April 13, 2026
Clustering of Default Profiles in Vehicle Financing
Sandra Ferreira Almeida Lopes; Miguel Ângelo Lellis Moreira
Summary prepared by the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing.
Default in vehicle financing constitutes one of the most significant obstacles to maintaining stability in the Brazilian automotive credit system, generating severe impacts on both the financial health of institutions and the economic vulnerability of consumers (Silva & Oliveira, 2021). In an environment characterized by the rapid digitalization of banking services and the expansion of credit offerings, the management of non-payment risk has become a strategic priority, requiring that lending policies be based on increasingly precise data analyses (Banco Central do Brasil, 2023). The complexity of delinquent consumer behavior is influenced by a multiplicity of factors ranging from basic demographic characteristics to socioeconomic conditions and specific contractual terms (Lima & Costa, 2020). Variables such as monthly income bracket, education level, borrower’s age, total financed amount, number of installments, and applied interest rate play crucial roles that, when analyzed in isolation, may not reveal the full extent of the risk involved. The interaction between these elements demands the use of advanced analytical approaches capable of identifying latent patterns and segmenting risk profiles with a level of detail superior to conventional statistical methods (Breiman, 2001).
The field of Data Science, driven by the development of machine learning algorithms, offers robust tools for predictive analysis and for supporting strategic decision-making in uncertain scenarios (Provost & Fawcett, 2013). Within this ecosystem, unsupervised learning stands out for its ability to group observations based on intrinsic similarities, without the need for prior labels or classifications (Jain, 2010). Algorithms such as K-means and DBSCAN allow for the exploration of large data volumes to discover hidden structures, enabling financial institutions to identify groups of customers with homogeneous behaviors. This segmentation is fundamental for personalizing collection strategies and for refining credit scoring models, contributing to more efficient portfolio management (Souza & Andrade, 2022). However, there is a noted lack of studies in the national literature that apply clustering techniques specifically aimed at delinquency in the automotive sector, as most research focuses on supervised classification models. This gap justifies the need for investigations that use the methodological rigor of Data Science to explore the descriptive richness of contractual and demographic data, aiming to prevent financial losses (Silva & Oliveira, 2021).
The theoretical foundation of this study is based on the premise that delinquency is not an isolated phenomenon, but the result of a combination of structural and conjunctural factors. Cluster analysis, or clustering, emerges as the ideal technique to deal with this multidimensionality, as it seeks to maximize homogeneity within each group and heterogeneity between them. The central objective lies in identifying recurring profiles of delinquent customers, relating variables such as age, income, and financed amount to understand the dynamics that lead to the breach of contractual obligations. By mapping these profiles, it becomes possible to anticipate risks and formulate more resilient credit policies, adapted to the particularities of each population segment.
The methodology adopted is characterized as a quantitative case study, with exploratory and explanatory objectives, structured to investigate the cause-and-effect relationships in vehicle delinquency (Gil, 2019). The research was developed remotely, with Piracicaba, São Paulo, as its operational base, during the period between March and June 2025. The target population consisted of customers who established vehicle financing contracts between 2020 and 2024, encompassing a period of significant economic fluctuations in Brazil. The final sample consisted of 9,300 valid records, selected through non-probabilistic convenience sampling, ensuring the integrity and diversity of the variables necessary for robust statistical analysis.
The data collection process used secondary sources from financial institutions operating in national territory. The variables selected for analysis included monthly income range, education level, client age, contract duration, total financed amount, number of installments contracted, previous default history, applied interest rate, and down payment amount. In strict observance of the General Data Protection Law (Law No. 13,709/2018), all records were subjected to a rigorous anonymization process before any manipulation, ensuring the confidentiality of personal information and the ethical compliance of the research.
Data processing and analysis were performed using the Python programming language, with the support of the Pandas and NumPy libraries for matrix manipulation and exploratory analysis (McKinney, 2017). The preprocessing workflow was detailed and included several critical steps to ensure the quality of the analytical inputs. Initially, duplicate records were removed and structural inconsistencies in the database were corrected. Missing values were handled using statistical imputation techniques, where numerical variables received the mean or median, while categorical variables were filled with the mode, depending on the observed distribution. To ensure that variables with different scales did not distort the clustering algorithm’s results, Z-score normalization was applied, which rescales the data to have a mean of zero and a standard deviation of one. Categorical variables were transformed into numerical formats using one-hot encoding, allowing their integration into Euclidean distance calculations. Furthermore, outlier identification was conducted through interquartile range analysis and boxplot visualization, resulting in the treatment of extreme values that could compromise centroid formation.
The modeling stage used the K-means algorithm, selected for its computational efficiency and suitability for distance-based segmentation problems (MacQueen, 1967). K-means operates through an iterative process that seeks to minimize the sum of squared distances between data points and the centroids of their respective clusters. To determine the optimal number of clusters, the elbow method was applied, which monitors the variation in the sum of internal squared errors in relation to the increase in the number of groups. The final decision on the number of segments was refined by the silhouette index, which assesses the quality of separation between clusters and the internal cohesion of each (Rousseeuw, 1987). The silhouette index ranges from -1 to 1, with values close to one indicating that observations are well-allocated to their groups and distant from neighboring groups.
Complementarily, Multiple Correspondence Analysis was used to validate the associations between categorical variables and the generated groupings. The technique allowed for the construction of a binary matrix and the calculation of total inertia, which represents the variability explained by the factorial axes. The inertia decomposition enabled the graphical projection of clusters onto a two-dimensional plane, facilitating the visual interpretation of statistical distances. To confirm the significance of the associations found, the Chi-squared test and Cramér’s V coefficient were applied, which measure the strength of the relationship between nominal variables. This entire methodological apparatus was designed to transform raw data into strategic knowledge, enabling a deep understanding of default profiles.
The analysis of the results revealed an overall default rate of 21.4% in the sample of 9,300 clients, highlighting the relevance of the problem in the context of automotive credit. Initial descriptive statistics showed an average income of R$ 4,280, with a considerable standard deviation of R$ 2,150, and an average age of 38 years, ranging from 18 to 75 years. The average financed amount was R$ 32,400, with an average payment term of 42 installments. These indicators suggest a heterogeneous population, which reinforces the need for segmentation to capture the nuances of financial risk (Lima & Costa, 2020).
The application of the elbow method indicated a clear inflection in the inertia curve when the number of clusters reached four, suggesting that this would be the ideal balance between model complexity and variance explanation capability. This choice was corroborated by the average silhouette index of 0.62, a value considered satisfactory for multidimensional data, indicating that the formed groups have consistent separation (Rousseeuw, 1987). Segmentation resulted in four distinct profiles, each presenting specific socioeconomic and behavioral characteristics that directly influence the probability of default.
The first group identified, named Cluster 1, presented the highest risk of default, with a rate of 38%. This profile is predominantly composed of young clients, with an average age of 26 years, and an average monthly income of R$ 2,100. The average financed amount for this group was R$ 42,800, with an extended payment term of 54 installments and an average interest rate of 2.4% per month. The combination of low income, high financed amount, and high interest rates creates a scenario of high financial vulnerability. These findings are in line with the studies by Silva & Oliveira (2021), which indicate that a lack of financial experience combined with long-term commitments significantly increases the probability of contractual default among younger individuals.
In contrast, Cluster 2 revealed itself as the lowest risk profile, with a default rate of only 9%. This group has an average age of 37 years and an average income of R$ 5,200. The average financed amount was R$ 28,300, distributed over 48 installments, with an interest rate of 1.9%. The greater financial stability of this segment, reflected in a higher income and a more moderate financed amount relative to payment capacity, justifies the low rate of delays. Age maturity also appears to act as a risk mitigation factor, possibly associated with a more consolidated credit history.
Cluster 3 presented a default rate of 15%, characterized by clients with an average age of 63 years and an average income of R$ 3,800. The average financed amount was the lowest among all groups, totaling R$ 18,600, with a reduced number of installments, an average of 22 months, and the lowest interest rate, 1.6%. This profile suggests a more conservative consumption behavior, focused on lower-value vehicles or with a larger down payment, which reduces risk exposure. However, the 15% default rate indicates that, even under favorable contractual conditions, external factors can affect this group’s payment capacity.
Finally, Cluster 4 presented a default rate of 22%, with an average age of 41 years and the highest average income in the sample, R$ 11,200. Despite the high income, this group financed high amounts, averaging R$ 65,000, over 60 installments and 2.2% interest. The high risk in this segment, despite apparent financial capacity, may be associated with a high commitment of income to luxury goods or sensitivity to economic variations affecting high-income professionals. This result demonstrates that income alone is not an absolute predictor of adherence, and should be analyzed in conjunction with the debt amount and contractual term.
Statistical validation through Multiple Correspondence Analysis confirmed the robustness of the groupings, with the first two factorial axes explaining 72.3% of the total data variance. The projection of the clusters onto the factorial plane showed that income, down payment, and number of installments are the variables with the greatest discriminatory power among the groups. The Chi-square test revealed a significant association (p < 0.01) between the clusters and the income and down payment ranges. Cramer’s V coefficient showed values of 0.42 for income, 0.39 for down payment, and 0.47 for the number of installments, indicating moderate to strong associations. These results confirm that the groupings are not random, but reflect real financial behavior structures present in the database (Souza & Andrade, 2022).
The discussion of the results allows us to infer that default in vehicle financing is a multifactorial phenomenon. The concentration of high interest rates and long terms in Cluster 1 suggests that financial institutions try to compensate for the risk inherent in the young and low-income profile with higher rates, which, paradoxically, can overload the customer’s budget and lead to non-payment. On the other hand, Cluster 4 shows that the appetite for high-value credit, even among affluent customers, carries considerable risk that cannot be ignored. Data visualization reinforced that profiles with lower down payment amounts and a higher number of installments tend to group in the segments with higher default rates, validating the practice of requiring larger down payments as a risk mitigation strategy.
Despite the consistency of the findings, it is necessary to acknowledge the limitations of this study. The model was based exclusively on historical and anonymized data, which prevents the inclusion of real-time behavioral variables, such as overdraft usage or checking account transactions, which could enrich the analysis. Furthermore, the K-means algorithm, by its nature, assumes that clusters have spherical shapes, which may not capture more complex or overlapping data distributions. The absence of subjective variables, such as payment intention or consumer risk perception, also limits the depth of the sociological explanation of the phenomenon. For future research, it is recommended to explore algorithms such as DBSCAN, which handles noise and irregular shapes better, and to integrate dynamic data sources to increase the predictive capacity of the models.
It is concluded that the objective was achieved, as the application of unsupervised machine learning techniques allowed for the identification and detailed characterization of four profiles of defaulting customers in vehicle financing. The segmentation demonstrated that credit risk is decisively influenced by the interaction between monthly income, borrower’s age, and contractual conditions, such as the number of installments and the interest rate. Statistical validation through Multiple Correspondence Analysis and association tests confirmed the consistency of the formed groups, highlighting that the use of clustering algorithms provides valuable subsidies for the formulation of more precise and personalized credit granting policies. The study reinforces the importance of Data Science as a strategic tool for financial risk mitigation and for the improvement of portfolio management in highly competitive environments.
Bibliographic References:
Banco Central do Brasil [BACEN]. 2023. Relatório de Economia Bancária. Disponível em: <https://www.bcb.gov.br>. Acesso em: 25 jun. 2025.
BRASIL. 2018. Lei nº 13.709, de 14 de agosto de 2018. Lei Geral de Proteção de Dados Pessoais (LGPD). Diário Oficial da União: seção 1, Brasília, DF, 15 ago. 2018.
Breiman, L. 2001. Statistical modeling: The two cultures. Statistical Science 16(3): 199-231.
Gil, A.C. 2019. Métodos e técnicas de pesquisa social. 7. ed. Atlas, São Paulo, SP, Brasil.
Jain, A.K. 2010. Data clustering: 50 years beyond K-means. Pattern Recognition Letters 31(8): 651–666.
Lima, R.; Costa, M. 2020. Fatores determinantes da inadimplência em crédito pessoal: Uma análise empírica. Revista Brasileira de Finanças 18(2): 45-62.
MacQueen, J. 1967. Some methods for classification and analysis of multivariate observations. In: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, 1967, Berkeley, California, USA. Anais… p. 281–297.
McKinney, W. 2017. Python for Data Analysis: Data Wrangling with Pandas, NumPy, and IPython. 2. ed. O’Reilly Media, Sebastopol, CA, EUA.
Provost, F.; Fawcett, T. 2013. Data Science for Business: What You Need to Know About Data Mining and Data-Analytic Thinking. O’Reilly Media, Sebastopol, CA, EUA.
Rousseeuw, P.J. 1987. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics 20: 53–65.
Silva, A.; Oliveira, T. 2021. Inadimplência em financiamentos de veículos: Uma abordagem exploratória. Revista de Administração Financeira 13(1): 77-92.
Souza, F.; Andrade, L. 2022. Aplicação de algoritmos de clusterização na análise de risco de crédito. Revista Brasileira de Inteligência Artificial 9(3): 112-130.
Executive summary from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq
Learn more about the course; click here:
