How accurate are online property valuations in Hong Kong?
It depends on the property. A well built model can price a liquid Hong Kong estate flat within about 10 percent of its eventual sale price most of the time, and is far less reliable on a village house or a Peak home. The number that matters is not the headline accuracy claim but how often a tool is wrong, and whether it will tell you.
By the QPV Founder. Published 05/07/2026.
Accuracy is three questions, not one.
"How accurate is it" sounds like one question. It is really three: how close does the valuation land to the eventual sale price, how often does it land that close, and how honest is the tool about the cases where it fails. A number without the other two is marketing.
As a benchmark, mature automated valuation models in liquid markets (Zillow in the US, Hometrack in the UK) publish accuracy of roughly 80 to 90 percent PPE10, meaning 80 to 90 percent of their valuations land within 10 percent of the actual sale price. That is a realistic ceiling for a statistical model on a liquid, well traded property. Hong Kong's most heavily traded estates (Taikoo Shing, Mei Foo, City One) sit closest to that ceiling. The problem in Hong Kong is not that good accuracy is impossible. It is that almost no consumer tool tells you what its accuracy is. For the full picture of how these models work, see our guide to the automated valuation model in Hong Kong.
The metrics that actually measure accuracy.
The valuation industry does not judge accuracy by a single percentage. It uses a small set of standard metrics, and knowing them is the fastest way to tell a serious tool from a marketing one.
| Metric | What it measures | What good looks like |
|---|---|---|
| PPE10 (hit rate) | Share of valuations within 10 percent of the sale price | 80 to 90 percent on liquid estates |
| MdAPE | Median absolute percentage error across all valuations | Lower is better; single digits on liquid stock |
| FSD | Forecast standard deviation, the spread of the error | Tighter spread means a more trustworthy range |
| COD | Coefficient of dispersion, the IAAO uniformity measure | Under about 15 for uniform residential valuations |
PPE10, MdAPE and FSD are the standard reporting metrics for automated valuation models. The coefficient of dispersion comes from the International Association of Assessing Officers (IAAO) ratio-study standards, the global reference for judging whether mass valuation is uniform and unbiased. None of these is exotic. They are the vocabulary any credible valuation provider should be able to speak. If you want the deeper version, we cover the accuracy maths in AVM accuracy explained: MdAPE, PP10 and the Hong Kong gap.
The practical rule: any tool that claims a precise accuracy figure, "95 percent accurate", without naming the metric, the sample size, and the market it was measured on, is overstating. Regulators are moving the same way. The United States Interagency AVM Rule, in force since 1 October 2025, requires lenders to hold their automated valuation models to documented quality-control standards. The direction of travel is disclosure, not a headline percentage.
An honest valuation is a range with a confidence level.
The single most useful thing to know about a valuation is not the number. It is the width of the range around it. A tool that returns "HK 9.2 million" and nothing else is hiding its own uncertainty. A tool that returns "HK 8.7 to 9.6 million, high confidence" is telling you both the estimate and how much to trust it.
This matters because two competent valuations of the same Hong Kong flat legitimately differ, and neither is wrong. The reasons are mechanical: different comparable sales, different weighting of floor and view, different treatment of saleable area. We unpack exactly why in why three surveyors give three valuations. The takeaway for accuracy is that a point estimate presented as a single truth is less honest than a range presented with a confidence level, because the range is where the real information lives. This is the argument for reading a valuation as a confidence band rather than a single figure.
Why accuracy varies so much across Hong Kong.
Accuracy is not a single property of a tool. It is a property of the tool applied to a specific building. The variable that moves it most in Hong Kong is comparable density: how many similar units in the same or a neighbouring block have sold recently. Where that evidence is thick, a model is accurate. Where it is thin, no model can be.
| Property type | Comparable density | Realistic online accuracy |
|---|---|---|
| Large liquid estate (Taikoo Shing, Mei Foo) | High, many recent same-block sales | Strong |
| Mid-size private estate | Moderate | Good, wider range |
| Older walk-up or single block | Low | Weak, treat as indicative |
| Village house or Peak home | Very low, few true comparables | Unreliable without a surveyor |
This is why a blanket accuracy claim is misleading in Hong Kong: the same tool can be strong on a Mei Foo unit and near useless on a village house. District matters too, which we cover in how property valuation differs by Hong Kong district, and the hardest case of all has its own guide, valuing a New Territories village house. The honest tool widens its range on these properties rather than pretending to a precision it does not have.
What Hong Kong valuation tools actually tell you.
Here is the gap. Reviewing the mainstream Hong Kong valuation sources against the disclosures that let you judge accuracy, most publish almost none of them. They return a number, sometimes a range, and stop.
| Source | Output | Confidence range | Method published | Accuracy metric published |
|---|---|---|---|---|
| Agency estimators (CentaEstimate, Midland) | Point estimate | No | No | No |
| Bank internal AVM | Point estimate | No (internal) | No | No |
| Chartered surveyor | Single figure with report | Narrative, not statistical | Per instruction | Not applicable |
| QPV | Range | Yes | Yes, at /methodology | Confidence per valuation; benchmark published as the dataset matures |
None of the incumbents is doing anything improper. Agency tools are lead generators, and bank models are internal risk tools that were never built to be shown to the public. But it does leave a real gap for anyone who wants to know how much to trust the number. Professional valuation standards already expect reasoning to be shown: the International Valuation Standards (IVS) and the RICS Red Book require a valuer to state the basis and the assumptions, and Hong Kong Institute of Surveyors (HKIS) members work to the same discipline. An online tool that shows its method and its confidence is simply applying that same standard to a statistical valuation.
Five things to check before you trust a number.
You do not need to be a statistician to judge an online valuation. You need to ask five questions, and to distrust any tool that will not answer them.
- Does it give a range, not just a point? A range with a confidence level is more honest than a single figure.
- Will it show the comparables? You should be able to see the recent sales the estimate is built on.
- Is the method written down? If you cannot read how it prices, you cannot judge whether to trust it.
- Does it get wider on hard properties? Honest accuracy degrades visibly on village houses and thin-comparable blocks.
- Is it independent of the lender? A valuation with no stake in the mortgage decision is a cleaner second opinion.
This is the standard QPV is built to. The full approach is set out at /methodology: standardised inputs, explainable outputs, a confidence range on every valuation, and the comparables behind it, with a published accuracy benchmark that grows with the dataset rather than a headline percentage asserted on day one. The QPV product walks through what a sample report looks like end to end.
Accuracy also has a limit that no online tool can cross. On high-value, unusual, or legally sensitive properties, the right move is a HKIS-registered chartered surveyor, and the AVM versus surveyor decision sets out when each is required. And if it is a low bank valuation that sent you here, the practical options are in what to do when your bank valuation comes in low.
What people ask about online valuation accuracy.
How accurate are online property valuations in Hong Kong?
It depends on the property. On a liquid estate flat a good model lands within about 10 percent of the eventual sale price most of the time, in line with mature-market AVMs at 80 to 90 percent PPE10. On village houses and Peak homes, where comparable sales are scarce, accuracy falls sharply. The reliable signal is whether the tool publishes how often it is wrong.
What accuracy number should I look for?
A published hit rate such as PPE10, stated with the sample and the market it was measured on. A claim of "95 percent accurate" with no metric, sample size, or period is not usable. Professional mass valuation is judged by IAAO ratio-study measures such as the coefficient of dispersion, not by a marketing percentage.
Why do two websites value my flat differently?
Different comparable sales, different weightings, and different treatment of floor, view and saleable area. Two competent valuations of the same property legitimately differ, and the thinner the recent evidence for that building, the wider the honest range. A point estimate hides this; a confidence range shows it.
Can I trust a free online valuation for my mortgage?
Treat it as a second opinion, not a binding figure. The bank sizes the mortgage on its own valuation. Use an independent online valuation to sanity-check that number, and on high-value or unusual properties confirm with a HKIS-registered chartered surveyor.
See how QPV reports accuracy.
Read the methodology behind every valuation, or go deeper on the accuracy maths and what the metrics mean.