A Comparative Study on Artificial Intelligence Based Automated Valuation Models

This paper presents a comparative study of traditional linear regression approaches, and a range of Artificial Intelligence (AI) based automated valuation models (AVMs) in the context of mass appraisal valuation. Using datasets from Adams County (USA), Belfast (UK), and Eindhoven (Netherlands), the research evaluates model performance across a diverse range of market structures and data characteristics. A suite of models including Decision Tree, Random Forest, Gradient Boosting, Light Gradient Boosting, Generalized Regression Neural Network, Multi-Layer Perceptron, and Dense Neural Network were developed and tested against standardised ratio study metrics from the International Association of Assessing Officers (IAAO), along with other statistical measures for evaluating model performance. The findings reveal that ensemble machine learning models consistently outperform traditional regression methods across all the jurisdictions examined, exhibiting superior accuracy, precision and equity even with limited predictor datasets. Furthermore, the application of hyperparameter tuning was observed to enhance model performance, particularly for boosting based algorithms. The study highlights the adaptability of AI-based AVMs to heterogeneous and non-linear property markets, whilst emphasising the continued need for transparency in variable selection and communication around the variables used within automated valuation models. The findings underscore the transformative potential of AI in advancing equitable and accurate property assessment practices, and for the development of industry guidance on the ethical and transparent integration of AI and generative AI technologies within valuation frameworks and assessment practices.