Comparing hedonic and machine learning models in residential property valuation: evidence from sub-market analysis in Italy
Gabriella Maselli, Laura Gabrielli, Antonio NesticòPurpose
This study aims to compare the predictive performance, robustness and interpretability of traditional hedonic regression and machine learning (ML) models in residential property valuation. While regression-based approaches estimate marginal prices directly, they may be limited in capturing non-linearities and multicollinearity. Conversely, artificial intelligence (AI) models provide greater flexibility but are more prone to overfitting and often lack transparency. Explainable AI (XAI) is integrated to improve the interpretation of price-formation mechanisms.
Design/methodology/approach
An exploratory case study compares multiple linear regression (MLR), artificial neural networks (ANN), random forest (RF) and a hybrid stacking model combining ANN and RF. The analysis includes variable definition, data collection and preprocessing, model specification and implementation and validation through training and k-fold cross-validation. To enhance interpretability, permutation feature importance and SHapley Additive exPlanations values are applied.
Findings
The empirical application focuses on two Osservatorio del Mercato Immobiliare zones (B1 and B2) of Padua (Italy). Model performance is strongly influenced by sample size, data heterogeneity and validation strategy. Stacking provides a favourable balance between predictive accuracy and robustness, particularly in the larger and more heterogeneous B2 sample, whereas ANN and RF are more prone to overfitting in smaller and less heterogeneous data sets (B1). MLR proves more stable across training and cross-validation phases. XAI confirms the central role of floor area in price determination, together with maintenance condition, construction period and service availability. Model performance should be interpreted as a trade-off between accuracy, stability and transparency.
Originality/value
This study provides an integrated comparison of econometric, ML and XAI-based approaches within a homogeneous housing market. It jointly evaluates predictive accuracy, robustness and interpretability, allowing different modelling strategies to be evaluated not only in terms of forecasting performance but also in relation to their transparency and stability. The empirical application to the Padua residential market provides useful evidence for researchers, valuers and decision makers interested in the use of advanced predictive tools for property value forecasting and market analysis.