Increasing the Transparency of Pricing Dynamics in the US Commercial Real Estate Market with Interpretable Machine Learning Algorithms

Abstract

This study proposes a holistic framework for the practical use of automated valuation models (AVMs) in a commercial real estate context that considers both accuracy and interpretability. The authors train a deep neural network (DNN) on a unique sample of more than 400,000 property-quarter observations from the NCREIF Property Index and perform model-agnostic analysis using Shapley Additive exPlanations (SHAP) to provide ex post comprehensibility of the algorithm’s prediction rules. They further assess the extent to which the inner workings of the DNN follow an economic rationale and set out how the proposed methods can add to the understanding of pricing processes in institutional investment markets. By addressing the caveats and illustrating the potential of machine learning in the field of commercial real estate, this article represents another important pillar in the practical use of AVMs.