What the data says

Studies

Questions about this market, answered from the brokerage's own sales rather than from received wisdom. Each one is reproducible: the code and the numbers live in the repository beside the workshop.

Orange County — places, and what lies under them

Three ways of dividing the same 12,992 closed sales, each painted as a region wash rather than left as dots. Two are shipped rungs: what the model reads as rung 2 — the HOA village where one exists (Ladera Ranch only, 483 sales) and the clustered tract area everywhere else — and the frozen area segment above it. Neither is refit.

The third is the derived taxonomy. Named subdivisions are the units, with a five-sale floor: a thin tract folds into its nearest named neighbour, never into a geometric box and never into a class of its own. Thirty-three columns — product per unit type, amenity, view by kind, land per home, terrain — reduce by PCA into ninety-six classes, and then adjacent units of the same class stick together within a district. 603 areas county-wide.

Colour is not arbitrary. With 600-odd areas over any ordinary palette, two different areas end up the same colour side by side and the map reads as noise. So hue is the product kind — custom estates, large detached, family detached, compact detached, the two mixed classes, townhome-and-condo, condo entry — and lightness is the price tier within it. A dot that differs from its neighbours now differs because the homes do. The ladder runs from “Custom estates, premium” (Nellie Gail, Ocean Ranch, CH Custom) and “Large detached, premium” at $1,100 per sqft (Emerald Bay, Monarch Bay Terrace) down to “Condo entry, value” at $466 (Leisure World, Laguna Village).

Underneath sit nine terrain fields from the geo workstream, each with the definition, the reading and the caveat that work authored: topographic position at 1000 m, height above neighbours, elevation, ruggedness, viewshed area, open azimuths, the road-noise field, distance to ocean, and parcel lot area. Switch one on and the taxonomy fades to a tint over it, so you can see whether a boundary follows a ridge, a road, or nothing at all.

21,530 sales › 14 cities 276 shipped villages › 9 segments › 703 derived areas in 77 classes 26 columns › 17 components
Open the map

The Price Model

Every coordinate the price model walks — what it was before this round, what it is now, and how many sales stand on each rung. Living area gains a ladder of its own with cuts that widen as they climb, the lot ratio is opened at the top, relative elevation becomes a coordinate, and one pool binary becomes two.

12 coordinates › 2 of them hurdles › 3 unit types pooling › interaction › where it gets thin U holds 3,283 slopes at 34 sales each — 8,125 at 16 derived
Read the specification

Comparability Map — derived taxonomy

The comp map built on the derived place ladder rather than the shipped one: 685 contiguous areas in 77 classes at the village and segment rungs, in place of 276 HOA villages and 9 segments. Same coordinates, same ladders, same corpus — only the taxonomy differs, which is what makes the two maps comparable.

25,081 properties › 20,190 with a recorded close 685 areas › 77 classes comp error 0.0598 vs 0.0633 shipped
Open the map

Archive — frozen versions

The pages below are dated snapshots, kept so that a rebuild can never quietly replace what an earlier reading was based on. Each is the file exactly as it stood on the date in its name.

Both correspond to the git tags model-v3-2026-08-29 and taxonomy-2026-08-29, so the code that produced them is recoverable too.

Comparability Map

Every property ever listed is a node. Click any home and it wears the black star; a fitted model then prices it from prior sales only — nothing that closed after the subject listed is allowed to vote. Its comps are the anchors the model steps from, and each one shows its chain: the comp's actual sale price, then one row per single coordinate changed — a place rung, or beds, size per bed, lot ratio, stories, year built, the quarter it sold in, a pool, a view — each priced by what the model has measured that one change to be worth here. Every comp ends at its own implied value, and their disagreement is the range.

A pool and a view became coordinates in their own right on 27 August, which means they are priced by area and unit type rather than at one flat rate: a pool runs from about +16% in one area to slightly negative in another, while a view prices much the same everywhere — ocean over none, about +11%. Condition is handled differently, because only about half of listings state it: it is applied between two homes only when both of them say what condition they are in, so a blank is never mistaken for a "no". That change took the model's median error from 5.5% to 5.1% across every sale in the corpus.

Place is a ladder of five nested rungs, neighborhood › tract › area › district › segment, where segments are classes of whole AREAS (product mix + value) — one segment links far-flung, alike markets across districts. Region washes paint tracts and the segment taxonomy on the street map. Nothing in the panel is asserted: press score these settings and the page re-prices homes that have already sold under whatever dials are on screen, and reports how far off it lands.

17,024 properties 12,648 recorded closes 53 districts › 363 areas › 1,348 tracts 9 area segments (value + era dialed) 14 cities, sales through Aug 2026
Open the map