System of modeling and forecasting real estate prices based on machine learning methods

Abstract

The article examines how the problem of forecasting real estate prices is solved using a systematic approach to modeling and forecasting. Machine learning methods were systematically used to solve the problem. The systematic approach to modeling and forecasting is based on the analysis of the studied processes, establishing the types of existing characteristic uncertainties, assessing the structure and parameters of the model, as well as forecasts based on the constructed model. It combines three groups of tasks on a single methodological basis: the task of data analysis and pre-processing; the task of building models and their evaluation; the task of building forecasts and their evaluation. The structure of the systematic approach to modeling and forecasting is developed and presented. An important aspect that affects the effectiveness of using machine learning methods is the process of pre-processing data. Improving the methods of pre-processing data is a complex task that must be solved systematically, taking into account the specifics of the real estate market. Therefore, in this study, considerable attention is paid to the process of pre-processing data and research aimed at increasing the effectiveness of predictive values. The architecture of an information system for solving modeling and forecasting problems is developed and presented. As an example of implementing an information system, the problem of forecasting real estate prices is considered. The results of the following stages are presented: data collection, research and data preparation, model training on data, determining model efficiency, improving the efficiency of basic models. The following groups of models were used to solve the forecasting problem: regression models, tree models. The effectiveness of forecasting solutions was assessed using the MAE, MSE, RMSE, MAPE metrics. To improve the quality of forecasts, a single-layer structure of a heterogeneous ensemble of models based on stacking is proposed.