Prediction accuracy for property tax mass appraisal: A comparison between Regularised Machine Learning and the Eigenvector Spatial Filter approach

Prediction accuracy for mass appraisal has evolved substantially over the last few decades, facilitated by the revolution in data availability and the advancement of computational software. Accompanying these advances, newer forms of geospatial approaches and machine learning algorithms have opened up new horizons for price prediction and mass appraisal assessment. The application of machine learning and Artificial Intelligence within mass appraisal has witnessed considerable debate and often perceived as impractical, due to limited explainability and defensibility required for assessment application, notably in challenge scenarios. This study compares a traditional multiple regression approach (MRA) with regularised (penalized) machine learning approaches and a more nuanced geo-statistical technique, the Eigenvector Spatial Filter (ESF) approach, applying datasets for two different urban residential areas within the U.K. and U.S.A. The findings show the efficacy of the geo-statistical ESF technique against the ML approaches – both of which outperform the traditional MRA. The findings also show the ESF approach to provide the basis of a more understandable alternative spatial method for mass appraisal aligned with the MRA approach, with the spatial filters easily incorporated as predictors into MRA to alleviate spatial autocorrelation. Further, the penalized ML regression approaches offer a more practical alternative to other forms of ML for assessors. Both methods produce reliable, yet understandable, regression models for mass appraisal assessment.