Automated models for value prediction: A critical review of the debate

Mass appraisal techniques are used in the valuation of large groups of real estate assets. Their use involves the use of common real estate data, a single evaluation protocol and result verification tests. Given the vast amount of information they have to process, they are entrusted to automatic value prediction models. If initially these models were based on the theory of implicit marginal prices, identified through regression analysis, now they can take radically different forms thanks to the novelties brought by statistical self-learning algorithms. The algorithms of automatic learning – known as machine learning models – autonomously learn the information contained in a dataset. They are able to acquire the existing relations between the characteristics of the assets and the values of price of the goods, even when these have forms well distant from the more traditional linear relation. Each model is first trained with the data of known cases, and then tested in its ability to predict unknown values. The scientific literature has followed the evolution of the machine learning models for the prediction of the value, investigating them under more analysis profiles. The most frequently found research theme concerns the comparison of several evaluation models on the same dataset of real estate data, compared in terms of accuracy in the prediction. The research provides a critical review of the debate in all publications in which the effectiveness of new value prediction models has been empirically investigated. The models prove to be effective in their predictive capacity, less effective in their inferential capacity, i.e. to evaluate the dependence of the price phenomenon on the causes explained by the variables. The debate confirms a higher accuracy of prediction of the new models with respect to the traditional regression analysis. However, it is not possible to rank the models in order of accuracy, as the effectiveness of each model depends on the data available to it. In the face of this undeniable advantage, these models present a limit in their characteristic of black box: the valuer cannot know with certainty what values and forms the variables assume in the learning processes. This makes the models ineffective for understanding the dynamics of formation and variation of value in relation to the characteristics of the good and external agents.