How pricing models see a building

To a pricing model, a building is a set of signals: transactions, location, size, floor, age, amenities, income, and the behaviour of the market around it. The quality of a valuation is decided less by the cleverness of the model and more by the breadth, freshness and provenance of the data it reads. Feed a brilliant model thin data and it produces a confident guess; feed a disciplined model rich data and it produces evidence.

Strip away the architecture and a valuation model consumes a few families of input. Transaction evidence: what comparable assets actually traded for, and when. Asset attributes: size, floor, orientation, age, condition, specification. Income evidence: rents achieved, not rents advertised. Market context: supply in the pipeline, absorption, the direction and speed of the local cycle. And location, which deserves more precision than it usually gets.

Most automated models treat location as a radius: draw a circle, collect comparables, average them. Geography does not work that way. Two assets a few hundred metres apart can sit in different rental universes if a railway corridor, a motorway or a river runs between them, because the people who rent and buy in each pocket do not cross it in daily life. A serious model grades comparables on real routing, drive time, walk time, transit time, and respects the invisible boundaries that define demand pockets, the way a local broker who has walked the streets would. A comparable on the wrong side of a boundary is not a slightly weaker comparable; it is evidence about a different market.

The public web is not enough. Much of what a serious valuation relies on lives beyond the search index: government registries, licensed data platforms, primary records, planning files, foreign-language portals. A model whose knowledge ends at what was crawled and cached is working from the visible fraction of the evidence. This is why source coverage is a structural advantage rather than a feature: the model that reads the original of record is answering a different question from the model that recalls an approximation of it.

Missing data rarely announces itself. The model still produces a number; the number is just quietly built on fewer legs. Thin transaction evidence forces wider inference from less similar assets. Stale rental evidence drags the estimate towards a market that no longer exists. Unread planning records hide the supply that will compete with the asset at exit. The honest response is not a wider error bar buried in an appendix but a visible confidence score on every output, so a thinly-evidenced number looks different from a well-evidenced one at the moment someone reads it.

Are more data sources always better?

Breadth only helps when provenance holds. A hundred sources that disagree require reconciliation against originals of record, otherwise breadth just multiplies noise. The goal is coverage with anchoring: many independent signals, each traceable to where it came from.

Can a model price a truly unusual asset?

Unusual assets are where evidence is thinnest and confidence scoring matters most. The model can still structure the problem, assemble what evidence exists and price the comparable dimensions, but a low-confidence flag on an unusual asset is the system working correctly, not failing.

Why do two models disagree about the same building?

Usually because they read different evidence: different comparable sets, different rental data, different treatment of location boundaries. Before trusting either number, ask each model the glass-box question: which sources produced this, and what happens when I trace the number back?

Related reading: Why two identical apartments price differently, and Inside the underwriting engine.