How Many Data for a Reliable Assessment? Accuracy of Models and Number of Comparables in Automated Valuation Models (AVMs)

Abstract

Automated valuation models are increasingly used in appraisals. In some circumstances, their use provides accurate estimates with advantages in terms of time and cost of the process. Besides the programming capability, the reliability of an automated model depends on the data it has to train and verify its outcomes. In literature, many researches test the predictive capabilities of models without investigating whether similar performance would have been achieved with less data. The issue is crucial for evaluators, who aim to optimise effectiveness (accuracy) while reducing cost (data retrieval) in order to gain a competitive advantage over the industry. The research aims to test whether there is a minimum dataset size with which AVMs achieve their maximum performance. Tests were conducted on increasing sizes of the dataset in order to observe the behaviour of the models as the number of data increases. Three econometric models (Linear, polynomial and logarithmic regression) and three machine learning models (Regression tree, Random forest and Nearest neighbors) were tested. The results confirm that there is a minimum dataset size only for the machine learning models, whereas the hedonic models show little correlation with the amount of data. However, the degree of accuracy of the models seems to depend not only on the number of samples but also on the amount and type of predictors.