The missing link: Comparing mass appraisal techniques and their implications for urban land valuation policies in Latin America

From a public policy perspective, this paper examines technical alternatives for urban land mass appraisal, employing quantile regression forest (machine learning), ordinary kriging (geostatistics), and a linear model. These techniques significantly impact automated mass appraisal policies. The study was conducted in Córdoba, Argentina. Although machine learning models offer advantages in out-of-sample prediction, their black-box nature complicates public policy. Linear models enable the reconstruction of the value assignment process, but they have high prediction variance. Ordinary kriging uses only sample geographic coordinates to predict land values, thus eliminating the cost of generating independent variables. The results show that machine learning provides lower out-of-sample errors and a consistent value map that reflects the structure of urban land values. Geostatistics yield acceptable results, but they struggle to accurately capture urban barriers. The linear model performs poorly outside the sample, with high dispersion in similar environments. This study emphasizes the importance of considering the strengths and limitations of technical alternatives for mass appraisal policies, particularly in Latin America, where multilateral credit agencies often fund and control such policies. Implementing successful methodologies requires analyzing the capacity of local administrations to absorb innovation. Future research should focus on hybrid models that combine techniques