Data Science Analytics
Financial Institutions
Technology
August 05, 2025
AI and default in Brazil: new tools
The use of predictive models is already revealing hidden layers in the data, allowing for the identification of patterns and the anticipation of behaviors
The Brazilian economic scenario in the first semester of 2025 showed an increase in defaults, a trend that has been observed since 2022 and has been a recurring concern for governments and financial institutions over the past decades (Leite Filho, 2025).
This indicator reflects, in part, the country’s economic fluctuations, credit policies, and socioeconomic conditions. On a global scale, economic stagnation and trade tensions have increased the risk of financial obligation defaults. Zhao et al. (2025), for example, observed a notable increase — of 15,000% — in corporate debt defaults in China between 2014 and 2023.
For financial institutions, the delay or failure in the payment of bills can generate substantial losses and, in cases of a large volume of overdue loans, can threaten financial stability (Liu, 2025).
In this context, the prediction of default risk becomes essential for banks and other credit entities. Understanding how to mitigate this risk is fundamental to ensuring stable and healthy economic functioning. This article explores how Artificial Intelligence (AI) and Machine Learning (ML) are being applied to manage the complexity of default, seeking to optimize the understanding and management of credit risk in Brazil.
AI and credit analysis
The growing scale of personal loans and the complexity of credit data, rapidly growing in the era of internet development and big data tools, have made the accurate assessment of credit scoring and personal loan default risk a central topic in the financial field (Liu, 2025).
Traditionally, credit risk analysis relied on historical data and static indicators, which hindered the capture of market dynamics and nonlinear characteristics (Yi, 2025). In contrast, the evolution of machine learning, especially the application of Recurrent Neural Networks (RNNs) and their variants in time series data analysis, has provided new perspectives and methods for measuring credit default risk.
AI has proven to be a promising alternative, not only for predicting default but also for redesigning the way credit risk is understood in Brazil. In the context of higher education, for example, entities such as the Community Institutions of Higher Education (ICES) in Rio Grande do Sul have undertaken efforts to maintain economic-financial balance in the face of student reduction and macroeconomic conditions (Lima et al., 2025). These institutions, characterized by high management complexity and detailed financial planning, demonstrate that careful working capital management is fundamental for the accuracy and security of financial processes, mitigating default.
AI’s ability to integrate and process a vast range of data, beyond payment history or declared income, allows for the capture of complex relationships between demographic, behavioral, contextual, and even psychometric variables. This forms a broader and more dynamic view of customers’ risk profiles. The accuracy in predicting loan default has been crucial for the stability of the financial market and the prevention of systemic risks (Huang et al., 2025).
The use of AI-based predictive models is already revealing hidden layers in the data, allowing the identification of patterns and the anticipation of behaviors before the delay consolidates into full default. With this, credit analysis, previously reactive, starts to function as a proactive approach, with the potential to directly impact the profitability and capital management of financial institutions (Bhandary & Ghosh, 2025). AI, in this context, becomes an indispensable tool for navigating a complex and changing economic environment.
Models that learn from data
The main transformation brought by AI to the study of default lies in its ability to integrate and process a vast range of data, which were previously little considered in traditional models. It’s not just about looking at payment history or declared income; machine learning models can capture complex relationships between demographic, behavioral, contextual, and even psychometric variables— such as impulsivity, risk aversion, or tendency to postpone bill payments —, forming a much broader and dynamic view of each client’s risk profile (Liu, 2025).
In comparative tests, algorithms such as Random Forest, XGBoost, and recurrent neural networks have shown superior results to classical methods, such as logistic regression or standardized score. For example, in an empirical analysis on credit card default prediction in Taiwan, Bhandary & Ghosh (2025) observed that modern machine learning methods, including XGBoost, Random Forest, and Deep Neural Networks (DNN), outperformed traditional statistical methods in predictive performance. The Deep Neural Network (DNN) demonstrated the best overall performance compared to other machine learning models, such as Logistic Regression, Support Vector Machine (SVM), and Naïve Bayes, as noted by Liu (2025). In the industrial sector, Long et al. (2025) observed that the Vector Autoregressive-Gated Recurrent Unit (VAR-GRU) model, which combines machine learning with macroeconomic factors, presents optimal predictive capacity for corporate bond defaults at one, three, and six-month horizons.
Furthermore, enhancements in models based on Recurrent Neural Networks (RNNs) have also shown promise. Yi (2025) proposed a model based on Gated Recurrent Unit (GRU) with structural improvements, such as the introduction of the Focal Loss function and the fuzzy clustering algorithm, to capture temporal patterns in credit series. This new model not only surpassed classical statistical techniques, such as the Logistic and Copula models, with an accuracy of 96.53%, but also proved especially efficient in predicting future variations in payment behavior, in addition to being significantly faster in execution.
Non-financial characteristics
The use of non-financial data has been crucial in enhancing risk prediction. Huang et al. (2025) utilized 38 multidimensional non-financial characteristics of SMEs in China, such as company registration information and news sentiment, to predict the default risk of promissory notes. They found that company age, the number of legal disputes, the proportion of negative news, the count of previous year’s defaults, and the number of individuals subject to consumption restrictions were significantly correlated with default. Text analysis, in turn, has also proven valuable; Wu et al. (2025) observed that an LLM like ChatGPT, when analyzing loan reviews, can significantly improve default predictions, generating higher profitability compared to human-written text.
This paradigm shift allows default to be modeled as a process with multiple causes and possible trajectories, offering means to recognize paths and, more importantly, to intervene before the worst-case scenario materializes.
From prediction to prevention
The greatest potential of artificial intelligence applied to credit does not lie solely in the ability to predict who will be late on a payment; the real value of the technology emerges when it is used to avoid such delays from occurring. This approach has been adopted by many financial institutions.
Machine learning models are effective in the early identification of customers at higher risk of default, as demonstrated by Liu (2025). By predicting customer behavior, these models allow institutions to offer tailored renegotiations, adjust credit limits, or even modify collection cycles. In this sense, AI functions as a preventive monitoring system.
The tool has also been applied to adapt interventions in real time. For example, the IGRU-FCM model, developed by Yi (2025), not only accurately assesses credit default risk but also performs statistical analysis in just 59 milliseconds (1 ms = 0.001 s). This speed is crucial for financial markets that require real-time monitoring.
In another context, Wang & Duan (2025) observed that Chinese commercial banks, facing economic policy uncertainty, adopt a more diversified loan structure to mitigate credit default risks, and this relationship remains robust even after rigorous endogeneity tests (a situation where an explanatory variable in a regression model is correlated with the model’s error term).
This change in posture redefines the role of predictive models. Instead of functioning as filters that exclude consumers based on rigid criteria, they become instruments for offering financial solutions that are more humane and sustainable. In this context, AI acts as a partner to both the lender and the borrower.
Exclusion risks
Although AI brings significant advances to credit management, it also raises a series of ethical and operational concerns that cannot be ignored. When applied without clear criteria of transparency, accountability, and oversight, AI can reproduce or even deepen existing inequalities in the financial system.
One of the most discussed risks is algorithmic bias. Models trained with historical databases tend to repeat exclusion patterns, penalizing, for example, individuals who have never had access to formal credit or who live in regions with low banking supply. Even seemingly neutral variables, such as postal code or cell phone type, can function as proxies for income or education level, which can lead to the indirect exclusion of vulnerable groups. In fact, Wang and Duan (2025) warn about this phenomenon when comparing traditional statistical approaches and AI models, suggesting that algorithmic sophistication does not eliminate the risk of biased decisions.
The application of AI in the financial system, therefore, cannot be treated solely as a technical innovation. It is a political and economic choice that defines who will have access to credit, under what conditions, and with what consequences. The challenge is not just to make credit more efficient, but to ensure that it is fairer, more transparent, and more sustainable. This requires a combination of technology and governance, in which the explainability of AI models is essential. Models that allow us to understand the weight of each variable in the final decision open up space for more effective audits and for the construction of more equitable models.
Final Considerations
Default in Brazil in 2025 reflects not only a challenging economic moment, but also the macroeconomic dynamics and credit models adopted in the market, according to the analysis by Leite Filho (2025). In this scenario, AI emerges as a powerful tool to understand risk in a more profound and personalized.
Recent studies highlight the algorithms’ ability to identify patterns that escape traditional analyses and to propose more effective solutions, both for predicting and preventing non-payment. Machine learning models, such as XGBoost, Random Forest, and Deep Neural Networks (DNN), have demonstrated superior performance in predicting default, as shown by Bhandary and Ghosh (2025) and Liu (2025). Furthermore, innovations like the IGRU-FCM model, which integrates GRU with Focal Loss and fuzzy clustering, not only improve the accuracy of credit risk assessment but also significantly optimize computation time, which is crucial for real-time monitoring (Yi, 2025).
However, technical sophistication does not eliminate ethical responsibility. The use of AI in the financial system needs to be accompanied by control mechanisms, explainability, and fairness. Algorithms are not neutral; they reflect the data with which they were trained and the choices of those who built them. Technology has much to contribute to a smarter credit system, but this will only be possible if it is used with awareness and purpose. Instead of widening the gap between those who have access and those who are left behind, AI should be oriented towards building bridges, promoting a future of credit that depends less on the ability to predict a delay and more on the willingness to prevent it from happening.
| To access the references of this text click here. |
Who wrote this column
José Erasmo Silva
