Machine learning in real estate valuation

As the built environment's information layer expands and computational methods advance, valuation, which underpins investment and risk decisions, relies on predictive models. These models can capture complex price relationships, but their value depends on whether their output translates into informative, comparable market measurement, a translation complicated by fragmented data, heterogeneous assets, and thin markets. This dissertation therefore investigates how machine learning and Automated Valuation Models (AVMs) improve real estate valuation and market measurement, through three papers on price index construction, uncertainty quantification, and distributional market dynamics.
The first paper proposes a model-agnostic index methodology based on out-of-time prediction errors, applied to 29,998 commercial transactions in New York (2000-2019). Machine learning improves prediction accuracy over linear benchmarks, but is more sensitive to calibration data in small samples. The second evaluates four prediction interval methods with OLS, XGBoost, and Cubist on 86,838 residential transactions in Amsterdam (2001-2021): the model mainly determines interval width, while the interval construction determines coverage. The third builds quantile price indices from 175,070 apartment transactions in Warsaw (2006-2024): the 10th and 60th percentiles lead much of the distribution, while the upper tail follows.
Across the three papers, the results show that predictive models can support market measurement well beyond point predictions, provided that their stability, interval coverage, and distributional behavior are evaluated explicitly.