Is It Possible to Overfit the Algorithm? Case Study of Mass Valuation of Land Properties in Szczecin

Purpose: The article presents the mass estimation of real estate value using the Szczecin Mass Real Estate Valuation Algorithm, with various sizes of the initial sample. The aim of the article is to investigate how the change of the sampling pattern affects the results of the valuation and to capture how acceptably small the sample may be, so that the algorithm overfitting does not occur. Design/Methodology/Approach: In the mass valuation algorithm used, correlation coefficients were used to estimate the influence of particular property attributes on the property value. Empirical research was conducted on the basis of the database of land properties in Szczecin. This database was divided into two groups: a training set and a test set. For the trainee set, appropriate correlation coefficients were calculated, and then, using the algorithm of mass valuation, the value of real estate in the test set was estimated. Findings: In all analysed cases, the MAPE error for the testing sample was greater than for the training sample. However, the smallest difference between the errors for the training and testing sample occurred in the case of using proportional stratified sampling. When using simple randomization, increasing the sample size by 40% resulted in a decrease in the MAPE error value. On the other hand, reducing the sample size in order to reduce the costs of mass real estate valuation will result in an increase in the error value and model overfitting. Practical Implications: Appropriate selection of the real estate sample on the basis of which the coefficients of the influence of individual attributes on the value of the real estate are calculated directly affects the cost-effectiveness of the entire mass valuation process. Originality/Value: The smaller the sample size, the less real estate has to be individually appraised by property appraisers. In addition, the precise selection of real estate for the training sample in the algorithm affects the convergence of the obtained real estate valuations with their market value.