Location, Location, Location: The Power of Neighborhoods for Apartment Price Predictions Based on Transaction Data

Land and real estate have long been regarded as stable investments, with property prices steadily rising, underscoring the need for accurate predictive models to capture the varying rates of price growth across different locations. This study leverages a decade-long dataset of 83,527 apartment transactions in Vienna, Austria, to train machine learning models using XGBoost. Unlike most prior research, the extended time span of the dataset enables predictions for multiple future years, providing a more robust long-term prediction. The primary objective is to examine how spatial factors can enhance real estate price predictions. In addition to transaction data, socio-demographic and geographic variables were collected to characterize the neighborhoods surrounding each apartment. Ten models, each varying in the number of input years, were trained to predict the price per square meter. The model performance was assessed using the mean absolute percentage error (MAPE), offering insights into their predictive accuracy for both short-term and long-term predictions. This study underscores the importance of distinguishing between newly built and existing apartments in real estate price modeling. By splitting the dataset prior to training, predictive models focusing solely on newly built properties achieved an average reduction of about 6% in MAPE. The best-performing models achieved an average MAPE of 15% for one-year-ahead predictions and maintained a MAPE below 20% for predictions up to three years ahead, demonstrating the effectiveness of leveraging spatial features to enhance real estate price prediction accuracy.