| Abstract [eng] |
In this study, historical periods of real estate price bubbles in four selected countries were identified, and a random forest model was created to forecast real estate price bubbles. The research uses monthly time series data for the years 2000 – 2022. Ten factors influencing the formation of real estate price bubbles were selected for analysis. To determine the historical periods of real estate price bubbles, the "Exuber" package in R-Studio was chosen, which applies the SADF, GSADF, and BSADF criteria. By applying these statistics to the country's real estate price index variable, the existence dates of price bubbles for the analyzed years were determined. The foundation of further model development is based on three main elements: the creation of a binary variable, the use of a sliding data window, and the selection of optimal parameters for the random forest model. The sliding data window method was chosen for forecasting the price bubble. To determine the most suitable combination of the sliding window, several variations were tested, and the best combination was selected: a 12-month window and a 6-month forecast. Five random forest models were created. In order to select the best model, the accuracy, confusion matrix, kappa, sensitivity, and specificity indicators of the created models were analyzed and compared. The results showed that the most suitable forecasting model was created based on standardized data from the three selected countries. When applying the model to Lithuanian data, it was found that a real estate bubble is not expected at the beginning of the third quarter of 2023. Furthermore, the factors that have the greatest influence on the model's predictions were identified. |
