MATHEMATICAL MODELING AND FEATURE IMPORTANCE ANALYSIS IN THE PROBLEM OF HOUSING PRICE FORECASTING IN THE SECONDARY MARKET USING MACHINE LEARNING

Abstract

The paper investigates the problem of price prediction in the secondary residential housing market of Kyiv through the construction and comparison of alternative regression models. The relevance of the study is determined by the multifactor nature of housing price formation, high market volatility, and the need to improve the accuracy of analytical valuation tools. Classical parametric approaches do not always adequately capture nonlinear relationships and interactions between property characteristics, which justifies the use of modern algorithmic methods. A comprehensive data preparation procedure was implemented, including data cleaning, missing value imputation using the k-nearest neighbors algorithm, Winsorization of extreme observations, logarithmic transformation of target variables, and normalization of numerical features. Particular attention was given to feature engineering based on textual property descriptions. Recurrent qualitative characteristics were identified and encoded as binary variables, enabling the incorporation of infrastructural and comfort-related attributes into the analytical framework. Additionally, geospatial enrichment was performed by calculating the distance of each property to the city center as a key location indicator. Multiple linear regression, ensemble algorithms (XGBoost, LightGBM, Random Forest), and an artificial neural network were developed and compared. Model performance was evaluated on a test dataset (20 % of observations) using R2, RMSE, MAE, and MAPE metrics. Boosting algorithms, particularly LightGBM, demonstrated the highest predictive accuracy. Feature importance analysis revealed that total area is the dominant determinant of overall housing price, together with location-related characteristics and building age. In contrast, price per square meter is more strongly driven by locational parameters and administrative affiliation. The results confirm the effectiveness of ensemble algorithms combined with extended feature engineering for quantitative analysis of price formation mechanisms in the local residential real estate market.