Comparing Housing Valuation Techniques and Stacked Generalization: Exploiting Explainable AI

Abstract This study investigates the use of diverse model selections in stacked generalization to improve the predictive accuracy of Automated Valuation Models (AVMs). The research focuses on three commonly used valuation techniques: the Comparable Sales Method (CSM), the linear Least Absolute Deviation (LAD), and the nonlinear XGBoost (XGB). A dataset consisting of 164,619 apartment transactions from Oslo between 2008 and 2022 is utilized for testing, with 25% of the data used for out-of-sample predictions. While the stacked model combining XGB, CSM, and LAD achieves the best performance with a Median Absolute Percentage Error (MdAPE) of 5.17%, the individual XGB model performs nearly as well, achieving an MdAPE of 5.24%. Analysis reveals that stacking provides marginal improvements and primarily relies on XGB predictions. However, the computational cost of stacking raises questions about its practicality. The research highlights the limitations and benefits of different housing valuation techniques and varying data sizes, offering practical insights to enhance the performance of future AI-AVMs. By utilizing explainable AI, this study contributes to a better understanding of how different models collaborate and diverge, revealing their competitive advantages in house price valuation.