Comparative analysis of the effect of various spatial information accounting methods on the accuracy of valuation models of real estate: the case of Tomsk two-room apartments

This paper considers different methods of accounting information on the spatial arrangement of real estate items to assess the accuracy of cost estimation. The study is focused on a relatively homogeneous group of two-room apartments in brick and bearing-wall houses in the city of Tomsk. Basic data (cost, area, etc.) on the apartments was obtained from ru09.ru. Yandex.ru was used to perform geocoding. The obtained data was filtered, cleaned and enriched. Duplicates and conflicting data were removed from the sample; values of the new variables were calculated: initial data set consisted of 5072 records; cleaned data set consisted of 1656 records. This was followed by a descriptive analysis of the downloaded data. Basic statistical characteristics were calculated and the distribution histograms were built for each variable. We considered several ways of accounting information on the spatial arrangement of the apartments on the city map: based on models with variable structure and based on k-nearest neighbor method. The effectiveness of each model was evaluated in comparison with the efficiency of the basic model that ignored spatial information. Comparison of the models was carried out using tq-Cross Validation method. It was established that the inclusion of information about the spatial arrangement of an apartment enables to increase the accuracy of the forecast. The approach based on k-nearest neighbor method was considered the most effective. The search for optimal parameter values was conducted in the range of [10, 200] in increments of 5 using tq-CrossValidation method. It was found that the distribution of the optimal values of the parameter K had a bimodal character. The most common parameter K equaled to 34 and 82. The comparative analysis of the models performance was conducted. Histograms of MSE parameter values distribution for each model were built. A comparative analysis of estimates of models coefficients was conducted.

Keywords

Authors

References