Machine
Learning Technologies in Preventing
[1] Ahmad, A.K., Jafar, A. & Aljoumaa, K. Customer Churn. Journal of Computer
Customer churn prediction in telecom using Science and Technology Studies, 6(2), 19–29.
machine learning in big data platform. J Big [Link]
Data 6, 28 (2019).
[Link] [6] Sharmila K. Wagh, Aishwarya A.
Andhale, Kishor S. Wagh, Jayshree R.
[2] Pahul Preet Singh, Fahim Islam Anik, Pansare, Sarita P. Ambadekar, S.H.
Rahul Senapati, Arnav Sinha, Nazmus Sakib, Gawande, Customer churn prediction in
telecom sector using machine learning Communication Systems (ICICACS),
techniques, Results in Control and Raichur, India, 2024, pp. 1-7, doi:
Optimization, 10.1109/ICICACS60521.2024.10498700.
Volume 14, 2024, 100342, ISSN 2666-7207,
[Link]

[7] De, S., & Prabu, P. (2022). Predicting

customer churn: A systematic literature
review. Journal of Discrete Mathematical
Sciences and Cryptography, 25(7), 1965–
1985.
[Link]
238

[8] Chang V, Hall K, Xu QA, Amao FO,

Ganatra MA, Benson V. Prediction of
Customer Churn Behavior in the
Telecommunication Industry Using Machine
Learning Models. Algorithms. 2024;
17(6):231.
[Link]

[9] Arefin, S., Parvez, R., Ahmed, T., Jahin,

F., Sumaiya, F., Ahsan, M. (2025).
Enhancing Bank Consumer Retention: A
Comprehensive Analysis of Churn Prediction
Using Machine-Learning and Deep Learning
Techniques. In: Goar, V., Kuri, M., Kumar,
R., Senjyu, T. (eds) Advances in Information
Communication Technology and Computing.
AICTC 2024. Lecture Notes in Networks and
Systems, vol 1075. Springer, Singapore.
[Link]
7_15

[10] P. Murugesan and P. Trivedi, "Tri-

Strategy Remora Optimization Algorithm
based Support Vector Machine for Customer
Churn Prediction," 2024 International
Conference on Integrated Circuits and