Accuracy, published
How accurate are our estimates?
Every estimate site asks you to trust a number. We measured ours against 38,116 real NSW sales and publish the results — including the segments where we're weak. As far as we know, no other Australian property estimate site does this.
14.3%
Median error
Half of estimates are closer than this
61.9%
Within ±20%
Share of estimates within 20% of the sale price
70.5%
Range hit rate
Sales that landed inside our shown low–high range
11.4%
Known sale history
Median error when we know the property's own last sale
How we measured it (and why you can trust it)
- Real sales, hold-out replay. We took a deterministic 1-in-100 sample of NSW Valuer General residential sales (2001–2026) and asked: what would our model have said the day each property sold — using only transactions recorded before that day? No look-ahead, no cherry-picking.
- The exact code that serves this site. The backtest imports the same estimator module the live pages call. When we improve the model, we re-run the backtest and update this page.
- Nothing excluded. Every recorded residential sale counts — including family transfers and other non-market deals we can't detect. Our real-world accuracy on ordinary sales is likely slightly better than these numbers.
- Metrics. “Median error” is the median absolute difference between estimate and actual sale price, as a percentage of the sale price.
Accuracy depends on what we know
When a property has sold before, we anchor to its own history adjusted for market movement — much more accurate than area comparables alone. Each estimate on this site tells you which basis it used.
| Segment | Sales tested | Median error | Within ±10% | Within ±20% | In shown range |
|---|---|---|---|---|---|
| Property with known sale history | 16,089 | 11.4% | 45.6% | 70.3% | 70.4% |
| Comparable sales only | 22,007 | 17.2% | 33.3% | 55.8% | 70.6% |
By dwelling type and price
| Segment | Sales tested | Median error | Within ±10% | Within ±20% | In shown range |
|---|---|---|---|---|---|
| Houses | 24,561 | 14.7% | 37.2% | 60.7% | 68.6% |
| Units / apartments | 13,549 | 13.6% | 40.8% | 64.1% | 74% |
| Price under $500k | 17,266 | 14.9% | 37.5% | 60.1% | 71.5% |
| Price $500k–1m | 13,517 | 11.9% | 44% | 68.6% | 76.2% |
| Price $1m–2m | 5,264 | 16.5% | 33% | 57.8% | 62.2% |
| Price over $2m | 2,063 | 24.1% | 25.4% | 43.9% | 46.4% |
Recent years
The model gets more accurate as transaction history accumulates.
| Segment | Sales tested | Median error | Within ±10% | Within ±20% | In shown range |
|---|---|---|---|---|---|
| 2019 | 1,324 | 13.9% | 37.6% | 64.1% | 70.7% |
| 2020 | 1,553 | 13.3% | 41.3% | 65.3% | 72.2% |
| 2021 | 2,078 | 14.9% | 37.6% | 60% | 64.8% |
| 2022 | 1,494 | 14.9% | 37% | 61.2% | 66.8% |
| 2023 | 1,523 | 14% | 39.5% | 62.8% | 68% |
| 2024 | 1,623 | 13.5% | 40.7% | 63.9% | 69.3% |
| 2025 | 1,695 | 13.5% | 39.3% | 63.6% | 70.1% |
| 2026 | 550 | 14.1% | 39.1% | 64.7% | 66.9% |
What we don't know (yet)
- No building attributes. Bedrooms, bathrooms, renovations and condition aren't in public transaction records. Commercial data providers with those attributes report median errors around 6–9%; publishing our number is how we'll prove we close the gap as richer data lands.
- Per-transaction sales cover 2001–2026 (through May 2026). Every estimate is labelled with its own data year, so you always see how current the evidence behind it is. Coverage refreshes as each new NSW Valuer General release lands.
- NSW only for now. Victorian suburb medians are live; per-property VIC/QLD estimates follow the same path once sales data is licensed.
Backtest generated 2026-07-10 · 38,116 target sales · source: NSW Valuer General property sales information (CC BY 4.0). Estimates are statistical model outputs, not professional valuations.
See it on a real address
Search any NSW address — the estimate shows its basis and evidence, free.