Do Kolmogorov-arnold Networks Have Great Genes? An Application to Sydney’s Housing Market

We benchmark Kolmogorov-Arnold Networks (KANs) against thirteen alternative methods—regularized linear, kernel, tree-based, and neural—for predicting suburb-level relative prices in Sydney’s housing market. Our data cover 2.5 million residential sales across 963 suburbs over 2001–2025, aggregated to suburb-level relative prices defined as the ratio of each suburb’s median sale price to the Greater Sydney median. A standardized evaluation design—common predictor set, common tuning protocol, and common cross-validation scheme—isolates model effects from data effects across two time horizons: a 25-year crosssection and a 5-year panel. Gradient-boosted trees dominate at both horizons, consistent with prior evidence on tree-based superiority for tabular data. Despite a stronger theoretical foundation—the Kolmogorov-Arnold Representation Theorem guarantees exact decomposition of any continuous function into univariate components—KANs never rank in the top four and are the most computationally expensive method. The bottleneck is not the theorem but its implementation: gradient descent on B-spline coefficients offers no guarantee of recovering the exact decomposition.