Random Forests with Economic Roots: Explaining Machine Learning in Hedonic Imputation

The increasing integration of machine learning techniques in hedonic imputation has significantly improved predictive capabilities for both price prediction and price index computation. Nevertheless, the lack of transparency, particularly regarding model explainability, remains a notable challenge. This paper undertakes a comprehensive investigation through simulation and empirical analysis, focusing specifically on linear hedonic regression and tree-based machine learning methods, with the aim of elucidating the economic rationale behind machine learning predictions. Our comparative analysis highlights the superior performance of random forests over traditional hedonic regression models. This enhanced performance is critically dependent on assumptions about consumer utility functional forms: tree-based algorithms outperform linear hedonic models most significantly when assuming limited substitutability between product attributes, a characteristic that aligns with the algorithm’s design. The core contribution of this research is the development of a novel model explainability framework that bridges the gap between economics and machine learning. By shedding light on the economic rationale behind machine learning outcomes, we address the growing demand for interpretable machine learning approaches within the economics community. This framework not only enhances the transparency of machine learning models but also provides valuable insights for their application in real-world economic scenarios.