Introduction to special issue

Abstract

In 2019, as guest editors of the Journal of Property Research, we called for contributions to the special issue Automated Valuation Services (AVSs). We were interested in particular in case studies that discuss the development, implementation, and operation of an AVS. We are very grateful to Bryan MacGregor, the editor of the journal, and to the many reviewers, who assessed the submissions and helped us with the selection of the four papers that have been included in the special issue. While there are already many papers that examine the performance of different statistical models for market value predictions of residential properties, only a few papers examine how to implement such models as a service for users on an ongoing basis. Users expect that such a service is easy to use, and they also expect that it is timely and robust. A service should provide a prediction of the market value, but should also indicate the uncertainty of this prediction in a manner that the user can understand. Methods from machine learning are increasingly used for these tasks and it can be difficult to explain these methods to non-experts. If it is important that details on the methods should be communicated to users, then this should be done as clearly as possible. The first paper by Hill et al. (2021) examines the importance of the performance measure used for the selection of the statistical model for an AVS. As there are usually competing statistical models, each should be fitted to transaction data with a rolling windows approach. Given the market value predictions from each of the competing models, sets of out-of-sample prediction errors can be computed. Obviously, the model with the ‘best’ prediction errors should be chosen. This requires, however, that each set of prediction errors is aggregated into measures that can be compared. Hill et al. (2021) provide a review and analysis of performance measures that have been proposed in the literature. Their classification of performance measures – and transformations of these – with respect to different aspects of the distribution of prediction errors underscores the necessity to align model selection with the application at hand. The authors examine this empirically with data from flat transactions from Graz, Austria. Based on their analysis, Hill et al. (2021) recommend seven core measures, each addressing a different aspect of the ‘best’ model. The second paper by Krause et al. (2020) addresses that every market value prediction – by its very nature – has an inherent uncertainty to it. The statistical model used in an AVS can provide estimates of uncertainty, such as prediction intervals, with ease and high accuracy. The authors start by unifying the terminology with which to discuss uncertainty. This is useful given the varied terminology in academic research