|Comparison of Models to Forecast Real Estates Index Introducing Machine Learning
|이주미(Lee, Ju-mi) ; 박성훈(Park, Sung-Hoon) ; 조상호(Cho, Sang-ho) ; 김주형(Kim, Ju-Hyung)
|Real Estates Index Forecasting; Machine Learning; Long Short-Term Memory
|As the real estates occupy major portion of domestic households assets, relevant issue has been dealt seriously by the Korean government.
However, apartment prices in downtown Seoul, the capital city, have soared despite various policies. Forecasting the real estate market trend
has become an important research topic in order to provide information for establishing policies. In the prediction of the real estate market in
the previous studies, two research directions were classified as follows: quantitative economic models and machine learning models. Regarding
this trend, there was a need for comparative research on machine learning models, emerging methods, that are used to compare and predict
various real estate indices. In this study, the machine learning model RF(Random Forest), XGBoost(eXtreme Gradient Boosting), and LSTM
(Long Short Term Memory) are used to select suitable machine learning models for selected real estate index and conduct a comparative
study to validate predictive power of machine learning models. Apartment sales index, land price index, charter price index, and real estate
psychological index using univariate variables are predicted. In addition, RF, XGBoost and LSTM models all tended to be generally marginal
with RMSE values of 0.0268, 0.0296, and 0.0259 in charter(Jeonse), Korean traditional pre-deposit rental system, price index data with linear
but small variants. This shows that the prediction of the real estate index is deviated from the prediction accuracy of machine learning
models depending on the periodic characteristics and data characteristics of the real estate index.