Customer Clustering of a Financial Institution for Generation of Personalized Economy Alerts

Ewerton Martins de Menezes; Beatriz Garcia Lopes

DOI: 10.22167/2675-6528-202602877

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Digital transformation and analytical intelligence in the banking sector have enabled a more consultative approach by financial institutions. The study aimed to develop a system for generating personalized savings alerts based on the history of financial transactions. The adopted methodology involved three stages: firstly, individuals were grouped using multiple clustering algorithms, notably K-means and TimeSeriesKMeans. Subsequently, contextual anomalies were detected, mapping statistically significant deviations in consumption behavior within groups and their respective time series, using the Isolation Forest, Mahalanobis Distance, and modified Z-Score algorithms. Finally, contextual alerts were generated, translating mathematical deviations into structured interventions. The results of this strategy, applied to a sample of 2,679 account holders from a large Brazilian financial institution, demonstrated its technical feasibility, identifying at least one personalized alert for 69.50% of the analyzed base. The main contribution consisted of operationalizing behavioral economics theories with machine learning algorithms. It was concluded that the convergence between data science and behavioral economics can enable scalable financial advisory services tailored to each client’s reality.

Keywords: Unsupervised learning; Anomaly detection; Personal finance; Time series; Nudge theory.

1. Introduction

The growing digitalization of payment methods and the decline in the use of cash have transformed the banking sector, granting financial institutions unprecedented visibility into their customers’ consumption habits. This scenario of abundant transactional data, however, contrasts with a low level of financial literacy in the population.

Data from the National Confederation of Commerce (CNC, 2026) indicate that the indebtedness of Brazilian families remains at high levels, evidencing a gap in financial literacy, especially among lower-income groups. Traditional financial advisory approaches, being generic and disconnected from individual reality, demonstrate limited effects in this context.

To overcome these limitations, recent literature indicates that data-driven strategies are fundamental to enhancing the customer experience and enabling personalized services at scale, overcoming the restrictions of traditional banking services (Adeniran et al., 2024). In this context, customer segmentation transcends a purely operational activity and consolidates itself as a strategic tool for retention and marketing.

Business analysis, as reinforced by Hossen et al. (2025), allows for the identification of hidden patterns in consumer behavior, aiding in more assertive managerial decision-making. The importance of dynamic segmentation, which combines classic models such as Recency, Frequency, Monetary Value (RFM) with time series techniques, is highlighted by Abbasimehr and Bahrini (2022). In addition to transactions, emotional and sociodemographic factors also influence loyalty, suggesting segmentations based on customer sentiments towards banking services (Calvo-Porral and Lévy-Mangin, 2020).

Among the unsupervised learning techniques for segmentation, the K-means algorithm is an industry benchmark due to its implementation simplicity, ease of interpretation, and computational efficiency on large databases (Hossen et al., 2025; Thompson et al., 2021). However, K-means has limitations, such as sensitivity to “outliers” and difficulty in identifying irregularly shaped clusters. To mitigate these problems, Yan et al. (2025) propose improvements in density-based algorithms, such as DBSCAN, which are more effective in identifying noise and separating complex groups, although they require greater effort in parameterization.

More recent advances explore the use of Deep Learning, with models like Deep Clustering that utilize autoencoders and neural networks to capture non-linear relationships in complex financial data (Choi et al., 2024). However, Wei et al. (2024) warn that the computational cost and the “black-box” nature of neural networks still impose barriers to the widespread adoption of these techniques compared to classical methods. The effectiveness of clustering algorithms is also affected by the high dimensionality and correlation between variables, a challenge addressed by Zhao et al. (2021) with regularized versions of K-means, and by Nisha and Pramod (2025) who highlight dimensionality reduction for segmentation and predictive models.

In the context of financial education, the personalization of guidance is a critical factor for user engagement. This work is based on the theory of “nudge” (Thaler and Sunstein, 2009), from behavioral economics, which advocates the possibility of guiding people to make better decisions without restricting their freedom of choice. Unlike traditional interventions that resort to generic social comparison, which can generate frustration (Beshears et al., 2015), the proposed strategy employs clustering algorithms to delimit customer segments with similar sociodemographic profiles and consumption habits. Thus, the “nudge” acts in a contextualized manner and is enhanced by the prior categorization of transactions, broadening the range of possible interventions and enabling highly targeted recommendations. Given the above, the objective of this work is to generate personalized savings alerts through customer clustering combined with anomaly detection using their sociodemographic profiles and consumption behaviors, with the aim of encouraging more conscious and sustainable financial decisions.

2. Material and Methods

This work adopted an exploratory quantitative approach, using transactional data from real clients of a large financial institution in Brazil. The research was conducted with a sample of 2,679 individuals, selected from a universe of 24 million individual clients. The sample calculation followed the methodology of Fávero and Belfiore (2025), considering an infinite population, 95% confidence, and a 5% margin of error.

The data collection covered approximately 4.4 million financial transactions, recorded between January and December 2025. To ensure ethics and the protection of sensitive data, the database was fully anonymized by the institution before academic access. Institutional authorization was formalized by a Term of Consent, ensuring compliance in data processing.

Data preparation involved the selection of sociodemographic and transactional attributes, categorized by an internal institutional solution. After cleaning invalid transactions and removing two clients with incomplete data, the final sample consisted of 2,679 clients and 4,065,152 valid transactions. Qualitative variables were encoded, including “One-Hot Encoding”.

Feature engineering condensed the transactional history into statistical metrics per client and per month, creating five consumption dimensions (Revenues, Expenses, Results, Investments, and Card Expenses) and time series variables. Data scaling and normalization were crucial to handle heterogeneity. Continuous variables were scaled with “StandardScaler”, asymmetric ones with “PowerTransformer”, and percentages with “MinMaxScaler”.

The clusterization modeling sought to create internally homogeneous and externally heterogeneous groups (Fávero and Belfiore, 2025), in two phases: individual optimization and global comparison. Five algorithms were selected: K-means, K-prototypes, HDBSCAN (Campello et al., 2015), Gaussian Mixture Model (GMM), and K-means for time series (TimeSeriesKMeans). The optimization of each algorithm used specific internal metrics, such as Elbow Method, Silhouette Coefficient, DBCV, BIC, and AIC. For K-means, Principal Component Analysis [PCA] was applied for dimensionality reduction.

K-means for time series employed the Dynamic Time Warping [DTW] metric for longitudinal similarity (Abbasimehr and Bahrini, 2022). In the global comparison phase, the algorithms were evaluated by agnostic metrics, including Kruskal-Wallis and Chi-Squared tests for practical significance, and Epsilon-Squared (€²) and Cramér’s V for effect size (Cohen, 1988).

Contextual anomaly detection focused on significant deviations from the group pattern. For critical variables, the Modified Z-score (Iglewicz and Hoaglin, 1993) was applied, with anomalies defined by Z-score above 3.5. Mahalanobis Distance (Mahalanobis, 1936) was used for inconsistencies in resource allocation, considering data covariance, with anomalies above the critical value of the Chi-squared distribution with 99% confidence. To monitor abrupt behavioral changes, the Isolation Forest algorithm (Liu, Ting and Zhou, 2008) was employed, trained individually by cluster with trend variables.

In anomaly detection in time series, the ideal behavior of each cluster was extracted by DTW Barycenter Averaging [DBA] (Petitjean et al., 2011). The DTW distance between the individual series and the group’s barycenter served as the anomaly coefficient, with the Modified Z-score applied to trigger alerts. The final step consisted of generating contextual alerts, translating the detected anomalies into predefined educational messages, based on the principles of behavioral economics and the cognitive bias of social proof. All steps were implemented in the Spyder IDE tool, using the Python language and libraries such as Scikit-learn, Pandas, and Numpy.

3. Results and Discussion

The present study aimed to develop a personalized savings alert generation system, based on the analysis of customers’ financial transaction history. The adopted methodological approach demonstrated the technical viability of the proposed solution, identifying at least one personalized alert for 69.50% of the analyzed base, composed of 2,679 account holders from a large Brazilian financial institution. The results obtained highlight the capacity to operationalize principles of behavioral economics, such as “nudge” theory, through machine learning algorithms, promoting scalable financial advisory tailored to the individual reality of each client.

Data Understanding

The analyzed sample revealed a diversified sociodemographic profile, with 66% men and 34% women. A significant concentration of clients with complete higher education and postgraduate degrees was observed, indicating a high level of education. Income profiles B and C were the most frequent, encompassing 99% of the sample. The clients’ ages ranged from 22 to 75 years, with an average of 44 years, reflecting the inclusion of both recent and long-standing clients in the institution. The majority of clients had only one active account, although the sample also included individuals with multiple accounts.

Regarding marital status, 55% of clients were married and 34% single. The average income observed was R$ 50,457.51, characterized by a high range and dispersion, encompassing clients with no formal registered income to those with incomes 20 times higher than the general average. Concerning transactional behavior, 4,065,152 valid financial transactions were analyzed. Of this total, 45.76% were made with credit cards, while 54.24% corresponded to transfers, payments, purchases, revenues, withdrawals, and cash deposits. The distribution of these movements remained relatively constant over the months, with an average of 338,000 monthly transactions.

Feature Engineering and Preprocessing

The feature engineering stage was crucial for transforming transactional history into metrics that captured consumption behavior. The data were grouped by customer and by month, resulting in the creation of five main dimensions: Revenues (RC), Expenses (GSTO), Result (RSTD), Investments (INVS), and Card Expenses (GSTO_CRT). Additionally, expenses were detailed by macrocategories (debt, essential, unclassified, and superfluous) and ten specific categories, calculating the percentage in relation to each customer’s total monthly spending. This granularity allowed for an in-depth analysis of consumption patterns.

To capture the behavior of expense dimensions and categories over 12 months, two extraction strategies were employed. The first consisted of statistical aggregations, generating 76 consumer behavior variables for each client, including mean (MED), standard deviation (DPDR), maximum (MAX), and trend (TND). The second strategy involved creating time series, sequencing the monthly averages of consumption dimensions and spending by category, resulting in 19 consumer behavior variables for each client, structured in a three-dimensional cube for the application of specific algorithms.

The heterogeneity of the collected attributes, which included ordinal, percentage, temporal metric variables, and monetary values with varied distributions and scales, required a robust preprocessing strategy. Sociodemographic variables, such as age, income profile, education level, years of relationship, and number of accounts, were scaled using StandardScaler. This procedure standardized the data, centering the mean at zero and setting the standard deviation to one, preventing attributes with larger numerical ranges from dominating the Euclidean distance calculation and preserving the original variance structure.

Volumetric financial variables, such as assumed income, profit, revenue, expenses, and investments, were normalized using PowerTransformer. This choice is justified by the asymmetric nature of these monetary attributes, which often exhibit right-skewed long-tail distributions due to the presence of high-transaction financial customers. PowerTransformer stabilized the variance and mapped these data to a distribution closer to the normal curve. Expenses by macrocategory and category, expressed as percentage values of total spending, were normalized using MinMaxScaler, which mapped these attributes to the interval of zero to one, preserving the proportions and relative distances of consumption without distortions.

Modeling and Selection of Clustering Algorithms

The clustering modeling was performed in two phases to create internally homogeneous and externally heterogeneous groups. In the first phase, five algorithms were individually selected and optimized: K-means, K-prototypes, HDBSCAN, GMM, and K-means for time series (TimeSeriesKMeans). Dimensionality reduction was an essential step, with the application of Principal Component Analysis (PCA) which compressed the original space to nine principal components, retaining 80.42% of the explained variance. In scenarios of nonlinear dimensionality reduction via UMAP, the “Trustworthiness” metric was maintained above 80%, ensuring the preservation of original proximity relationships among customers.

The optimization of the number of clusters (k) for K-means and K-means for time series, using the Elbow Method, did not reveal a sharp slowdown point in the inertia curve, suggesting a continuous and complex distribution of financial data. The Silhouette Coefficient analysis for these algorithms indicated a global maximum value at k = 2, which, from a business perspective, would result in an excessively generic and unviable binary segmentation for personalized alerts. Therefore, the ideal k was defined based on local peaks, adopting k = 5 for K-means, k = 4 for K-prototypes, and k = 6 for K-means for time series.

In the case of the GMM algorithm, the optimization converged to k = 7, minimizing the BIC and AIC information criteria, which suggests that seven Gaussian distributions better represent the probability of customer belonging. The HDBSCAN optimization involved evaluating 324 hyperparameter combinations and dimensionality reduction via UMAP. The optimal configuration, with UMAP of 2 components, 10 nearest neighbors, minimum distance of 0.01, and Euclidean distance, combined with HDBSCAN parameters of minimum cluster size of 30 and neighbor density rigor of 3, resulted in the formation of four clusters, classifying only 1 customer as noise and achieving a DBCV of 0.75.

In the second phase of modeling, the global comparison of algorithms revealed that K-means for time series and GMM presented a more balanced population distribution. In contrast, HDBSCAN concentrated 72.52% of the sample (1,943 clients) in a single group, and K-means and K-prototypes generated clusters with less than 1% of the sample base, indicating the presence of extreme “outliers”. Although the Kruskal-Wallis and Chi-Squared statistical tests showed significant differences between clusters for most variables across all algorithms, the effect size measurement (Epsilon-Squared and Cramér’s V) provided more relevant methodological contrasts.

HDBSCAN stood out in alignment with qualitative business labels, while GMM performed better in spending categories. K-means demonstrated analytical superiority in partitioning volumetric financial variables, with an average effect size of 0.138. This capability to maximize the separation of quantitative magnitudes ratified the choice of K-means as the main algorithm for cluster formation, providing a robust foundation for the subsequent anomaly detection phase. The effectiveness of K-means, as evidenced by Zhao et al. (2021), is sensitive to high dimensionality and correlation between variables, which justifies the preprocessing and dimensionality reduction steps applied.

Characterization of Behavior Clusters

The five clusters generated by the K-means algorithm revealed distinct behavioral profiles, despite some overlap in intermediate income and expense profiles. This segmentation ensured the necessary heterogeneity for generating personalized alerts. Cluster K0, named “Debtors”, comprised 639 clients with an average age of 46 years, predominantly married men, with a high level of education and high income. This group presented the worst average financial result in the sample, with a deficit of R$ -3,458.61 and a negative investment balance of R$ -569.37, suggesting a dependence on credit lines to maintain their standard of living.

Cluster K1, “New clients”, was the youngest group in the sample, with the highest proportion of women (38%) and the lowest income profiles. With 837 clients, it generated the lowest financial volume, but managed to balance revenues and expenses, maintaining a positive investment balance. It stood out for having the highest proportion of spending on personal expenses (17.1%) and the largest share of the budget committed to debt (5.1%). Cluster K2, “Average client”, was the largest group numerically, with 1,138 clients of intermediate income and education levels. This group presented the highest budget commitment to essential expenses (37.4%), a small average deficit of R$ -867.22, and small investment reserves of R$ 920.12, leading expenses in categories such as Transportation (9.3%), Home (8.6%), and Food (8.5%).

Cluster K3, “Saver”, with 47 clients, presented the highest average age, strong male predominance, high socioeconomic profile, and long relationship history. The income of this group was significantly higher than resource outflows, generating an average result of R$ 33 thousand and an investment volume of R$ 73 thousand. It showed the lowest proportion of essential expenses (29.70%), indicating greater income availability for investments and leisure. Furthermore, it recorded the highest proportion of expenses in fees and taxes (5.1%) and unclassified expenses (36.94%), suggesting frequent use of cash. Cluster K4, “Elite/High Income”, was the smallest group, with 18 clients, representing the top of the financial pyramid. It held the highest volumes of income, spending, and revenue, with an individual financial surplus of R$ 110 thousand, directed towards investments. It was the only group with a relevant presence of the A income profile and the highest concentration of master’s and doctoral degrees, in addition to presenting the highest proportional expenses in leisure (7.9%) and health (6.6%).

The analysis of the barycenters of the six clusters generated by K-means for time series (T0 to T5) revealed distinct temporal patterns in financial behavior throughout 2025. A revenue peak in February was noticeable across all groups, influencing net balances and investments, suggesting a strong seasonal event in the database. Cluster T0, the smallest in numerical terms, was characterized by a strong downward trend in activity throughout the year, with continuous retraction of revenue and expenses from the second half onwards. Cluster T1 presented secondary revenue peaks in April and September, with April concentrating the group’s highest expenditure, reflected in negative results and investment balances.

The T2 Cluster, unlike the others, had its revenue peak in September, accompanied by a spending peak, with no financial surplus or investment balance, and fluctuating credit card consumption. The T3 Cluster, the largest group in the base, also showed a secondary revenue peak in September, with positive results and investment balance, and stable credit card consumption, but with an upward trend at the end of the year. The T4 Cluster registered a secondary revenue peak in September without a spending peak, resulting in a financial surplus and investment balance. Finally, the T5 Cluster presented several revenue peaks throughout the year that were not reflected in the other financial metrics, with a clear spending peak in July that reduced investment balances.

Anomaly Analysis and Alert Generation

Contextual anomaly detection was a central component for alert generation, differentiating itself from traditional approaches that only seek global “outliers”. For alerts on critical variables, the Modified Z-score was used, which replaces the mean with the median and the standard deviation with the median absolute deviation (MAD), making the estimators more robust. Z-score values above 3.5 were considered anomalies. To detect inconsistencies and imbalances in the resource allocation structure, such as the relationship between income, revenues, expenses, and investments, Mahalanobis Distance was employed, which considers the data’s covariance matrix, a fundamental aspect in finance. Observations exceeding the critical value of the Chi-squared distribution with 99% confidence were classified as anomalies.

To monitor abrupt behavioral changes, the Isolation Forest algorithm was applied, isolating anomalies through random tree partitions, proving effective in capturing complex patterns without assuming a normal distribution. This algorithm was trained individually for each cluster, based on outcome trend, revenue, spending, and investment variables, allowing for the identification of clients whose financial situation had abruptly altered. For alerts based on K-means clustered consumption time series, anomaly detection was structured from the sequential alignment of data, using the DTW Barycenter Averaging (DBA) method to calculate the weighted average temporal sequence of each cluster. The DTW distance between the individual client’s time series and their group’s barycenter was used as the anomaly coefficient, and the Modified Z-score was applied to quantify the deviation and trigger the alert.

The first set of alert rules, focused on structural and trend anomalies, identified financial inconsistencies and trend breaks in relation to peers, reaching 18.66% of the sample (500 clients) with the generation of 544 alerts. Specifically, 161 clients received alerts about “Downtrend in budget”, detected by Isolation Forest and a negative balance trend result. Another 153 clients were alerted about “Structural imbalance between income and expenses”, based on Mahalanobis Distance, indicating a significant deviation from their profile pattern. Additionally, 230 clients were identified with “Growth in debt spending”, based on Z-score above 3.5 and an average above 5% for debt variables.

The second category of alerts, focused on excessive spending in specific categories and credit card usage, impacted 30.27% of the base (811 users), generating a total of 862 alerts. Examples include 20 clients with “Excessive credit card spending”, with an average of R$ 5,450.45, and 14 clients with “Excessive lifestyle spending”, representing 23% of total spending. Alerts about “Excessive debt spending” reached 624 clients (17% of the total), “Excessive communication spending” 64 clients (24% of the total), “Excessive personal expenses spending” 13 clients (32% of the total), “Excessive leisure spending” 78 clients (36% of the total), and “Excessive transportation spending” 49 clients (46% of the total), all based on modified Z-score and/or percentage average above defined limits.

The detection of deviations in total spending and credit card spending time series, using the modified Z-score, identified abrupt changes in consumption behavior in 3.43% of the base (92 clients), resulting in 93 alerts. Of this total, 3 clients were alerted about “General spending deviation” and 90 clients about “Credit card spending deviation”. In addition to alerts based on statistical deviations, alerts were generated to encourage good financial practices, covering 20.38% of the sample (546 clients) and totaling 614 interventions. This included 307 clients with “Investment potential”, identified by available balance above average and investments below expectations, and 303 clients with “Unclassified spending”, with more than 50% of their spending uncategorized.

From the perspective of positive reinforcement, alerts were generated for maintaining healthy habits and surplus scenarios, impacting 35.91% of the base (962 individuals) with 1,095 alerts. This category included 526 clients with “Investment share growth”, identified by investments and positive investment trends, and 551 clients with “Low indebtedness level”, characterized by high average revenue and low debt proportion. Proactive recognition of good behaviors acts as a trigger to strengthen user relationship, as suggested by Calvo-Porral and Lévy-Mangin (2020), who highlight the influence of positive emotions on customer satisfaction and retention in banking.

In summary, the proposed solution demonstrated its technical viability to operate at scale, generating a total of 3,118 alerts and making 69.50% of the sample (1,862 clients) eligible to receive at least one contextual alert across different dimensions of financial life. The transition from a static segmentation model to a dynamic monitoring structure, based on anomaly detection not only in the total transaction volume but also in categories, proved fundamental to broadening the range of possible interventions. The results highlight the analytical capability to identify structural anomalies, excessive spending, and significant morphological deviations in time series, automating the identification of vulnerabilities before they culminate in default. The approach, grounded in “nudge” theory (Thaler and Sunstein, 2009), allows deviations to be calculated by comparing the client only with peers from their own mathematical grouping, ensuring a fair and contextualized personalization of financial guidance.

4. Conclusion

The present study aimed to develop a personalized savings alert generation system based on financial transaction history. The technical feasibility of the proposed solution was verified, identifying at least one personalized alert for 69.50% of the analyzed base, composed of 2,679 account holders. The employed methodology, which involved clustering clients into distinct sociodemographic and behavioral profiles, combined with the detection of contextual anomalies through algorithms such as K-means, Isolation Forest, Modified Mahalanobis Distance, and Z-score, allowed for the identification of structural financial inconsistencies, excessive spending in specific categories, and significant morphological deviations in consumption time series. This approach resulted in the generation of 3,118 alerts, making 1,862 clients eligible for targeted interventions. The main contribution of this work lies in the operationalization of behavioral economics principles, such as “nudge” theory, through machine learning algorithms, enabling scalable financial advisory tailored to the individual reality of each client, and promoting financial education.

Although the methodological architecture for alert generation has demonstrated its analytical effectiveness, a limitation of the study was the absence of measuring the behavioral effectiveness of these messages in the field. For future studies, it is suggested to apply controlled practical tests with end-users, aiming to validate the real impact of “nudges” on changing consumption habits. Additionally, it is recommended to investigate aspects related to the prioritization and frequency of sending these notifications, in order to mitigate informational fatigue, especially in cases of customers eligible for multiple alerts. The convergence between data science and behavioral economics, as evidenced by this work, has the potential to transform the performance of financial institutions, making it more consultative and proactive in promoting customers’ financial health.

Bibliographic References

Abbasimehr H.; Bahrini A. 2022. An analytical framework based on the recency, frequency, and monetary model and time series clustering techniques for dynamic segmentation. Expert Syst Appl 192: 116373.

Adeniran, I. A.; Abhulimen, A. O.; Obiki-Osafiele, A. N.; Osundare, O. S.; Agu, E. E.; Efunniyi, C. P. 2024. Data-driven approaches to improve customer experience in banking: techniques and outcomes. International Journal of Management & Entrepreneurship Research 6(8): 2797-2818.

Beshears, J.; Choi, J. J.; Laibson, D.; Madrian, B. C.; Milkman, K. L. 2015. The Effect of Providing Peer Information on Retirement Savings Decisions. The Journal of Finance, 70(3): 1161-1201.

Calvo-Porral C.; Lévy-Mangin JP. 2020. An emotion-based segmentation of bank service customers. International Journal of Bank Marketing 38 (7): 1441-1463.

Campello, RJ.G.B; Moulavi, D.; Zimek, A.; Sander, J. 2015 Hierarchical Density Estimates for Data Clustering, Visualization, and Outlier Detection. ACM Trans Knowl Discov Data. 10(1):1-51.

Choi I.; Koh W.; Koo B.; Chang Kim W. 2024. Network-based exploratory data analysis and explainable three-stage deep clustering for financial customer profiling. Eng Appl Artif Intell, 128: 107378.

Cohen, J. 1988. Statistical Power Analysis for the Behavioral Sciences. 2. ed. Routledge. New York.

Confederação Nacional do Comércio de Bens, Serviços e Turismo [CNC]. 2026. Pesquisa de Endividamento e Inadimplência do Consumidor [PEIC]. Disponível em: <https://pesquisascnc.com.br/pesquisa-peic/>. Acesso em: 01 abr. 2026.

Fávero, L. P.; Belfiore, P. 2025. Manual de Análise de Dados: Estatística e Machine Learning com Excel®, SPSS®, Stata®. 2. ed. LTC, Rio de Janeiro, RJ, Brasil.

Hossen, M. A.; Bhattacharjee, B.; Dey, S. K.; Jamee, S. S.; Obaid, M. O.; Mia, M. S. I.; Khan, S. I.; Sharif, M. K. 2025. Business analytics for customer segmentation: a comparative study of machine learning algorithms in personalized banking services. International Journal of Economics Finance & Management Science, 10(03): 1-13.

Iglewicz, B.; Hoaglin, D. C. 1993. How to Detect and Handle Outliers. ASQC Quality Press.

Liu, F. T.; Ting, K. M.; Zhou, Z. H. 2008. Isolation forest. Proceedings – IEEE International Conference on Data Mining, ICDM: 413-422.

Mahalanobis, P. C. 1936. On the generalized distance in statistics. Proceedings of the National Institute of Sciences of India 2(1): 49-55.

Nisha T.N.; Pramod D. 2025. Improving bank customer churn prediction with feature reduction using GA. Scientific Reports 15(1): 40035.

Petitjean F.; Ketterlin A.; Gançarski P. 2011. A global averaging method for dynamic time warping, with applications to clustering. Pattern Recognit. 44(3): 678-693.

Thaler, R. H.; Sunstein, C. R. 2009. Nudge: Improving Decisions About Health, Wealth, and Happiness. Penguin Publishing Group.

Thompson J. R. J.; Feng L.; Reesor R. M.; Grace C. 2021. Know your clients’ behaviours: a cluster analysis of financial transactions. Journal of Risk and Financial Management 14(2): 50.

Wei X.; Zhang Z.; Huang H.; Zhou Y. 2024. An overview on deep clustering. Neurocomputing, 590: 127761.

Yan X.; Li Y.; Nie F.; Li R. 2025. Bank customer segmentation and marketing strategies based on improved DBSCAN algorithm. Applied Sciences 15(6): 3138.

Zhao H.; Luo X.C.; Ma R.; Lu X. 2021. An extended regularized k-means clustering approach for high-dimensional customer segmentation with correlated variables. IEEE Access 9: 48405-48412.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform