Data and sources
What Crescira shows, where each part comes from, how fresh it is, and what it leaves out.
The dates on this page are read from the data now being served, which was built on 2026-11-16.
How fresh it is
| What | Covers | Newest record | Records |
|---|---|---|---|
| Registered sales | since 1995-07 | 2026-11-13 | 2,535,806 |
| Monthly price trends | months that have ended | 2026-10 | — |
| Provisional sale agreements | 2026-03-26 to 2026-11-16 | 2026-11-16 | 4,846 |
| Rentals | 2021-01-01 to 2026-11-16 | 2026-11-16 | 211,108 |
| Buildings | — | — | 63,468 |
| Units | — | — | 1,273,315 |
| Crescira Index | 1995-07 to 2026-10 | 2026-10 | — |
| Unit valuations | as of that month | 2026-10 | — |
The data is rebuilt every night. Trend charts leave out the month in progress. The newest month they do show is still filling in, because sales reach the register some weeks after they are agreed.
Where it comes from
Sale, rental, building and unit records are licensed from our data supplier, credited at the foot of every page.
The places near each building, which feed the quality-of-life scores and the maps, come from:
- Food and Environmental Hygiene Department: licensed restaurants · 17,284
- Lands Department: iGeoCom places (schools, shops, sports grounds, bus stops, car parks and more) · 13,484
- Supplied with the property records · 1,457
- Other Hong Kong government open data (parks, ferry piers) · 72
How to read the figures
- Every price is a registered sale, not an asking price. Sales reach the register some weeks after they are agreed, so the newest weeks are always incomplete.
- A monthly median describes whatever sold that month, so it moves when the mix of flats sold moves. On the Hong Kong and district charts, the bars show how many sales the line rests on, and median price and median price per square foot are computed over the same sales.
- The Crescira Index values the same flats through time, so it is steadier than a median and can point the other way. Year-on-year figures come from the index and describe a district, never a single building.
- A valuation is a model's estimate from registered sales. It is not an offer, and not a bank's or a surveyor's valuation. Its accuracy has been measured, and each figure is quoted with how it was tested.
How accurate the valuations are
These are all the measurements we have made, not the best of them. The two tests answer different questions, and neither figure means much without how it was tested.
| How it was tested | Data of | Median error | Within 10% | Sales tested |
|---|---|---|---|---|
| Forecasting 2026-08, a month the model had never seen | 2026-10-10 | 10.03% | 49.7% | 2,262 |
| Forecasting 2026-07, a month the model had never seen | 2026-08-29 | 8.97% | 55.5% | 2,598 |
| Filling in held-back sales within a period it knows | 2026-10-10 | 4.14% | 78.1% | 6,697 |
| Filling in held-back sales within a period it knows | 2026-08-29 | 4.46% | 77.0% | 6,484 |
- Forecasting a month never seen: the model is trained only on sales before that month, then values that month's sales. This is the stricter test.
- Of the 2,262 sales tested for 2026-08, 271 were in newly launched buildings the model had never seen a sale in, the hardest case there is.
- Filling in held-back sales: a random tenth of the last 24 months' sales is held back from the model, which then values them. Other sales from the same weeks stay in its training data, so this is the easier test.
- Which sales are tested: both tests draw on the sample the model is trained from, about half of the residential sales that have a published saleable area.
- Median error: half the valuations were closer to the registered price than this, and half were further away.
What we don't show
- Asking prices and listings. The site holds registered sales, and provisional sale agreements that are labelled as such.
- The month in progress. Trend charts and headline figures stop at the last month that has ended, and that month's own count keeps growing for some weeks.
- A price per square foot where no saleable area is published. Subsidised-housing sales (Home Ownership Scheme and similar) carry a price and no area, so they have no price per square foot and are left out of its median.
- A valuation for a single-unit building (a house). The model values a flat by comparing it within its block: floor, position and the building's own sales. A house has none of those comparisons, so none is shown. A house recorded with more than one unit still gets one.
- A year-on-year change for a single building. One building sells too few flats a year for a figure worth trusting, so the site shows its district's figure and says it is the district's.