Predicting land valuations within 4% using spatial ML

Mass appraisal has run on hedonic regression for decades, and for good reason: it is transparent, defensible, and easy to explain to a valuation tribunal. But regression assumes value is a smooth, mostly linear function of a property's attributes, and on the urban fringe that assumption falls apart. A parcel's worth there depends heavily on what surrounds it, and those effects are neither linear nor independent.

Why residuals blow out on the fringe

Two problems compound. First, spatial autocorrelation: nearby parcels share unobserved value drivers — a view, a school catchment, a coming rezoning — that a per-parcel model treats as noise. Second, amenity gradients are non-linear: proximity to a train line adds value up to a point and then subtracts it once you are close enough to hear it. Hedonic models paper over both, and the residuals show it exactly where valuations are most contested.

Two models, two jobs

We split the problem. A gradient-boosted tree ensemble handles the per-parcel attributes — land area, zoning, frontage, improvements — where trees capture non-linear interactions that regression cannot. A Graph Neural Network handles the neighbourhood: each parcel is a node, adjacency and shared frontage are edges, and the network learns to propagate value signals between neighbours the way the market actually does.

• Gradient-boosted trees: non-linear per-parcel attribute effects.
• Graph Neural Network: spatial spillover across the parcel adjacency graph.
• A blend layer combines both into a single valuation with an uncertainty band.

No hand-engineered distance features

The result that mattered most to us: the GNN made manual distance-to-amenity features redundant. We did not have to compute "metres to nearest school" or "metres to CBD" and guess at their functional form. The adjacency graph let the model discover those gradients itself, which is both less work and less bias.

Across fourteen cadastral districts the combined model reached a median absolute percentage error of 3.8%, with the largest improvements exactly where hedonic regression was weakest — the heterogeneous fringe.