Gradient Boosting Decision Tree for House Price Prediction with Google Trends

Predicting house price accurately can reflect the popularity of the housing market and help making decisions for investors and policymakers. Statistics of macro factors are commonly used for house price forecasting; however, macro factors obtained from government reports have defect of time lag and may impair the prediction performance. Google Trends data can serve as a leading sentiment indicator of people’s attitudes and expectations toward the housing market and help improve house price prediction. Therefore, this study proposes a new methodology framework for house price prediction with Google Trends data. Recursive Feature Elimination (RFE), a feature selection method, is utilized to remove noisy data and improve feature quality. Gradient Boosting Decision Tree (GBDT) is adopted to establish models for house price forecasting. Real estate-related Google Trends data, along with the fundamental house price index (HPI) data are collected to predict the growth rate of HPI in the United States. Results show that RFE can effectively remove irrelevant features and improve the model performance. GBDT has higher and more stable prediction accuracy than other prediction models, especially when the predicted time span is long. Compared with models including fundamental HPI data only, models containing Google Trends data can exhibit higher and more stable prediction accuracy for long time span forecasting. Three categories of Google Trends indices, including “house rent”, “housing market & real estate market”, and “mortgage & real estate agency” are found to be the most important indicators of the variation of HPI growth rate.