Deep learning accurate predictions of house prices are essential for prospective homeowners, investors, appraisers, and insurers. However, some studies lack accuracy as they overlook critical factors like accessibility and economic attributes that influence house prices. This paper aims to predict house prices by considering structural, locational, accessibility, and economic attributes, while also exploring the effect of accessibility on housing prices. The dataset contains 2,019,663 real estate transaction records from 1975 to 2018 in the Washington metropolitan area, obtained from the Zillow website. In this study, the accessibility index is calculated using Distance, Cumulative Opportunities, and Gravity measures, with the gravity measure surpassing others due to its consideration of both land use and transportation aspects. Economic attributes are then utilized to predict the average monthly house price using deep learning algorithms such as LSTM, GRU, and Simple RNN, with the Simple RNN demonstrating superior performance. Following the amalgamation of structural and locational attributes with the accessibility index and average house prices, various machine learning algorithms—including Linear Regression, Lasso, Ridge, Random Forest, GBM, LightGBM, XGBoost, Decision Tree, AdaBoost, Artificial Neural Network, and Stacked Generalization—are employed for prediction. Subsequent evaluation reveals that Stacked Generalization (ANN + LightGBM) provides the best performance, with an R-squared value of 0.96 and RMSE of $23,290. Moreover, this paper identifies accessibility index thresholds (80,003 for large buildings and 160,103 for small buildings) and demonstrates that a higher accessibility index leads to lower housing prices, attributed to noise pollution, decreased privacy, and increased supply responses.
Read full abstract