Predicting the Unseen: Transductive Transfer Learning in Real Estate Price Prediction

Abstract

This study examines transfer learning for time-dependent newly built apartment price prediction using spatial features only, motivated by their higher temporal stability compared to sociodemographic or economic variables. We evaluate whether a model trained in data-rich settings can generalize to unseen areas and to a different city. In Vienna, the transfer setup achieved a mean absolute percentage error (MAPE) of 20%–30% for 1-year predictions in unseen areas, with a maximum negative error (transfer-induced performance loss relative to the non-transfer baseline) of 11%, and performed particularly well in developing areas (MAPE ≈10%). When transferred to another city, the maximum negative error remained below 8% under complete spatial feature coverage, with a minimum MAPE of 2.08% for 1-year predictions. Finally, reducing the spatial feature set from 118 to approximately half preserved predictive performance while lowering computational complexity and improving practical applicability. Overall, the results suggest that spatial feature-based transfer learning can provide competitive short-term accuracy for apartment price prediction while supporting principled feature reduction.