Advisor: Ioannis Giannopoulos, Gerhard Navratil

Predicting real estate prices remains challenging due to the heterogeneity of real estate, the complexity of cities, and the temporal dynamics of the market. While existing research has advanced the modeling of real estate valuations, most approaches treat spatial characteristics as additional control variables rather than structural determinants, use validation strategies that overestimate predictive performance due to temporal leakage, and lack a systematic assessment of the temporal and spatial generalizability of the models.
This dissertation develops a spatially oriented modeling framework for predicting real estate prices that is tailored to the constraints and uncertainties of real markets. Using 171,040 verified residential real estate transactions from Vienna (2010–2022), this work constructs a data pipeline that integrates transaction records, socioeconomic indicators, infrastructure networks, crime statistics, and points of interest into a unified feature space with 165 variables. Novel geocomputational methods formalize the neighborhood context through isochrone-based accessibility measures, network-derived distances, and descriptions of built-up area derived from building area data.
The empirical part of the work addresses five research questions. First, long-term model stability is assessed through multi-year prediction experiments, which show systematic MAPE growth from 15% after one year to over 36% after eight years for tree-based models. Meanwhile, ensemble architectures reduce error by about 11 percentage points after eight years compared to tree-based baselines. Second, validation bias is quantified by comparing conventional random splitting with a strictly forward-looking evaluation, showing that conventional validation can substantially bias the assessment of predictive performance over the same period—an ensemble showed promising validation metrics but deteriorated significantly under temporal separation. Third, the feasibility of purely spatial predictions is demonstrated: removing socioeconomic variables costs less than 1 percentage point on average, while eliminating the need to predict these variables themselves, with isochrone-based features proving particularly robust at 300 meters in the Vienna experiments. Fourth, transfer learning experiments show that spatial models can be generalized across urban contexts—including a cross-city transfer from Vienna to Innsbruck with a negative transfer of typically less than 5 percentage points—although errors occur when vertical information is missing or the transportation infrastructure differs significantly. Fifth, a controlled experiment with 13 real estate experts shows that models outperform experts in 63% of comparisons when information is equal. Meanwhile, hybrid collaboration improves results for 62% of experts when predictions are appropriate. However, anchoring effects propagate errors when this is not the case.
The dissertation contributes to a reproducible urban computation pipeline, a temporally robust valuation methodology, systematic evidence for the transferability of spatial features, and an empirical basis for algorithmic predictions compared to professional judgment. The key finding is that spatial features—accessibility, urban morphology, neighborhood infrastructure—can serve as a stable basis for price predictions and that the transition from interpolation-based validation to true predictive evaluation is essential for an honest assessment of model capabilities.