● Methodology Read in ~4 minutes

Priced like the market, tested like it matters.

Three commitments hold everything else up: the model never sees the seller's ask, it is only ever graded on sales it never trained on, and a new model ships only when it proves itself on real closes.

Rule one

No list price. Ever.

List price is what a seller hopes. Training on it produces a tool that echoes the ask. It is banned from the model's inputs, the valuation comes from what the property is, where it sits, and what comparable buyers actually paid.

Rule two

The past predicts the future

Testing is chronological: the model trains on earlier sales and is graded on later ones, the way it will actually be used. Random splits flatter a model. Ours is scored on 3,126 held-out BC sales, refreshed as the market moves.

Rule three

Better or it doesn't ship

Every retrain competes against the reigning model on the same held-out sales, and is promoted only when it wins. What reaches you is always the strongest model we've produced.

● Inputs

A signal for every question
a good appraiser would ask.

What is it? Where is it? What's the land worth versus the building? What's the market doing this month? What did nearly-identical properties close at? Each signal earns its place, every one measurably sharpens the estimate.

Each signal earns its place, the full feature list is proprietary.

PropertyLiving area, beds, baths, lot, age, garage, waterfront, strata, floor
ViewsOcean, city, mountain, lake, from MLS records and listing remarks
ConditionKitchen and bath quality scored from listing photos; renovation state from remarks
AssessmentBC Assessment value, land-to-total ratio, assessment staleness
ComparablesNearest-sold medians, prior sales of the same property, repeat-sale trend
LocationSub-area, postal code, and distance to water per parcel
MarketAbsorption, DOM, price-cut depth, supply shift, as of the offer date, not the close
MacroBoC overnight rate, mortgage benchmark, yield curve, CPI, BC employment

● Continuous improvement Validated on held-out BC sales

Sharper every
month.

The production model is refined continuously, each improvement validated on thousands of BC sales it never trained on before it ever reaches you.

Every release is proven on ~3,100 chronologically held-out BC sales: the model is always tested on the future, never the past.

Jun 29 Baseline model
Jul 02 Log-price training + nearest-sold comparable features
Jul 02 Market state joined at offer date instead of close date
Jul 03 Kitchen & bath condition scored from listing photos
Jul 06 Deeper BC Assessment features, land/improvement split and staleness
Jul 06 Leasehold and fractional tenures excluded for accuracy
Jul 06 Suites, renovation state, and views read from listing remarks. Current model

● Two models, two questions

“What is this property worth?”

The core valuation model. Prices any Island property from its fundamentals, no listing required, no ask consulted. This is the number for CMAs, acquisition targets, and portfolio marks.

R² 0.80

“Will it sell at this ask?”

A second model reads every active Island listing against its ask: predicted close price, over/under-pricing flags, and the odds of an offer in the next 30, 60, and 90 days.

Typically within 1.75% of close price

● Prove it to yourself

Pick the address. We'll show the work.

Request a demo