Applying Data Science and Machine

Learning to Segment Customers and

Optimize Bank Term Deposit
Subscriptions

1 Introduction
In recent years, the banking industry has faced increasing competition and rapidly changing
customer behavior. One of the main challenges for banks is to identify customers who are
likely to deposit money through marketing campaigns. Deposits play a crucial role in a
bank’s financial stability, as they provide a primary source of funds for lending and
investment activities.

Traditionally, banks rely on manual assessment or simple customer segmentation based on

experience and basic rules when conducting deposit marketing campaigns. However, these
approaches are often inefficient, time-consuming, and may lead to inaccurate targeting
decisions. As a result, banks may invest significant resources in contacting customers who
have a low intention of subscribing to a deposit product, while overlooking potential
customers with higher interest.

Because of the rapid development of digital banking systems, banks are now able to collect
large volumes of customer data, including demographic information, financial status and
historical interaction records. Therefore, it creates an opportunity to apply data-driven
approaches to better understand customer behavior and improve decision-making
processes.

In the banking field, machine learning can be applied to predict whether a customer will
subscribe to a term deposit based on historical marketing and customer information. These
predictions can help banks optimize marketing strategies, improve customer targeting and
enhance campaign effectiveness.

Therefore, this study focuses on applying machine learning techniques to evaluate and
predict the likelihood of customers subscribing to a term deposit using a real-world bank
marketing dataset.

1.1 Problem formulation​

Despite the availability of detailed customer and marketing data, banks still face difficulties in
accurately predicting customer responses to deposit marketing campaigns. Customer
decisions to subscribe to a term deposit depend on numerous factors, such as demographic
characteristics, financial conditions, loan status and previous marketing interactions. These
factors interact in complex and non-linear ways, making it difficult to evaluate customer
potential using traditional rule-based approaches.

In practice, banks often apply broad marketing strategies without effectively distinguishing
between customers who are likely to subscribe to a deposit and those who are not. This can
lead to inefficient use of marketing resources, increased operational costs and lower
campaign success rates. Moreover, inaccurate prediction of customer behavior may
negatively impact customer satisfaction and long-term relationships.

Therefore, the problem addressed in this study is the need for a data-driven and systematic
approach that can analyze customer information and predict the likelihood of customers
subscribing to a term deposit. Machine learning models provide a solution by learning
patterns from historical marketing data and generating predictions to support more accurate
and objective decision-making.

1.2 Research Questions / Hypotheses

Based on the problem formulation and the characteristics of the bank marketing dataset, this
study aims to address the following research questions:

●​ RQ1: What are the key factors that influence customers’ decisions to subscribe to a
term deposit?​

●​ RQ2:Which customer segments are most likely to subscribe to a term deposit, and
what are their typical characteristics?

These research questions guide the design of the methodology, experimental setup and
result analysis in this study.

2 Related works
1. A data-driven approach to predict the success of bank telemarketing (Moro et al.
2014). [Link]

Objective: Propose an effective data mining approach to forecast the success of long-term
deposit telemarketing campaigns. The authors aim to support bank managers in reducing
wasted resources on ineffective calls and minimizing customer annoyance, particularly in the
context of the economic recession in Portugal at that time.

Methods: Four machine learning models representing different analytical approaches are
used: Logistic Regression, Decision Trees, Artificial Neural Networks and Support Vector
Machines. The main evaluation metrics include the Area Under the ROC Curve (AUC) and
the Area under the LIFT curve (ALIFT).

Results: The Neural Network (NN) model demonstrates superior performance compared to
the other models, achieving an AUC of 0.8 and an ALIFT of 0.7. Analysis of the LIFT curve
shows that if the bank focuses its calls on only the top 50% of customers ranked by the
model, it can reach up to 79% of successful subscribers. This means that the bank can
reduce calling efforts by half while retaining nearly 80% of the total revenue.

Limitations: At the time of the study (2014), data imbalance handling techniques such as
SMOTE were not yet widely applied. Moreover, the dataset is limited to a single bank in one
country, which may restrict the generalizability of the findings

2. Jiang et al. (2025): Term Deposit Prediction using Logistic Regression

[Link]

Objective: Jiang et al. (2025) aim to predict whether a customer will subscribe to a term
deposit using bank marketing data, framing the problem as a binary classification task to
support more effective customer targeting.

Method: The study uses the Portuguese Bank Marketing dataset from the UCI Machine
Learning Repository and applies Logistic Regression as a baseline model. Feature selection
based on correlation analysis is performed to remove irrelevant variables and improve model
interpretability.

Result: Results show that Logistic Regression achieves stable and improved performance
after feature selection and is less prone to overfitting than Decision Tree models, confirming
its effectiveness as a baseline for term deposit prediction.

Limitations: Rely on Logistic Regression as a baseline model, which may not capture
complex non-linear relationships in customer behavior. In addition, class imbalance in term
deposit subscription is not explicitly addressed, potentially affecting prediction performance.

3. Applying machine learning to the development of prediction models for bank

deposit subscription (Hou et al., 2022). [Link]

Objective: This study aims to address the difficulty banks face in identifying target
customers for term deposit products. The main objective is to build and compare the
performance of five supervised machine learning models to predict whether a customer will
subscribe to a term deposit.

Methods: The authors use the Bank Marketing dataset from the UCI Machine Learning
Repository. Five algorithms are applied: Naive Bayes, Decision Tree, Random Forest,
Support Vector Machine and Neural Network.

Results: Random Forest achieves the highest accuracy (91.89%), followed by Decision
Tree (91.79%) and SVM (91.72%). The Neural Network model shows the highest sensitivity
(99.99%) but lower accuracy (88.86%) because it overpredicts the “no subscription”
outcome.
Limitations: the study does not apply attribute selection techniques to further optimize
model performance.

4. Machine Learning Performance on Predicting Banking Term Deposit (Tuan, 2022).

[Link]

Objective: Tuan (2022) studies the prediction of customer subscription to banking term
deposits in order to improve marketing effectiveness, focusing on advanced deep learning
approaches for the same business problem addressed in this study.

Method: The research uses the Bank Marketing dataset from Kaggle, which is similar in
structure to the UCI dataset and applies several deep learning models, including
SimpleRNN, LSTM, GRU, BiLSTM and BiGRU, to model complex relationships in customer
data.

Results: The results show that deep learning models achieve high accuracy, with GRU
performing best (around 90.8%), followed by BiLSTM and BiGRU, indicating the potential of
deep learning for term deposit prediction.

Limitations: The study primarily evaluates model performance using accuracy, without
detailed analysis of precision, recall, or class imbalance, which are critical in banking
applications. In addition, the deep learning models provide limited interpretability, making it
difficult to explain predictions to banking practitioners. Finally, the dataset is tabular rather
than truly sequential, which raises questions about the suitability of RNN-based models such
as LSTM and GRU for this problem.

5. Interpretable Data Mining Approaches to Predict Term Deposits Subscriptions —

Chen (2023) [Link]

Objective: The study aims to predict customers’ subscription to bank term deposits and
improve model interpretability. It focuses on identifying key customer features that influence
subscription decisions and avoids using unrealistic predictors such as call duration.

Methods: The author uses telemarketing data from a Portuguese bank (UCI dataset) and
applies two interpretable machine learning models: Decision Tree and Random Forest.
Categorical variables are converted using one-hot encoding, the data is split 80:20 for
training and testing and the variable “duration” is excluded from predictive models. Model
performance is evaluated using confusion matrices and AUC-ROC scores, along with
importance analysis.

Results: Random Forest outperforms Decision Tree, achieving an AUC-ROC of about

73.7% without using duration. Although accuracy decreases compared to models that
include duration, the results are more realistic for real-world prediction. Feature importance
analysis shows that balance and age are the most influential factors in term deposit
subscriptions.

Limitations: The predictive performance is lower than previous studies because a highly
correlated variable, which is duration, is removed. The study also relies on only two
tree-based models and does not explore more advanced algorithms such as deep learning.
3 Background
The Importance of Data Analytics in Bank
In the era of rapid digital transformation, financial institutions are increasingly
relying on data analytics to optimize business operations and
decision-making. One of the most critical areas is the segmentation ,
particularly term deposits. However, the major challenge today is the high cost
of campaign execution coupled with low response rates, primarily due to the
lack of synchronized data exploitation.
Targeting the wrong audience not only wastes resources but also negatively
impacts brand image. For instance, a customer who currently has an active
loan at Techcombank (TCB) and is under the pressure of monthly interest
payments may still repeatedly see Facebook ads from TCB inviting them to
open a savings account. This "mismatched" marketing approach
demonstrates a lack of customer insight:
●​ Wrong Audience: The customer needs debt solutions, not investment
products for idle cash.
●​ Poor Timing: This creates annoyance and leads the customer to
perceive the bank as unprofessional in data management.
●​ Wasted Costs: The bank is paying for ad impressions that yield zero
value.

Consequently, building predictive models to identify the right customer

segments with genuine potential and demand has become an urgent
requirement. This helps banks transition from mass marketing to personalized
marketing, ultimately optimizing budgets and enhancing brand credibility in the
eyes of customers.

Related Business Operations

This research is closely linked to several core operations in banking and data analytics,
including:

1. Customer Relationship Management (CRM)

CRM goes beyond simple data collection; it is the strategic integration and analysis of
customer information to deeply understand behaviors and needs.
●​ Omnichannel Integration: CRM integrates data from multiple touchpoints (branch
transactions, website visits, phone calls, and emails) to create a comprehensive
360-degree customer profile.
●​ Advanced Segmentation: It helps categorize customers based on their lifetime value,
behavior, and potential, enabling the design of highly personalized strategies.
 ●​ Customer Scoring: CRM systems involve scoring models to predict engagement
probability, propensity to buy, or credit risk.

1.1. Case Study: CRM at Techcombank

Techcombank implements a modern CRM system (powered by platforms like Salesforce) to
unify customer data across various channels:
●​ Branch Transactions: Tracking account balances, deposit/withdrawal history, and
existing loan portfolios.
●​ Digital Banking & Website: Monitoring login frequency, navigation patterns, and
online product interests to deliver timely and relevant financial offers.
●​ Call Center and Email Marketing:Recording customer feedback and marketing
call history.

Through omnichannel data integration, Techcombank builds a comprehensive 360-degree

customer profile to:
●​ Categorize customers by value, behavior, and potential: For example, customers
with a high average balance who have not yet opened a term deposit but interact
positively with previous campaigns are classified as "high-potential customers."
●​ Design personalized strategies: The bank delivers messages, offers, or product
recommendations tailored to each customer segment, enhancing campaign
effectiveness.
●​ Customer Scoring: Utilizing historical and behavioral data to predict the likelihood of
a customer signing up for a new product or their credit risk. For instance, a customer
with positive interaction history but a large existing loan may be assigned a specific
risk score, allowing relationship managers to provide appropriate consultation.

Practical Results:
Through CRM, Techcombank has increased campaign response rates, improved customer
retention, and reduced communication costs with low-potential leads.
●​ Increased lead distribution to sales teams: According to the 2024 Annual Report,
by combining CRM with data analytics and AI, Techcombank increased the number
of leads allocated to the sales force by 45% year-over-year. CRM enables a deeper
understanding of customers and ensures the right leads are assigned to the right
[Link]://[Link]/FileStore2/File/2025/04/09/20250409_TCB_250409_Ann
ual_Report_2024.pdf?utm_source=[Link]
●​
●​ Improved conversion rates: The report also indicates that CRM and data insights
helped the bank achieve a conversion rate of approximately 29.3%, representing a
~41.5% year-over-year increase. This reflects that CRM does not just collect data but
effectively leverages it for sales and marketing
[Link]://[Link]/FileStore2/File/2025/04/09/20250409_TCB_25040
9_Annual_Report_2024.pdf?utm_source=[Link]

2. Direct Marketing – From "Mass" to "Targeted"

Telemarketing is no longer about making random calls from a list. In the
digital era, it has been transformed through technology:
●​ Resource Optimization: Instead of employees making hundreds of
calls a day with low success rates, the system only pushes
"Leads" with high priority scores. This reduces operating
expenses (OPEX) for Contact Centers.
●​ Personalized Scripts: Based on CRM data, staff do not just offer a
generic product. They know if a customer has a recently matured
savings account or a large amount of idle cash to provide a
tailored consultation script.
●​ Reducing Intrusiveness: Contacting the right person at the right
time reduces customer annoyance, protects brand reputation, and
maintains long-term relationships instead of being perceived as
"phone spam."

Based on your key points, combined with insights from academic

research (such as the study by Moro et al.) and practices at banks like
Techcombank, here is an in-depth analysis of these three pillars in
modern banking strategy:

3. Business Decision Support – Transforming Data into Value

The ultimate goal of analysis and prediction is not just dry numbers, but
strategic decisions:
●​ Marketing Budget Allocation: Banks will concentrate their budgets
on segments with the highest conversion rates (such as Cluster 2
in the research) and reduce costs in low-potential segments
(Cluster 1 – young individuals with housing or car loans).
●​ Utilizing LIFT
○​ LIFT (Area of LIFT): Helps determine what percentage of total
expected registrations can be achieved by reaching only the
top 20-30% of customers. For example: Reaching the top
50% of customers could yield 79% of total deposits.
●​ Cross-selling Strategy: For groups with no demand for savings
(due to high debt), the bank can pivot to introducing financial
support products or debt restructuring instead of persistently
offering term deposits.

Các thuật toán sử dụng (ngoài phạm vi bài học)

Bên cạnh các phương pháp truyền thống, nghiên cứu còn áp dụng một số thuật toán học
máy nâng cao chưa được đề cập trong nội dung bài học nhằm cải thiện độ chính xác dự
đoán và khả năng xử lý dữ liệu lớn. Các thuật toán này có khả năng:

●​ Khai thác mối quan hệ phi tuyến giữa các biến đầu vào
●​ Xử lý tốt dữ liệu có tính đa dạng và mất cân bằng
●​ Nâng cao hiệu suất phân loại so với các mô hình thống kê cơ bản

Việc sử dụng các thuật toán ngoài chương trình học không chỉ giúp mở rộng phạm vi nghiên
cứu mà còn cho thấy tiềm năng ứng dụng của khoa học dữ liệu trong các bài toán nghiệp
vụ thực tế của ngành ngân hàng.

Dựa trên những ý chính bạn đã nêu, kết hợp với các kiến thức từ tài liệu chuyên ngành
(như nghiên cứu của Moro et al.) và thực tế tại các ngân hàng như Techcombank, dưới đây
là phân tích chuyên sâu về ba trụ cột này trong chiến lược ngân hàng hiện đại:

4 Methodology
4.1 Method
This study adopts a quantitative research approach based on the CRISP-DM framework.
The analysis follows six phases: business understanding, data understanding, data
preparation, modeling, evaluation, and deployment. All data mining tasks were conducted
using Orange Data Mining.

Figure 1. CRISP-DM model

Link of image: [Link]

4.2 . Data
The Bank Marketing dataset was obtained from the UCI Machine Learning Repository. It
contains records of direct marketing campaigns conducted via telephone by a Portuguese
bank between 2008 and 2010. The objective is to predict whether a client subscribes to a
term deposit (target variable y). The dataset includes 45,211 observations with 16 predictor
variables and one binary target. There are no null (NA) values; however, several categorical
variables include an “unknown” category. Table 1 summarises the variables used in this
study. Table 2 reports descriptive statistics for key numerical variables, while Figures 3 and 4
illustrate selected categorical distributions, including the imbalanced subscription outcome.

Figure 2. Sample of the Bank Marketing Dataset

Table 1 presents the description of the variables used in this study.

Type Description
Variable Name

age Integer Age of the client

job Categorical Type of job

marital Categorical Marital status

education Categorical Level of education

default Binary Has credit in default (yes/no)

balance Integer Average yearly account balance

 housing Binary Has housing loan (yes/no)

loan Binary Has personal loan (yes/no)

contact Categorical Contact communication type

day_of_week Categorical Day of the week of the last contact

month Categorical Month of the last contact

duration Integer Duration of the last contact in seconds

campaign Integer Number of contacts performed during the current

campaign

pdays Integer Number of days since the client was last contacted
from a previous campaign (-1 means the client was
not previously contacted)

previous Integer Number of contacts performed before the current

campaign

poutcome Categorical Outcome of the previous marketing campaign

y Binary Indicates whether the client subscribed to a term

deposit (yes/no)

Table 1. Description of Variables

Table 2 presents the descriptive statistics of the main numeric variables,

while Figures 1 and 2 illustrate the frequency distributions of selected
categorical variables.

Feature Mean Median Mode Standard deviation Min Max Missing

age 40.93621 39 32 10.61876204 18 95 0

balance 1362.272 448 0 3044.765829 -8019 1021 0
27

duration 258.1631 180 124 257.5278123 0 4918 0

campaign 2.763841 2 1 3.098020883 1 63 0

pdays 40.19783 -1 -1 100.128746 -1 871 0

previous 0.580323 0 0 2.303441045 0 275 0

Table 2. Descriptive Statistics of Numeric Variables

Figure 3. Frequency Distribution of Martial status

 Figure 4. Subscription Outcome Distribution

5 Experiment

5.1 Experimental Setup

Figure 5. Experiment on Orange (implementation)
 1.​ Experiment 1: Factors Influencing Term Deposit Subscription

First, the dataset was prepared for modelling. Categorical variables were encoded and
numerical variables were standardised using z-score normalisation (μ = 0, σ = 1) to ensure
consistent feature scales. Potential outliers were inspected using the Outliers widget to
reduce the influence of extreme values on model training.

Next, the dataset was split into training and testing sets using a hold-out approach with an
80:20 ratio (Data Sampler). Three classification models were trained and compared: Logistic
Regression, Neural Network, and Decision Tree. Model performance was evaluated on the
test set using multiple metrics to account for the imbalanced target distribution.

To answer the first research question on influencing factors, Logistic Regression was
selected as the primary interpretable model. Feature effects were interpreted using a
nomogram (log-odds scale), where positive contributions indicate an increased likelihood of
subscribing to a term deposit and negative contributions indicate a decreased likelihood.

2.​ Experiment 2: Customer Segmentation

To address the second research question, clustering was applied to group customers with
similar characteristics. Before clustering, categorical variables were encoded and numerical
features were standardised. The target variable (y) was excluded from the clustering process
and used only after clustering to examine subscription tendencies across segments.

K-means clustering was used to segment customers into distinct clusters. The clustering
results were then analysed by comparing subscription rates across clusters and profiling
each cluster using key attributes such as age, balance, and loan indicators.

5.2 Evaluation criteria

Experiment 1: Classification (Term Deposit
Subscription Prediction)
Experiment 1: Classification Evaluation Metrics

Because the subscription outcome is imbalanced (far fewer “yes” cases than “no”), multiple
evaluation metrics were used instead of relying only on accuracy:

●​ Accuracy measures the overall proportion of correct predictions.​

●​ Precision measures how many predicted subscribers actually subscribed, which is

important to reduce false positives.​

●​ Recall measures how well the model identifies actual subscribers, which is important
to avoid missing potential customers.​

●​ F1-score is the harmonic mean of precision and recall, providing a balanced view
under class imbalance.​

●​ AUC (Area Under the ROC Curve) evaluates the model’s ability to distinguish
between subscribers and non-subscribers across thresholds.​

●​ Matthews Correlation Coefficient (MCC) provides a robust overall measure for

binary classification and is particularly suitable for imbalanced datasets.

Experiment 2: Clustering (Customer Segmentation)

To assess the quality and usefulness of the customer segmentation, the following criteria
were applied:

●​ Silhouette Score evaluates how well-separated the clusters are; higher values
indicate better-defined clusters.​

●​ Within-Cluster Sum of Squares reflects cluster compactness; lower variation

suggests customers within a cluster are more similar.

6 Results and Discussion

6.1 Các kết quả
 1.​Experiment 1: Classification (Term Deposit
Subscription Prediction)
1. Test and score

Figure 6. Model Evaluation Metrics

Using 5-fold cross-validation, Logistic Regression and Neural Network achieve very similar
overall performance (AUC 0.908 vs 0.904; F1 ≈ 0.900), while the Decision Tree performs
substantially worse (AUC 0.585). Therefore, the final model selection depends on the
trade-off between detection performance and interpretability

2.​ Confusion matrix

Figure 7. Confusion matrix of Neutral Network

 Figure 8. Confusion matrix of Decision Tree

Figure 9. Confusion matrix of Logistic Regression

At the default threshold (0.5), the Neural Network identifies the most subscribers (TP = 1646)
and has the lowest false negatives (FN = 1777), which is suitable when the objective is to
capture as many potential subscribers as possible. In contrast, Logistic Regression produces
the fewest false positives (FP = 653), reducing unnecessary targeting, and remains easier to
interpret. Given the research goal of identifying key influencing factors, Logistic Regression
is selected for further interpretation.

3.​ ROC analysis

 Figure 10. ROC analysis

The ROC curves confirm that Logistic Regression and Neural Network have comparable
discrimination ability, consistent with their similar AUC values. The Decision Tree curve lies
closer to the diagonal, indicating weak classification ability.

4.​ Nomogram

Figure 11. Factors affecting customers’ decisions to subscribe to a term deposit

The nomogram highlights duration, previous contacts, campaign intensity, previous

campaign outcome, and month as the most influential predictors. Duration shows the
strongest association with subscription; however, as it is observed during the call, it mainly
reflects interaction strength rather than a customer attribute known before contacting

Experiment 2: Customer segmentation

Figure 11. Term Deposit Subscription by Cluster

Figure 12. Average Customer Age

Default status by Customer Segmentation

The variable default shows no variation in this sample (all ‘no’), so it does not help
differentiate clusters.

Figure 13. Loan Status by Customer Segment

Housing by customer segmentation

 FIgure 14. Balance Profile by Customer Segment

Figure 15. Mean Previous Contacts by Customer Segment

6.2 Trả lời câu hỏi nghiên cứu
1.​ What are the key factors that influence customers’ decisions to subscribe to a
term deposit?

Based on the logistic regression nomogram (target class = 1, “subscribed”), the five most
influential variables are duration, previous, campaign, poutcome, and month. The results
suggest that customers’ subscription decisions are driven largely by interaction intensity and
campaign history, rather than purely demographic factors.

1) Duration (call duration in seconds).​

Duration is the strongest predictor in the model. Longer calls are associated with a higher
likelihood of subscribing to a term deposit, which likely reflects stronger engagement during
the conversation (e.g., the customer asks questions, requests clarification, or is actively
considering the offer). However, it is important to note that duration is recorded after the call
begins, so it captures the outcome of interaction quality rather than a customer characteristic
known in advance.

2) Previous (number of contacts before the current campaign).​

This variable represents prior outreach history. Customers who have been contacted
previously tend to show a different propensity to subscribe compared with those who have
not. In practice, this can be interpreted as an indicator of relationship familiarity and prior
exposure to the product, which may increase readiness to subscribe. At the same time, it
may also reflect repeated past attempts, so it should be discussed as a contact-history factor
rather than a purely positive driver.

3) Campaign (number of contacts during the current campaign).​

Campaign measures how many times the bank contacted the customer within the ongoing
marketing campaign. The model indicates that higher campaign intensity is generally
associated with a lower likelihood of subscription. A reasonable interpretation is that
customers who require many repeated contacts are less interested or may perceive the
repeated outreach as intrusive, making them less likely to convert.

4) Poutcome (outcome of the previous campaign).​

Poutcome captures whether the previous marketing campaign was successful,
unsuccessful, or unknown for a given customer. This factor is highly informative: customers
with a positive previous outcome are more likely to subscribe again, while unsuccessful or
unknown outcomes are typically linked to lower likelihood. This aligns with marketing
intuition that past positive responses signal higher affinity toward the product.
5) Month (month of last contact).​
Month reflects timing effects and potential seasonality in customer behavior and campaign
execution. The model suggests that subscription likelihood varies across months, implying
that customers may be more receptive at certain times of the year (e.g., budget planning
periods, end-of-year saving behavior, or months when promotions are stronger). Therefore,
month acts as a proxy for seasonal demand and campaign timing.

Conclusion ​
In summary, the logistic regression results indicate that term deposit subscription is most
strongly associated with call engagement (duration) and marketing contact dynamics
(previous, campaign, and poutcome), with seasonality (month) also playing an important
role. These findings imply that improving interaction quality and learning from prior campaign
outcomes may be more impactful than simply increasing contact frequency.

2.​ Which customer segments are most likely to subscribe to a term deposit, and
what are their typical characteristics?

K-means clustering was performed using customer attributes only (excluding the target
variable y). The target y was used afterward to compare subscription rates across clusters.

To understand which customers are more likely to subscribe to a term deposit, the dataset
was segmented into three clusters (C1–C3). The subscription rate for each cluster was
calculated as:

Subscription rate = (number of “yes”) / (total customers in the cluster)

Segment C2 – High-value, low personal-loan customers (most likely to

subscribe)

C2 shows the highest subscription rate (16.3%), suggesting this is the most responsive
segment for term deposit offers. Customers in this cluster have the highest average account
balance (5093) and the lowest personal-loan prevalence (8.2%). Their average age is
around 42.5, placing them in a middle-aged group. They also have the highest average
number of previous contacts (0.758), which may indicate stronger prior engagement with the
bank.

Summary: financially stronger customers with low personal debt and slightly higher prior
engagement are the most likely to subscribe.

Segment C3 – Younger, housing-loan-heavy customers (moderate likelihood)

C3 has a subscription rate of 10.5%, which is lower than C2. This cluster is the youngest
group (33.9 years) and has a low average balance (523). A key feature of this segment is
the high housing-loan proportion (67.3%), suggesting many customers are already
managing major financial commitments. Their average previous contacts (0.624) is
moderate.

Summary: younger customers with lower balances and higher housing-loan burden appear
less likely to commit funds to term deposits.
Segment C1 – Older, lower-engagement customers (lowest likelihood)

C1 records the lowest subscription rate (10.1%). Customers in this cluster are the oldest
(51.9 years) with a moderate average balance ( 692). They also have the lowest average
previous contacts (0.440), which may reflect weaker engagement with prior marketing
efforts. Compared with the other groups, C1 has the lowest housing-loan proportion (41.7%),
but this does not translate into higher subscription.

Summary: although older, this group shows lower engagement and does not subscribe more
than the younger group.

Conclusion
Overall, Segment C2 is the most likely to subscribe to a term deposit. The profile of this
segment suggests that customers with higher balances, lower personal debt, and stronger
prior engagement are more responsive to term deposit products. In contrast, C1 and C3
show similarly lower subscription rates, with C3 likely constrained by higher housing-loan
commitments and C1 potentially limited by weaker engagement.

7 Conclusion
This study used the Bank Marketing dataset from the UCI Machine Learning Repository,
consisting of 45,211 records collected from telephone-based marketing campaigns
conducted by a Portuguese bank between 2008 and 2010. The dataset includes 16 predictor
variables describing customer demographics, financial conditions (e.g., balance, housing
loan, personal loan), and campaign-related interaction features, with a binary target variable
indicating whether a customer subscribed to a term deposit.

Following the CRISP-DM framework, two main experiments were conducted. First, a
classification task was performed to predict term deposit subscription using Logistic
Regression, Neural Network, and Decision Tree models. Model performance was evaluated
using multiple metrics—AUC, accuracy, precision, recall, F1-score, and MCC—to account for
the imbalanced distribution of the target class. The results show that Logistic Regression
and Neural Network achieved similarly strong predictive performance (AUC around 0.90),
while the Decision Tree performed substantially worse. Logistic Regression was selected as
the final model because it provides competitive performance while remaining interpretable
for factor analysis. Based on the logistic regression nomogram, the most influential
predictors include call duration, previous contacts, campaign intensity, previous campaign
outcome, and the month of contact. Overall, the findings suggest that subscription behaviour
is driven primarily by interaction intensity and campaign history, with seasonality also playing
a meaningful role.

Second, K-means clustering was applied to segment customers into three groups and to
identify which segments were more likely to subscribe. The analysis shows that Cluster C2
has the highest subscription rate (16.3%) and is typically characterised by a higher average
account balance, a lower prevalence of personal loans, and slightly higher prior engagement
(more previous contacts). In contrast, the other clusters exhibit lower subscription rates and
are associated with lower balances or heavier loan commitments.

Despite these insights, the study has several limitations. The target variable is imbalanced,
which can affect model calibration and the interpretation of accuracy. In addition, some
predictors—especially call duration—are only observed during the interaction and are not
available before contacting the customer, limiting their usefulness for pre-call targeting.
Future work could address these issues by applying imbalance-handling techniques (e.g.,
class weighting or resampling), optimising decision thresholds based on the cost trade-off
between false positives and false negatives, and developing a “pre-call” model that uses
only variables known before contacting customers. Further extensions may also include
testing stronger ensemble methods (e.g., gradient boosting) and using advanced
explainability tools such as SHAP to validate and deepen the interpretation of key drivers.

Abstract
This study investigates customers’ subscription decisions for term deposits using the Bank
Marketing dataset from the UCI Machine Learning Repository. The dataset contains 45,211
observations collected from telephone-based direct marketing campaigns conducted by a
Portuguese bank between 2008 and 2010. It includes 16 predictor variables covering client
demographics (e.g., age, job, education), financial status (e.g., balance, housing loan,
personal loan), and campaign interaction features (e.g., contact duration, number of contacts
in the current campaign, and previous campaign outcomes). The target variable indicates
whether the customer subscribed to a term deposit.

Following the CRISP-DM framework, the analysis consists of two experiments. First, three
classification models—Logistic Regression, Neural Network, and Decision Tree—were
developed to predict term deposit subscription. Models were evaluated using AUC,
accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC) to account
for the imbalanced class distribution. Logistic Regression and Neural Network achieved
comparable predictive performance (AUC approximately 0.90), while the Decision Tree
performed substantially worse. Logistic Regression was selected as the final model because
it provides competitive performance and strong interpretability for factor analysis. Using a
nomogram, the most influential factors were identified as call duration, previous contacts,
campaign intensity, previous campaign outcome, and month of contact.

Second, K-means clustering was applied to segment customers into three groups.
Subscription rates were calculated for each segment, revealing that Cluster C2 had the
highest subscription rate (16.3%). This segment typically featured higher account balances,
lower personal-loan prevalence, and slightly higher prior engagement. Overall, the findings
suggest that subscription behaviour is driven more by campaign interaction dynamics and
financial readiness than by demographics alone, providing actionable insights for improving
targeting strategies in future campaigns.