Evaluating avms performance. beyond the accuracy

Automated Valuation Models (AVMs) are regularly used in mass appraisal techniques. Thanks to developments in artificial intelligence, machine learning algorithms are increasingly being used alongside traditional econometric models. The final phase in the definition of the models consists in the verification phase of the results elaborated by Avm. The predictive effectiveness tests evaluate the models trained on part of the dataset (the training set) and then measure their ability to predict the remaining values of the dataset (testing set). This verification methodology provides as final output the accuracy parameter, i.e. the difference between predicted prices and actual prices. According to many authors this parameter, if considered alone, is insufficient. The research consists in an accuracy test of 5 Avm in the ability to predict the values of 1038 properties in the city of Padua. To the accuracy results of the test are added the results of cross-validation and the use of different statistical indicators for the measurement of predictive effectiveness. The results provide useful information that broadens the framework of model knowledge. They can be used in the analysis and description of automated evaluation models.