Every report shows its working. This page explains each part: where the number comes from, how the "why" amounts are calculated, how the verdict is decided, and how accurate it has been on sales it never saw.
We match the address against the NSW address file (G-NAF) and the Valuer General's sales records. Small typos in the street or suburb are corrected, and the report tells you when that happens. The match gives us the exact map position and, if the property has sold before, its last sale.
The model looks at 25 facts about the property:
The model is a LightGBM gradient-boosted tree model: 1,250 small decision trees. It starts from the postcode price level, and each tree adds a small correction based on the property's profile. It learned those corrections from 3,851,833 cleaned NSW residential sales from 1990 to November 2026. Non-market sales are left out of training: prices more than four times above or below the postcode level, and properties resold within 90 days at a very different price.
The result is the estimated sale price today, rounded to the nearest $1,000. The calculation runs in your browser: the address you type isn't sent anywhere.
Share of the model's decisions driven by each kind of information (LightGBM split gain):
The street and neighbourhood, and the property's own last sale, do most of the work. That's why estimates are more reliable for properties that have sold before.
Each line in "What moves the price" answers one question: how much would the estimate change if this property were typical for its area in that one respect? We run the model again with that single input swapped for the typical value, and the difference is the dollar amount.
Street, location and size are measured as if the property had no previous sale, because its own sale already reflects its street and position and would hide their effect. The amounts don't add up to the estimate; each shows one factor on its own.
Two messages you may see:
We compare the middle of the price guide (for "$800k – $850k", that's $825k) with our estimate. Whether a gap is big depends on how accurate the estimate usually is, so the bands come from real results: how far actual sale prices landed from our estimates on 16,308 recent sales the model never saw.
Half of all sales landed between -10% and +9% of the estimate (properties that had sold before). A guide inside that middle half is fairly valued. The further outside it, the stronger the verdict:
| Verdict | Where the guide falls among real outcomes | Sold before | No earlier sale |
|---|---|---|---|
| Undervalued | Below the 10th percentile | more than 21% below | more than 30% below |
| Possibly undervalued | 10th–25th percentile | 10–21% below | 13–30% below |
| Fairly valued | 25th–75th percentile | -10% to +9% | -13% to +14% |
| Possibly overvalued | 75th–90th percentile | 9–21% above | 14–33% above |
| Overvalued | Above the 90th percentile | more than 21% above | more than 33% above |
Properties with no earlier sale on record have wider bands, because their estimates are less certain. The verdict also tells you how unusual the gap is, for example "only 12% of sales land this far below our estimate".
In NSW, guides are often set below the expected sale price to draw buyers, especially at auction. When a guide looks undervalued, the report reminds you of this. Check the recent sales before you read it as a bargain.
"$950,000", "950k", "$900k – $950k", "1.2m", "1.2 – 1.3m", "Offers over $850,000". With a range, we use its midpoint.
Sales of the same type (houses with houses, apartments with apartments) from the last 3 months, in this order:
The summary price per m² is the median of the similar-sized sales, multiplied by this property's area. It's a quick cross-check on the model: if the two are far apart, the property or its recent neighbours may be unusual.
To test it fairly, the model was trained only on sales before 1 August 2026, then asked to price the 16,308 sales from August to November 2026. The test sales were never used to tune it.
| Group | Sales | Median error | Within 10% | Within 20% |
|---|---|---|---|---|
| All sales | 16,308 | 10.1% | 50% | 75% |
| Houses | 9,984 | 10.5% | 48% | 73% |
| Apartments | 6,324 | 9.4% | 52% | 77% |
| Sold before | 11,648 | 9.1% | 54% | 79% |
| No earlier sale on record | 4,660 | 13.3% | 40% | 65% |
So a typical estimate is about 10.1% from the eventual price; half land within 10% and three in four within 20%. Each week we also score the model on the newest sales before adding them to training: 10.5% median error for the week of 9 Nov 2026 and 10.9% for 16 Nov 2026.
Property sales: NSW Valuer General bulk property sales information, 1990 to the latest weekly file (CC BY-NC-ND 4.0). Addresses and locations: G-NAF © Geoscape Australia (CC BY 4.0). The model is retrained each week as new sales are published.
This is an automated estimate, not a formal valuation. Do your own research and get professional advice before buying.