Reliable Prediction Intervals for Automated Rental Valuations

Automated valuation models (AVMs) are widely used for large-scale residential rent appraisal, yet standard models do not provide predictive uncertainty measures with guaranteed out-ofsample coverage at prespecified nominal levels, creating risks for institutional decision-making in valuation, risk management, and policy design. Using a transaction-level dataset covering the Flemish rental market in Belgium, we study AVM performance and uncertainty quantification in a large-scale, heterogeneous, and feature-poor setting, where only location, property type, energy performance, number of bedrooms, and rent prices are observed. We show that industry-standard point-prediction accuracy can be achieved by exploiting non-linear spatial structure using coarse geospatial units such as boroughs. For uncertainty quantification, we compare ensemble quantile regression and Inductive Conformal Prediction (ICP). While both improve empirical coverage, ICP is preferred as it guarantees finite-sample marginal coverage without distributional assumptions at substantially lower computational cost. Conditioning ICP calibration on bedroom count (Mondrian ICP) yields the largest efficiency gains, reducing 95% coverage prediction interval width by up to 5.3% relative to absolute residual split conformal prediction. Overall, our results demonstrate that valuation uncertainty can be materially reduced in large-scale, feature-poor housing data with minimal additional modeling complexity.