Consistency, Interpretability, and Governance of Automated Valuation Models: Evidence from Spanish Institutional Residential Portfolios Under Regulatory Transition

Order ECM/599/2025 establishes the first statutory basis for Automated Valuation Models (AVMs) in Spanish regulated residential valuation, embedding transparency and traceability requirements alongside the EU Artificial Intelligence Act and Basel III/CRR3 prudential standards. Operational guidance from the Banco de España remains pending, leaving three questions unresolved: whether AVMs achieve sufficient consistency with existing reference valuations for lending decisions, whether their internal structure can be recovered with enough fidelity to satisfy traceability requirements, and whether practitioners navigating the reform regard current tools as operationally fit. This study addresses these questions through three complementary empirical tracks applied to institutional residential portfolios supplied by an anonymised Spanish real estate services provider. Track A tests inter-valuation agreement between an automated-style Broker Opinion of Value and a client-provided Bank AV reference across 1,107 Barcelona-Madrid dwellings, finding a median absolute divergence of 14.98 per cent that is overwhelmingly property-idiosyncratic rather than systematically structured. Track B applies TreeSHAP to a LightGBM surrogate trained to reproduce 4,132 professionally determined Market Values, recovering a location-dominated (55.5 per cent of attribution) non-linear hedonic structure at a documented out-of-sample fidelity of R-squared 0.64. Leave-one-province-out validation reveals that this fidelity does not transfer to markets absent from training, with error rising two- to fourfold relative to random-split baselines. Track C situates these findings against practitioner perceptions gathered through a structured questionnaire. We derive governance recommendations for regulators, lenders, and AVM providers, including geographic validation criteria, hybrid human-algorithm protocols, and disclosure standards for surrogate explainability.