4th industrial revolution discussed in the World Economic Forum forecasts an era in which AI applied to a wide range of technologies can replace human brains. This, not only economically but also socially, is expected to be a turning point that will bring about an important change. In the real estat...
4th industrial revolution discussed in the World Economic Forum forecasts an era in which AI applied to a wide range of technologies can replace human brains. This, not only economically but also socially, is expected to be a turning point that will bring about an important change. In the real estate field, there is a call for new ways of research to forecast future real estate price change taking into account more diverse macro and micro variables.
This research aims at finding an optimal research model by comparing traditional statistical methods applying deep-learning technology based on Seoul apartments’ real transaction price index that represent general housing prices, forecasting changes in the apartments’ real transaction prices, and discovering factors that have impact on housing prices. 125 variables were collected that impact real estate prices among big data released in accordance with the government’s 3.0 policy and final 65 variables were selected as explanatory variables for analysis.
First, prediction rates of real transaction price index were compared applying SVM, random forest, and artificial neural network or ANN that are used for machine running to the singular time series model. Comparison of MSE(Mean squared error) values of 1 step-ahead placed the ANN multilayer perceptron model at 0.38, 11 times lower error rate compared to 4.16 of linear regression model, and superior to SVM (2.20) and random forest (0.86). Linear regression model and SVM expected a simple trend showing that future prediction values will steadily increase, random forest expected the prices will end up remaining steady, and the multilayer perceptron model predicted the prices will go up until the first half of 2018 and then go down.
Second, housing transaction volume time series analysis put random forest at the lowest error rate, 0.43, ANN at 0.6, SVM 1.95, and linear regression model at 2.0 respectively. In case of time series analysis, its nature being fluctuating and unpredictable, ANN has been found inappropriate due to difficulty in reading regular patterns of the past data.
Third, we have selected the 40-5 model from discrete data having looked for the best model in terms of accuracy based on the number of hidden layers and nodes through real transaction price index using deep-learning. The accuracy of training data recorded 93.6%, AUC value being 0.98 and that of test data 85.7%, its AUC value being 0.97, which proved excellence both in sensitivity and specificity as well as adequacy in complexity of the model with 45 nodes being used. MAE value of the 30-20 model in successive data represented the lowest, 0.2, and 0.985 of a correlation coefficient which is close to the real value. DNN(Deep Neural Network) with more than 3 hidden layers showed less accuracy as the number of layers went up, which proved its unsuitability for analysis. It has been concluded that ANN utilizing multilayer perceptron with 2 hidden layers is effective in analyzing typical macro economic variables.
Fourth, variables from discrete data that have a significant impact on real transaction price index listed total construction completion value, Seoul apartments’ transaction volume, average KOSPI trade cost per day, and company default rate in order and those from successive data represented profit-making securities, economic sentiment index, and company default rate. Given the characteristics of the variables, it can be said that the discrete model is effective in short term prediction of housing market and the successive in long term.
Finally, analysis of real transaction price index for the next 2 years applying 65 variables to the 30-20 model that was picked as the best showed a modest increase until April 2017 and decrease afterward and that housing prices of the 1st quarter of 2018 will form around the lowest point in December 2008 when the prices collapsed in the wake of financial crisis.
Real estate prices can not be estimated only considering data from the past, macro economic variables, and partial characteristics of housing. They fluctuate depending on various factors such as government policies, consumer sentiment, and foreign economic landscape. Therefore further review is required to conclude whether they should be considered for decision making though results from the study explain the reality by and large. More accurate real estate market prediction models need to be developed applying many different analytic methods based on this study.
4th industrial revolution discussed in the World Economic Forum forecasts an era in which AI applied to a wide range of technologies can replace human brains. This, not only economically but also socially, is expected to be a turning point that will bring about an important change. In the real estate field, there is a call for new ways of research to forecast future real estate price change taking into account more diverse macro and micro variables.
This research aims at finding an optimal research model by comparing traditional statistical methods applying deep-learning technology based on Seoul apartments’ real transaction price index that represent general housing prices, forecasting changes in the apartments’ real transaction prices, and discovering factors that have impact on housing prices. 125 variables were collected that impact real estate prices among big data released in accordance with the government’s 3.0 policy and final 65 variables were selected as explanatory variables for analysis.
First, prediction rates of real transaction price index were compared applying SVM, random forest, and artificial neural network or ANN that are used for machine running to the singular time series model. Comparison of MSE(Mean squared error) values of 1 step-ahead placed the ANN multilayer perceptron model at 0.38, 11 times lower error rate compared to 4.16 of linear regression model, and superior to SVM (2.20) and random forest (0.86). Linear regression model and SVM expected a simple trend showing that future prediction values will steadily increase, random forest expected the prices will end up remaining steady, and the multilayer perceptron model predicted the prices will go up until the first half of 2018 and then go down.
Second, housing transaction volume time series analysis put random forest at the lowest error rate, 0.43, ANN at 0.6, SVM 1.95, and linear regression model at 2.0 respectively. In case of time series analysis, its nature being fluctuating and unpredictable, ANN has been found inappropriate due to difficulty in reading regular patterns of the past data.
Third, we have selected the 40-5 model from discrete data having looked for the best model in terms of accuracy based on the number of hidden layers and nodes through real transaction price index using deep-learning. The accuracy of training data recorded 93.6%, AUC value being 0.98 and that of test data 85.7%, its AUC value being 0.97, which proved excellence both in sensitivity and specificity as well as adequacy in complexity of the model with 45 nodes being used. MAE value of the 30-20 model in successive data represented the lowest, 0.2, and 0.985 of a correlation coefficient which is close to the real value. DNN(Deep Neural Network) with more than 3 hidden layers showed less accuracy as the number of layers went up, which proved its unsuitability for analysis. It has been concluded that ANN utilizing multilayer perceptron with 2 hidden layers is effective in analyzing typical macro economic variables.
Fourth, variables from discrete data that have a significant impact on real transaction price index listed total construction completion value, Seoul apartments’ transaction volume, average KOSPI trade cost per day, and company default rate in order and those from successive data represented profit-making securities, economic sentiment index, and company default rate. Given the characteristics of the variables, it can be said that the discrete model is effective in short term prediction of housing market and the successive in long term.
Finally, analysis of real transaction price index for the next 2 years applying 65 variables to the 30-20 model that was picked as the best showed a modest increase until April 2017 and decrease afterward and that housing prices of the 1st quarter of 2018 will form around the lowest point in December 2008 when the prices collapsed in the wake of financial crisis.
Real estate prices can not be estimated only considering data from the past, macro economic variables, and partial characteristics of housing. They fluctuate depending on various factors such as government policies, consumer sentiment, and foreign economic landscape. Therefore further review is required to conclude whether they should be considered for decision making though results from the study explain the reality by and large. More accurate real estate market prediction models need to be developed applying many different analytic methods based on this study.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.