Assessing the robustness of machine learning-based automated valuation models across heterogeneous residential markets in Malaysia

Purpose This study aims to examine the comparative robustness of machine-learning-based automated valuation models (AVMs) for residential property valuation across different Malaysian market settings. It evaluates random forest (RF), extreme gradient boosting (XGBoost) and artificial neural network (ANN) models by considering predictive accuracy, temporal consistency and computational feasibility, while distinguishing comparative performance across independently modelled data sets from genuine cross-market transferability. Design/methodology/approach A quantitative comparative approach is applied using an empirical residential property data set from Johor Bahru and simulated scenarios representing Kuala Lumpur and Penang in Malaysia. RF, XGBoost and ANN models are independently developed and evaluated within each case using prediction-error measures and reported temporal performance. Computational execution is assessed separately as an indicator of technical feasibility. The analytical framework interprets model performance as conditional on the interaction between algorithm characteristics, data set structure, preprocessing and validation design rather than as an inherent property of a particular modelling technique. Findings The findings show that relative model performance varies across the evaluated data sets, with different machine-learning approaches demonstrating stronger predictive performance under different conditions. This indicates that there is no universal model that can be assumed to perform consistently across heterogeneous residential market settings. Predictive accuracy and temporal consistency also represent distinct dimensions of model performance, while computational feasibility alone does not establish operational efficiency within professional valuation practice. The findings therefore support context-specific model evaluation, independent validation and continued professional oversight before AVMs are adopted for operational valuation, lending, portfolio management, taxation or mass appraisal applications. Originality/value This study contributes an exploratory framework for evaluating AVMs beyond a simple comparison of predictive accuracy by integrating accuracy, temporal consistency, computational feasibility and the evidential status of the underlying data sets. It further clarifies the distinction between cross-market comparative robustness and model transferability, while positioning AVMs as components of a broader valuation system involving data governance, validation, professional judgement, transparency and accountability. For Malaysia and the wider valuation profession, this study provides a basis for developing more rigorous validation practices, professional guidance and governance frameworks for the responsible integration of machine learning into property valuation.