BCB Requires Explainability in Real Estate Credit Scores
In June 2026, the Central Bank of Brazil published Resolution BCB No. 4,657. The text requires that all real estate credit scoring models be "explainable and auditable" by regulatory bodies. Penalties for non-compliance can reach 2% of the financial institution's revenue (Source: Central Bank of Brazil).
The measure is not merely technical. It represents a paradigm shift in the real estate credit sector, which handles over R$ 1 trillion in active portfolios in the country. Machine learning models that operate as black boxes—where neither the bank nor the client knows exactly why a loan was denied—are on borrowed time.
What Resolution BCB No. 4,657/2026 Requires in Practice
The resolution does not prohibit the use of artificial intelligence in real estate credit. On the contrary, it creates a regulatory framework that encourages more robust and transparent models. The main requirements are:
- Mathematical Explainability: The model must generate a comprehensible explanation for each credit decision. A numerical score is not enough. It is necessary to point out which variables (income, history, property location) most influenced the final score.
- Mandatory External Audit: Financial institutions must hire independent audits to validate the models every 12 months. The report must be sent to the Central Bank of Brazil (BCB).
- Bias and Discrimination Testing: It is mandatory to test whether the model penalizes specific groups (by race, gender, geographic region). If bias is identified, the institution must correct the algorithm or suspend its use.
- Progressive Penalties: Fines of up to 2% of revenue for institutions that fail to comply. In case of recurrence, the use of the model may be permanently prohibited.
The resolution applies to banks, fintechs, and credit cooperatives operating with real estate financing. Institutions like Creditas and Serasa Experian have already announced compliance with the new rules.
Practical Results: Default Rates Fell 18% with Explainable Models
One of the BCB's central arguments for the new regulation is that explainable models are also safer models. A study by ABECIP (Brazilian Association of Real Estate Credit) titled "Impact of Algorithmic Transparency on Brazilian Real Estate Credit" (2026) confirms this thesis.
In institutions that had already adopted explainable AI models before the resolution, default rates on real estate credit fell by 18% in 2026 (Source: ABECIP, 2026). The reason is simple: when a bank understands why a model denied credit, it can recalibrate variables and avoid errors that lead to defaults.
The table below, based on the ABECIP study, compares indicators before and after the implementation of explainable models in a sample of 12 major Brazilian financial institutions:
| Indicator | Before Explainable AI (2025) | After Explainable AI (2026) | Change |
|---|---|---|---|
| Default rate (90+ days) | 3.2% | 2.6% | -18.7% |
| Appeals against credit denial | 12,400/month | 8,100/month | -34.6% |
| Average credit approval time | 8.5 days | 5.2 days | -38.8% |
| Customer satisfaction score (NPS) | 62 | 74 | +19.3% |
According to the ABECIP study, transparency did not compromise efficiency. On the contrary: explainable models reduced approval time and improved the customer experience.
How Fintechs and Banks Are Adapting
Implementing the resolution is not trivial. Deep learning models, such as deep neural networks, are naturally opaque. To meet the explainability requirement, institutions are turning to techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations).
Serasa Experian, for example, announced that it has reformulated its real estate scoring algorithms to incorporate explanation layers. Each score now comes with an "influence report" showing the three main variables that impacted the final score.
Creditas, a real estate credit fintech, adopted a hybrid approach: explainable machine learning models (such as gradient boosting with decision trees) combined with traditional business rules. The company states that computational costs increased by 12%, but the correct approval rate rose by 22%.
Traditional banks, such as Itaú and Bradesco, are investing in internal "algorithm audit" teams. These teams are responsible for generating explainability reports and ensuring that models do not exhibit bias.
Conclusion
Resolution BCB No. 4,657/2026 is not an obstacle to innovation. It is a watershed moment. It forces the real estate credit market to abandon black-box models and adopt systems that combine mathematical efficiency with regulatory transparency. ABECIP data shows that this choice is not only ethical—it is profitable. Institutions that have already migrated to explainable models have reduced defaults, accelerated approvals, and improved customer satisfaction. For those seeking real estate financing, the resolution brings direct benefits: the right to a clear justification in case of credit denial, allowing the consumer to understand exactly what needs improvement to have the loan approved in the future. The 34.6% reduction in appeals against credit denial (Source: ABECIP, 2026) shows that explanations are being accepted more frequently. Another important point is protection against algorithmic discrimination, with bias tests that prevent models from penalizing specific groups without technical justification. The compliance deadline for all market players ends in December 2026. Those who are not ready by then will face severe fines and, potentially, a ban on operating with real estate credit in Brazil. The BCB's message is clear: in the future of credit, the algorithm must be as transparent as the contract the client signs.
