Automated valuation models (AVMs) provide efficient means for local government to determine fair and equitable property taxes, for mortgage providers to limit risks, and for asset owners to make complex investment decisions. Traditionally, AVMs have been econometric models, such as linear regression models. However, recent advancements in the field of machine learning (ML) have opened up a new, and in many fields successful toolbox, providing additional methods for the same data, as well as approaches to access new sources of information and to create new variables.
An important distinction between traditionally applied methods and more recently introduced techniques lies in the structure definition of a model. Econometric models require a model specification – transformation of variables, selection of functional form, interaction effects, and distributional assumptions – prior to estimating parameter values, whereas most ML algorithms determine the model’s structure and parameter values simultaneously (Athey, 2018).
This fundamental difference has theoretical consequences that are naturally reflected in practical applications. The main goal of this paper is therefore to discuss how ML algorithms compare to econometric models for residential real estate valuation in theory and to show what these theoretical differences mean in practice.
In Section 2 we discuss the position of ML algorithms within the landscape of AVMs by comparing econometric models and ML algorithms from a theoretical perspective. Section 3 shows two different ML applications within residential property valuation to highlight the advantages and disadvantages of ML algorithms. Finally, Section 4 concludes and provides routes for future research.
Authors: David Kroon and Marc Francke