Customer Segmentation for Credit Card in Cooperatives with Machine Learning

Bernardo Augusto Zucco; Adriana Camargo de Brito

DOI: 10.22167/2675-6528-202602718

Article derived from a Final Course 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 the Pecege Institute for textual synthesis and organization.

Summary

The segmentation of individual members of a Brazilian credit cooperative was proposed to support the personalized offering of credit cards in a competitive financial scenario. The study used a database with 59,108 credit card transaction records. The methodology employed the Python language for data processing and unsupervised machine learning techniques. Initially, the data were standardized with StandardScaler and dimensionality was reduced by Principal Component Analysis (PCA), retaining five components. For clustering, the K-means algorithm was applied, with scenarios of two and five clusters defined by methods such as Elbow, Silhouette, Calinski-Harabasz, and Davies-Bouldin. The two-cluster model offered a macro view between standard and intensive consumers, while the five-cluster model presented greater strategic granularity, identifying a “dormant consumer” segment with high conversion potential. The results provide subsidies for targeted marketing strategies, aiming at retaining high-value customers and strengthening institutional relationships.

Keywords: Cluster; K-means; PCA; Customer segmentation.

1. Introduction

The base of credit card users in Brazil has experienced an expressive expansion in recent years, driven by growing competition resulting from the entry of new financial institutions. As a consequence, a significant portion of the population has gained access to one or multiple credit cards (Central Bank of Brazil, 2022). The market offers different types of cards, from basic ones, exclusively for payments, to differentiated ones, which may include benefit and reward programs, generally with a higher annual fee (Central Bank of Brazil, 2023).

In this highly competitive scenario, the idea of customer segmentation has progressed significantly. Previously, segmentation was based primarily on characteristics such as demographics, age, income, and occupation. Over time, the credit card industry has become more competitive and data-driven, making segmentation significantly more sophisticated, focusing on the customer’s actual behavior, their effective actions, mindset, aspirations, economic importance, and financial needs (Krishnan, 2024).

The adoption of segmentation strategies becomes imperative to identify the target audience for each type of card. By aligning the offer with the specific needs of the consumer, financial institutions not only increase satisfaction levels but also contribute to strengthening the long-term relationship. Conceptually, customer segmentation consists of dividing the market into distinct groups that share common particularities, allowing the development of targeted tactics for specific profiles (Kansal et al., 2018; Güçdemir and Selim, 2015).

In the scope of data science, the grouping of similar profiles is enabled by unsupervised learning techniques, specifically through clustering algorithms. Unlike supervised models, clustering tools provide the identification of intrinsic patterns in the data without the need for prior labels, based on the similarity between samples for group formation (Harrison, 2019). The use of the K-means algorithm allows for the segmentation of customers into distinct groups, enabling the development of specific marketing strategies for each segment. In conjunction with this algorithm, it is possible to use Principal Component Analysis (PCA), an unsupervised statistical technique that aids in dimensionality reduction, commonly used before applying K-means to extract principal component features from sample variables (Umuhoza et al., 2020; Abdulhafedh, 2021; Wang, 2025).

In this context, the present study aims to identify the groupings of individual cardholders who are credit card users, as well as their distinctive characteristics. It seeks to support the refinement of marketing and communication strategies for offering differentiated cards. Additionally, it aims to identify high-value customers, contributing to the strengthening of institutional relationships for the offering of complementary products within the credit card ecosystem.

2. Material and Methods

This study was characterized as a quantitative research, of an exploratory and descriptive nature, with the purpose of identifying clusters of individual clients who are credit card users and their distinctive characteristics. The methodological approach employed unsupervised machine learning techniques, aligning with the objective of subsidizing the refinement of marketing and communication strategies for the offering of differentiated cards, as well as the identification of high-value clients.

The unit of analysis consisted of individual members of a credit cooperative operating in the southern region of Brazil. A database provided by the institution itself was used, comprising 59,108 records of members who conducted transactions using credit in the last six months, specifically from February 2025 to July 2025. The data were selected in conjunction with the product management area, based on criteria for limit “upgrades”, card categories, and portfolio segmentation.

To ensure compliance with the General Data Protection Law and the cooperative’s internal guidelines, the database was subjected to anonymization procedures. Sensitive data was suppressed and unique records were excluded, mitigating the risk of individual re-identification. After this processing, the dataset concluded with thirteen variables, of which the “IncomeBracket” variable was excluded for being categorical and not suitable for the K-means algorithm, resulting in twelve numerical variables for analysis.

The twelve remaining numerical variables were organized into four groups to structure the analysis. Two relationship and reciprocity variables were included, which measured registration time and the number of contracted products. Six credit card consumption variables detailed outstanding balances in other institutions and the cooperative, utilized and contracted limits, and values and quantities of transactions performed. Three financial profile variables covered average outstanding balances, investments, and current account balance. Finally, one demographic variable represented the member’s age.

The data processing, analysis, and algorithm application were performed using the Python programming language, which offers access to various libraries for statistics, mathematics, table manipulation, and machine learning. This tool allowed for efficient data manipulation and visualization of intermediate results.

Due to the distinct magnitudes of the variables, which could disproportionately influence the distance calculations, the data were standardized. This process, according to Fávero and Belfiore (2024), adjusted the variables to the same scale by subtracting the mean value of the observations and dividing by the standard deviation. Standardization was implemented using the `StandardScaler()` function from the Scikit-Learn library (Géron, 2021), resulting in data with a mean of zero and a standard deviation of one.

For dimensionality reduction, Principal Component Analysis (PCA), an unsupervised statistical technique, was applied. PCA aimed to extract the most relevant factors from the data, compressing the dataset and simplifying its description (Abdi and William, 2010; Fávero and Belfiore, 2024). The technique identifies intrinsic patterns and creates new sets of uncorrelated variables, maximizing the explained variance in each component.

The definition of the number of principal components to be retained was based on the Kaiser criterion, which recommends retaining components with eigenvalues greater than one (Fávero and Belfiore, 2024). Based on this criterion, the decision was made to retain the first five principal components, which together explained 73.07% of the cumulative variance of the dataset. Additionally, the communality of the variables was checked to ensure that the information of each variable was adequately preserved by the extracted components.

To determine the ideal number of clusters (K), the Elbow method, Silhouette Coefficient, Calinski-Harabasz Index, and Davies-Bouldin Index were employed (Harrison, 2019; Xu and Yusheng, 2019; Abdulhafedh, 2021; Géron, 2021). The Elbow method evaluates inertia, the Silhouette Coefficient measures cluster cohesion and separation, the Calinski-Harabasz Index assesses the ratio between internal dispersion and separation, and the Davies-Bouldin Index quantifies internal compactness and separation between groups.

For data clustering, the K-means algorithm was used, a technique of unsupervised learning that segments data into a pre-defined number of clusters (Géron, 2021; Fávero and Belfiore, 2024). The algorithm operates iteratively, starting with random centroids, assigning each point to the nearest centroid, and recalculating the centroids as the mean of the instances in each cluster, repeating the process until the centroids stabilize. The K-means implementation was performed with the Scikit-Learn library, using inertia as the performance metric.

After applying the algorithm, the results were evaluated to confirm if the variability between clusters was significantly greater than the internal variability of each group. For this analysis, the F-Test was employed, which tests the null hypothesis that the means of the variables are equal in all clusters versus the alternative hypothesis that at least one group has a statistically different mean (Fávero and Belfiore, 2024). The rejection of the null hypothesis, with a significance level of 5%, indicated that the groups formed were statistically distinct.

3. Results and Discussion

The data analysis began with the exploration of the database characteristics and the application of dimensionality reduction techniques, culminating in the segmentation of associates. This process was fundamental to identifying distinct groups of customers, based on their credit card usage patterns and financial profiles. The central objective was to support the refinement of the cooperative’s marketing strategies, aligning product offerings with the specific needs of each associate segment.

The initial statistical analysis of the database revealed important characteristics of the variables. It was observed that the variable “YearsOfRelationship” presented a larger standard deviation, indicating considerable dispersion in the data. In contrast, the variable “NumberOfProducts” exhibited a smaller standard deviation, suggesting that most associates have a quantity of contracted products relatively close to the average. These preliminary observations were crucial for understanding the distribution of information before applying the clustering algorithms.

When examining the credit card consumption variables, all presented a high standard deviation, which signaled the presence of extreme values in the database. Variables such as “VSFNCARTAO_outras_IF” and “VLimiteUtilizadoContaCartaodoCliente” showed a concentration of data in low or zero values, but also with significantly high extremes. Financial balance and investment variables, such as “VSomaSaldoMedio_Poup_RDC_LCA”, exhibited extreme heterogeneity, with standard deviations that exceeded their respective means by more than ten times, indicating a strong concentration of resources in a small portion of the associates.

Main Components

Before applying Principal Component Analysis (PCA), Pearson’s correlation coefficient was calculated between the variables. It was observed that the variables related to credit card consumption showed high positive correlations among themselves. This high collinearity suggests that these variables tend to have similar behaviors and contain redundant information, which could impact the performance of clustering algorithms. PCA was, therefore, an essential tool to mitigate this redundancy and reduce data dimensionality (Abdulhafedh, 2021).

The application of PCA aimed to extract the most relevant factors from the dataset, simplifying its description and the analysis of the structure of observations and variables (Abdi and William, 2010; Fávero and Belfiore, 2024). The definition of the number of principal components to be retained was based on Kaiser’s criterion, which recommends retaining components with eigenvalues greater than one, as they explain a variance equivalent to or greater than that of an original standardized variable.

Based on this criterion, the retention of the first five principal components was chosen. These components, together, explained 73.07% of the accumulated variance of the dataset, demonstrating their significant representation capacity. The analysis of the loading coefficients of these five principal components provided information about the contribution of each original variable to the formation of each component, indicating which variables were most relevant for each.

The first principal component (PC1) had the greatest contribution from transaction and credit card limit variables, mainly reflecting the members’ consumption behavior. PC1 was the one that carried the largest variance among all analyzed variables. Components PC2 and PC4, in turn, showed a stronger relationship with the member’s age and relationship duration with the cooperative, indicating that these demographic and bond characteristics are well represented by these components.

The PC2 and PC3 components also maintained relevant information about the current account balance and the associates’ investment balance, reflecting aspects of financial liquidity and investment capacity. The PC4 and PC5 components, on the other hand, contained information related to the use of credit cards outside the cooperative and the profile of credit-taking associates. PC5, specifically, presented a high loading coefficient for the credit-takers variable, highlighting its importance in this aspect.

The verification of variable commonalities was performed to ensure that the choice of the five principal components was relevant to the study. Commonality indicates the proportion of a variable’s variance that is explained by the set of extracted components (Fávero and Belfiore, 2024). It was observed that no variable presented a low level of commonality, with all values above 0.6. This confirmed that most of the information from each original variable was preserved by the selected components, validating the choice of the first five principal components.

Choice of the Number of Clusters

To identify the ideal number of clusters (K), several validation methods were employed. The elbow plot, which uses inertia, showed a smooth curve, with the decay rate slowing down from five clusters, suggesting an inflection point. The Calinski-Harabasz index presented a peak at two clusters and a local peak at five, indicating well-defined clustering configurations at these points (Xu and Yusheng, 2019).

The Davies-Bouldin index, which seeks lower values to indicate high-quality groupings, revealed its “valleys” at two and five clusters, corroborating the suitability of these configurations. Additionally, the silhouette coefficient analysis suggested that the ideal would be to start with two clusters, observing a drop and stabilization between the third and sixth cluster (Harrison, 2019; Géron, 2021). This partial convergence between the criteria led to the decision to evaluate in detail the solutions with two and five clusters, considered the most plausible for the analytical and business objectives of the study.

Analysis of the Solution with Two Clusters

The evaluation of the two-cluster solution through the F-Test indicated that most of the principal components contributed significantly to the distinction between the groups. It was noted that only the fourth principal component (PC4) presented a p-value of 0.2039, higher than the 5% significance level, suggesting that its contribution to the differentiation of the groupings was not statistically relevant. In contrast, the first principal component (PC1) exhibited the highest F-statistic, demonstrating superior discriminant power in the formation of these two groups of associates (Fávero and Belfiore, 2024).

Internal descriptive statistics of the groupings revealed that Cluster 0 comprised 92% of the associates, totaling 54,089 records. This group was characterized by being large and compact, with a low average distance to the centroid, indicating homogeneous segmentation and few instances that could be considered *outliers*. In contrast, Cluster 1, which represented 8% of the associates, with 5,019 records, showed greater internal dispersion and substantially higher extreme values, suggesting greater heterogeneity and more tail behavior in its characteristics.

Regarding the variables of relationship with the cooperative, Cluster 1 demonstrated, on average, a longer relationship duration and a greater number of contracted products compared to Cluster 0. However, the analysis of quartiles indicated that, although these differences were present, they were not sufficient to sustain, by themselves, a direct and exclusive association between the greater use of credit card and a greater reciprocity or longevity of the relationship with the institution.

The analysis of credit card principalities revealed that, although both clusters used cards from other financial institutions (45% for Cluster 0 and 42% for Cluster 1), Cluster 1 exhibited greater loyalty to the cooperative, with 75% of card spending occurring internally, compared to 61% for Cluster 0. This data is crucial, as it indicates that Cluster 1 members, despite using other cards, concentrate most of their spending within the cooperative.

In terms of credit limits, both clusters used, on average, a small portion of the available limit. The main difference lay in the magnitude of these values, with Cluster 1 having, on average, double the contracted limit and also double the limit effectively used compared to Cluster 0. This suggests that, although Cluster 1 has a greater credit capacity, it does not yet fully exploit it, which may represent an opportunity for the cooperative.

Although Cluster 1 is significantly smaller in number of associates, representing only 8% of the base, it was responsible for 38.55% of the total value transacted in the six months analyzed. This finding is of great relevance, as it highlights a group with a substantially more intense consumption pattern. Furthermore, this group presented an average purchase frequency approximately four times higher than that of Cluster 0, maintaining a high average ticket per transaction.

In the financial profile, Cluster 1 concentrated a relevant participation in the total values of the financial variables. The sum of the average balance of investments in savings, RDC, and LCA (V10) reached 34.24% for this group, and the other two financial profile variables (average debtor balance in loans, financing, and overdraft – V9, and average current account balance – V11) exceeded 25%. These results indicate that Cluster 1 has greater liquidity, significant participation in investments, and greater use of credit operations, characterizing it as a group of Intensive Consumers.

Based on these findings, Cluster 0 was interpreted as the Standard Consumer, representing the majority of the analyzed base. This group is characterized by lower purchase frequency, lower transaction volume, and lower utilization of available credit limits. However, it presents potential for increased spending concentration on the cooperative’s card. Cluster 1, on the other hand, was identified as the Intensive Consumer, a minority group but of high economic relevance, with a consolidated relationship, greater loyalty to the institution’s card, intense consumption, high purchase frequency, high average ticket, and significant participation in financial variables.

Analysis of the Solution with Five Clusters

The five-cluster solution, after excluding a residual cluster of only two observations considered *outliers* due to their lack of statistical representativeness, provided a more in-depth strategic granularity. The F-Test confirmed that all principal components contributed significantly to the differentiation between clusters, with p-values below 5%, and the discriminant power of the components followed a decreasing order from PC1 to PC5, as evidenced by the F statistics (Fávero and Belfiore, 2024).

Internal basic statistics of the clusters indicated a reduction in the average distance in some clusters (0, 1, and 3), suggesting greater internal compaction of the groups. In terms of relationship with the cooperative, Cluster 0 presented the shortest bond time and the smallest quantity of contracted products compared to the other clusters. In contrast, Cluster 1 stood out for presenting the longest average relationship time, while Clusters 2 and 3 exhibited intermediate values, indicating a moderate pattern of bonding with the cooperative.

The principal component analysis with five clusters revealed distinct behaviors. Cluster 0 showed lower concentration of cooperative card usage, with 53% *share*, and almost half of the spending (49%) made at other financial institutions. Cluster 1, in turn, presented 63% *share* in the cooperative, even with 39% usage in other institutions. Clusters 2 and 3 stood out for a high *share* in the cooperative, both above 70%, despite a significant portion of their members (40% to 46%) also using cards from other institutions.

Regarding credit limits, all groupings demonstrated moderate usage. Clusters 0 and 1 presented low limits consistent with consumption. Cluster 2, a select group, exhibited substantially higher limits in relation to its usage, being much higher even than Cluster 3, which also had high limit levels. This suggests that Cluster 2, despite having high credit capacity, does not fully exploit it, indicating potential for the cooperative.

The total transaction value reinforced the importance of smaller clusters in terms of financial contribution. Cluster 3 was responsible for 39.58% of the total transaction value, and Cluster 2, despite representing less than 2% of the base, contributed 14.68% of the total volume. These results highlight striking differences in consumption behavior between the groups, with the higher-value clusters (2 and 3) showing a higher purchase frequency and a higher average ticket.

In the financial profile, Cluster 0 showed lower adherence to the cooperative’s financial products. Cluster 1, despite having a consumption behavior similar to Cluster 0, showed relatively higher averages in financial variables, maintaining an intermediate level of service utilization. Cluster 3 stood out with a significantly higher average in the variable of average outstanding balance in loans, financing, and overdraft (V9), consistently staying ahead in the use of products that the cooperative provides.

The demographic variable “Age” also revealed significant differences between the audiences. Cluster 0 was the youngest and, as observed previously, the one with the shortest relationship time with the cooperative. Cluster 1, in turn, was the oldest group, with an average age difference of approximately 20 years compared to Cluster 0. Clusters 2 and 3 presented intermediate average ages, indicating distinct age profiles that may influence communication and product offering strategies.

Cluster Profiles

From the results obtained for the solution with five clusters, the profiles of the associates can be described in detail. Cluster 0, named Potential Consumer, is an audience with little time in the cooperative, characterized by low frequency of credit card use and still modest limits. This group represents a strategic opportunity for the cooperative to become their main financial institution, through targeted actions that encourage engagement and the use of its products.

Cluster 1, identified as Dormant Consumer, contains associates with longer relationship times and moderate levels of financial product usage. This segment has excellent potential to be explored, especially in offering differentiated cards, aiming to convert those who use cards from other financial institutions. The cooperative can focus on *cross-selling* and *up-selling* strategies to maximize the value of these clients.

Cluster 2, named Exclusive Consumer, is an extremely select group, distinguished by an excellent relationship and reciprocity with the cooperative. This profile can be further enhanced through the offering of more exclusive and personalized cards and financial products, which reinforce their status and loyalty to the institution. Maintaining and deepening this relationship are crucial for the cooperative.

Finally, Cluster 3, characterized as Intensive Consumer, includes the associates who generate the largest part of the transacted value. For this group, the strategy should focus on more robust benefit and scoring programs, further encouraging the use and loyalty to the cooperative’s card. The personalization of rewards and the communication of added value are essential to maintain the high engagement of these clients.

Comparison and Implications of the Models

The comparison between the two- and five-cluster models revealed that, although the two-cluster model offers a macro view, distinguishing between standard and intensive consumers, the five-cluster configuration provided superior strategic granularity. The five-cluster model allowed for the identification of more specific groups, such as the “Dormant Consumer,” representing an immediate opportunity to convert spending from other institutions to the cooperative, and the “Exclusive Consumer,” who demands high-value offers.

Both approaches, however, are valuable for the cooperative. The two-cluster model can be used for initial screening, quickly identifying high-value members for personalized offers and more generic communication. The five-cluster model, on the other hand, allows for the development of highly targeted marketing and communication strategies, optimizing the offering of complementary products and strengthening institutional relationships, in line with the study’s objective. Segmentation, therefore, fulfills its role by capturing the member’s real behavior, providing a guide for personalized communications and reward programs that increase the cooperative’s competitiveness in the financial market.

4. Conclusion

This study sought to identify the groupings of individual cardholder members of a cooperative, as well as their distinctive characteristics, with the purpose of supporting the refinement of marketing and communication strategies for offering differentiated cards and strengthening institutional relationships. To this end, the Python language was employed for data processing, standardization with StandardScaler, and dimensionality reduction by Principal Component Analysis (PCA), which retained five components explaining 73.07% of the variance. Subsequently, the K-means algorithm was applied, with the evaluation of scenarios of two and five clusters.

The analysis with two clusters revealed a macro distinction between Standard Consumers, comprising the majority of the base with lower purchase frequency and transacted volume, and Intensive Consumers, a minority group, yet responsible for a substantial portion of the total transacted value and characterized by greater loyalty and engagement with the cooperative. The five-cluster solution, in turn, provided deeper strategic granularity, allowing the identification of profiles such as Potential Consumer, Dormant Consumer, Exclusive Consumer, and Intensive Consumer. It was noted that the Dormant Consumer, despite having a consolidated relationship and resources within the cooperative, still uses cards from other institutions, representing a significant opportunity for spending conversion. This detailed segmentation offers valuable subsidies for the cooperative to develop personalized marketing and communication strategies, optimize the offering of complementary products, and strengthen institutional relationships, thereby increasing its competitiveness in the financial market.

Bibliographic References

Abdi e William, 2010 [Referência completa não encontrada no documento original]

Abdulhafedh, A. 2021. Incorporating “K-means”, Hierarchical Clustering and PCA in Customer Segmentation. Journal of City and Development 3(1): 12–30.

Banco Central do Brasil [BACEN]. 2022. Boxe 3 – Perfil de utilização de cartões de crédito no Brasil. Disponível em: <https://www.bcb.gov.br/content/publicacoes/boxe_relatorio_de_economia_bancaria/reb2022b3p.pdf>. Acesso em: 08 de setembro de 2025.

Banco Central do Brasil [BACEN]. 2023. Cartilha – Cartão de crédito. Disponível em:<https://www.bcb.gov.br/pec/appron/cartilha.pdf>. Acesso em: 08 de setembro de 2025.

Fávero, L.P.; Belfiore, P. 2024. Manual de Análise de Dados: Estatística e machine learning com Excel®, SPSS®, Stata®, R® e Python®. 2ed. LTC, Rio de Janeiro, RJ, Brasil.

Géron, A. 2021. Mãos à Obra: Aprendizado de Máquina com Scikit-Learn, Keras & TensorFlow: Conceitos, Ferramentas e Técnicas para a Construção de Sistemas Inteligentes. 2ed. Alta Books, Rio de Janeiro, RJ, Brasil.

Güçdemir, H.; Selim, H; 2015. Integrating multi-criteria decision making and clustering for business customer segmentation. Industrial Management & Data Systems 115(6): 1022-1040.

Harrison, M. 2020. Machine Learning – Guia de Referência Rápida: Trabalhando com Dados Estruturados em Python. 1ed. Novatec Editora, São Paulo, SP, Brasil.

Kansal, T.; Bahuguna, S.; Singh, V.; Choudhury, T. 2018. Customer Segmentation using “K-means” Clustering. In: International Conference on Computational Techniques, Electronics and Mechanical Systems (CTEMS), 2018, Belgaum, Índia. Anais… p. 135-139.

Krishnan, G.U. 2024. Customer Segmentation for Credit Card Customers: A Comprehensive Guide. Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol 10(5): 286–296.

Umuhoza, E.; Ntirushwamaboko, D.; Awuah, J.; Birir, B. 2020. Using Unsupervised Machine Learning Techniques for Behavioral-based Credit Card Users Segmentation in Africa. SAIEE Africa Research Journal 111(3): 95-101.

Wang, G. 2025. Customer segmentation in the digital marketing using a Q-learning based differential evolution algorithm integrated with “K-means” clustering. Plos One 20(2): e0318519.

Wang, X.; Xu, Y. 2019. An improved index for clustering validation based on Silhouette index and Calinski-Harabasz index. IOP Conference Series: Materials Science and Engineering 569(5): 052024.

Article originating from the Final Course Work 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