PERFORMANCE, ACCURACY AND EQUITY IN AUTOMATED VALUATION MODELS: APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN BRAZIL UNDER THE PERSPECTIVE OF THE SDGS

Abstract

This study evaluates the performance, interpretability, and algorithmic fairness of Automated Valuation Models (AVMs) applied to urban residential properties in Itabuna, Bahia, Brazil. A real and anonymized dataset comprising 100 properties marketed or transacted between 2023 and 2025 was modeled using Multiple Linear Regression (MLR) and an Artificial Neural Network (ANN). Physical, locational, and market attributes were subjected to normalization, outlier treatment, categorical encoding, and multicollinearity control. Model performance was assessed through R², mean absolute error (MAE), root mean square error (RMSE), residual analysis, and the Shapiro–Wilk test. Both approaches achieved high predictive accuracy, with R² values above 0.90 and MAE below 5%. The ANN slightly outperformed MLR, reaching R² = 0.922, MAE = BRL 10,830, and RMSE = BRL 13,760. However, MLR provided greater interpretability, auditability, and legal-technical defensibility, which are essential in judicial and regulatory valuation contexts. Residual analysis indicated no systematic bias across neighborhoods or price ranges. The findings support a hybrid approach that combines statistical transparency with AI-based predictive capacity, contributing to more accurate, equitable, and auditable valuation practices aligned with SDGs 10, 11, and 16.