Using Machine Learning Methods to Estimate the Cost of Housing

Abstract

Introduction. Nowadays, the state has enshrined at the legislative level the definition of appraised value for tax purposes in sales of real estate as mandatory. The comparative approach most often used by appraisers has disadvantages such as the inability to find analogues in some cases and the need to make corrections, which affects the reliability of the results. The module of electronic determination of appraisal value (Module) similar to the object of property appraisal of the Unified database of appraisal reports works on the same approach and quite often overestimates appraisal value that leads to increase in the size of the tax during sales as the real estate cannot be sold for the price less than the estimated cost.\n\nToday to determine the price of an automated system correctly, it is necessary to fill the Unified Valuation Database in the State Property Fund with large knowledge bases - a huge IT system. So far, the thoughtless machine still determines the price by the average value. Currently there are often situations when the appraised value of real estate, determined by the Module, exceeds its real market value. Given that the approach used by the Valuation Module does not always give the correct result, there is a need to find a better method to determine the value of housing that could be used by the Module.\n\nThe purpose of the paper. In this paper, an approach based on fuzzy logic was used to estimate the cost of housing in Kyiv. Fuzzy methods allow to apply a linguistic description of complex processes, to establish fuzzy relationships between concepts, to predict the behavior of the system, to create a set of alternative actions, to formally describe fuzzy decision-making rules.\n\nResults. The software implementation of the model in Python programming language was performed. Data for modeling were taken for the period July – October 2020 from a single database of property valuation reports. The sample contained 2133 records, it was filtered, divided into training and testing in the proportion of 85 : 15. To assess the quality of the program, the average relative error of the developed model was calculated.\n\n \n\nKeywords: fuzzy logic, machine learning, Python programming, linguistic variables, predictive model.