Analysis of Temporal Changes in Performance of Housing Price Estimation Models

Abstract

To accurately estimate housing prices, various models have been proposed in existing studies, but in practical use, it is necessary to consider model deterioration over time. However, the effect of time on models is not clear. This study estimates housing prices using Ridge regression, SVR, LightGBM and kNN for newly built houses and rental apartments located the wards of Tokyo or the government-designated cities in the LIFULL HOME'S dataset. A period of disuse was set aside between training and test data to evaluate model degradation over time with MAPE. The results suggested a decrease in prediction accuracy over time for all models. In particular, kNN achieved the best prediction accuracy, 6.70% for newly built houses and 7.56% for rental apartments, immediately after training, but the worst deterioration was observed 12 months after training, with +4.53% for newly built houses and +3.53% for rental apartments.