Towards a Better Uncertainty Quantification in Automated Valuation Models

The lecture focused on applying uncertainty quantification (UQ) methodology to non-linear automated valuation models (AVMs) in the real estate industry. While linear AVMs have been preferred due to their interpretability, this paper aims to demonstrate that more accurate non-linear algorithms also can effectively estimate prediction uncertainty. Using a dataset of 51,747 historical apartment transactions in Oslo, Norway, an AVM is trained using an XGBoost algorithm. Out-of-sample predictions are made, and three UQ approaches are evaluated: direct loss estimation, the standard deviation of a bootstrapped ensemble, and the width of a quantile regression prediction interval. A stacked generalization model that combines the three UQ estimators is proposed for improved performance. The findings indicate that all three UQ methods provide reliable uncertainty metrics, showing moderate positive correlation with the absolute prediction error and high correlation among themselves. A SHAP analysis of the stacked generalization model reveals that the bootstrapping ensemble estimator performs better for most apartments, while the quantile regression estimator excels for larger and more expensive units. The paper contributes by applying UQ methodology in real estate, highlighting the compatibility of non-linear AVMs with uncertainty interpretation. It also offers a comparative analysis of three UQ estimators and their suitability for specific types of dwellings.