Predicting the housing prices correctly is one of the major problems in urban analytics, especially when the area experiences high micro-geographic heterogeneity. The spatial, structural, and socioeconomic relationship is often nonlinear and is complex and traditional models do not give an accurate picture of the relationships of variables at the level of the neighborhood. In this paper, a deep learning architecture is proposed to accommodate hyper-local data sources and multi-dimensional sets of features to predict the prices of housing in the Tri-County Region of Seattle (which includes the counties of King, Pierce, and Snohomish). The goal is to build a data driven model which could advance the accuracy of price (forecast) with higher precision at the sub regional level to provide finer granularity of valuations. We use deep neural network (DNN) model that is trained on property-level data presented by Zillow and Redfin which are supplemented with demographic, locational and temporal features provided by the U.S. Census Bureau and OpenStreetMap. The locational resolution is ensured with the help of data normalization (date normalization and feature encode nodes), and geospatial tagging. Empirical evidence shows that the model suggested performs much better than the classical regression and ensemble methods, as it has lesser root mean square error (RMSE) and mean absolute error (MAE) in all the three counties. It entails impressive spatial generalisation and resistance to changes of the neighborhood properties of the model. Paper contributes to the growing research of urban AI by the innovation of ,strokes projective approach for hyperlocal housing markets forecasting. These results have real-life application on real estate investors, urban planners, and policymakers who would like to encourage equitable and practice-based housing policy.
Aditya Kasturi (Wed,) studied this question.