How Hong Kong AVMs should report confidence (and why most do not).
Every Hong Kong AVM gives you a number. Most do not tell you how confident they are. Here is how confidence should be reported, what the global standards say, and how to read a Hong Kong property valuation.
By the QPV Founder. Published 25/05/2026.
The Hong Kong AVM transparency vacuum.
Hong Kong largest retail banks all surface automated property valuations. None of them publish how accurate those valuations are. The HSBC Property Valuation Tool surfaces a number sourced from Cushman and Wakefield Limited; the disclaimer reads "for reference only and is not binding on the bank" with no confidence interval, no methodology summary, and no backtest disclosure. Hang Seng e-Valuation surfaces "instant property valuation for your initial reference" via the same disclaimer language; Multiple Mortgage instant valuation runs off the same underlying Hang Seng service provider. None of the three publishes how the number was computed or how often it has been right.
This is the opposite of the global direction. The US Consumer Financial Protection Bureau (CFPB) AVM Final Rule, effective 01 October 2025, requires any AVM used in a mortgage credit decision to meet five published quality control standards: confidence in estimates, anti manipulation, conflicts of interest avoidance, random sample testing and review, and non discrimination compliance. The Mortgage Industry Standards Maintenance Organization (MISMO) released its AVM Common Confidence Score Standard on 21 August 2025 (formal Recommendation status April 2026), anchored on the PP10 rule (the share of valuations falling within plus or minus 10% of true market value, on a zero to 100 confidence scale). The Australian Prudential Regulation Authority APG 223 has permitted AVMs since 2014 under a risk tiered hierarchy with mandatory back testing and audit.
Hong Kong has not yet imported these standards. The HKMA Supervisory Policy Manual module CR-G-7 on Collateral and Guarantees, revised 11 April 2025, sets the framework for property collateral valuation under Basel III standards but contains no AVM specific disclosure mandate. The vacuum is the position: an independent AVM that publishes its methodology, confidence basis, and known weaknesses is, on day one, the most transparent valuation source operating in Hong Kong.
The four metrics the AVM industry uses.
Global AVM vendors publish their accuracy using a small set of standardised metrics. These are not opinions. They are defined in the IAAO Standard on Ratio Studies, the MISMO 2025 Common Confidence Score, and the CFPB rule.
- MdAPE (Median Absolute Percentage Error). Across all backtest transactions, the median of the absolute percentage gap between predicted value and actual sale price. Zillow publishes 1.74% on market and 7.20% off market (Zestimate, 2026). HouseCanary publishes 2.8% on quarterly third party tested data. ATTOM, after launching its AI rebuilt AVM in May 2026, publishes 2.9% national MdAPE. Below 5% is institutional grade.
- PP10 (Percentage within plus or minus 10%). The share of valuations that land inside a 10% band of actual sale price. ATTOM publishes 80% or above on the rebuilt model. The MISMO 2025 standard anchors on this rule. Above 70% is the institutional threshold.
- FSD (Forecast Standard Deviation). The model own confidence interval per valuation: an FSD of 0.10 means the model expects roughly 68% of true values to fall within plus or minus 10% of its estimate. CoreLogic, Veros, HouseCanary, and Clear Capital publish FSD per valuation. The CFPB rule expects regulated AVMs to report it.
- COD (Coefficient of Dispersion). The IAAO mass appraisal uniformity measure. The IAAO Standard on Ratio Studies sets a 5.0 to 15.0 acceptable range for single family residential property, with 5.0 to 10.0 representing the tighter sub range for higher uniformity portfolios. COD survives the objection that any single sale price is noisy by measuring the calibration tightness across the full distribution.
A responsible AVM publishes at least MdAPE and PP10 at headline level, and FSD at the per valuation level. Cotality Australia (formerly CoreLogic, rebranded March 2025) publishes AVM accuracy reporting across 96% of Australian residential properties, with monthly model performance and independently audited methodology. The Hong Kong incumbents publish none of these numbers. For a plain-English overview of what these metrics mean and what to demand before trusting a number, see how accurate online property valuations in Hong Kong are.
QPV benchmarks its model against recorded sale prices at every recalibration, and each valuation carries a calibrated confidence range. Those benchmark results are shared directly with clients and partners on request rather than published, because a self-reported accuracy figure is not something a reader can independently verify. The framework, the confidence outputs, and the comparable transparency described in the next section are live today.
How QPV reports confidence per valuation.
A QPV valuation report does not stop at a point estimate. Every report carries three confidence layers visible on the front page.
The confidence badge. HIGH, MODERATE, or LOW. Two things drive it: how closely the selected comparable sales match the subject property, and how wide the resulting confidence interval is. Close matches on nearby, recent, similar stock earn a HIGH badge. Thin data, distant proxies or sparse comparables push it down to LOW. The exact cut-offs are part of the model and are walked through with clients under evaluation rather than published. The badge is the first thing a broker or buyer should look at, before the point estimate.
The confidence interval. A price range, not a flat percentage. The width is derived from three components: the base similarity width (tighter for direct comparables, wider for cross district proxies), the observed basis risk between the target district and any proxy district, and a sample size penalty applied when the comparable set is sparse. The output is genuine market uncertainty, not a model comfort placeholder.
The comparable transparency. Every report lists the top 15 to 20 transactions used to build the estimate: date, floor, price, size, similarity score, building age. Anyone reading the report can verify the comparables are recent, on district, and structurally similar to the target. A confidence number without the underlying comparables is unfalsifiable; we publish both.
How a real Causeway Bay valuation reads.
A Causeway Bay apartment valued on 22 July 2026: 360 saleable square feet, mid floor, apartment type, in a building of about 53 years. The QPV report front page returned:
- Point estimate: HKD 4.11M (about HKD 11,420 per saleable square foot).
- Confidence interval: HKD 3.10M to HKD 5.13M (about plus or minus 25%).
- Confidence badge: MODERATE.
- Comparables used: 30 registered sales, the closest 15 listed in the report.
- District benchmark: priced against the Causeway Bay distribution, where the district median runs at HKD 14,898 per square foot. The subject sits about 23% below that median.
- Disclosed weakness: part of the comparable evidence comes from the neighbouring district, because the building sits near a district boundary. The report says so on the front page, and the badge is downgraded accordingly.
Point estimate inside its confidence band. The width, about plus or minus 25 percent, is the honest uncertainty.
The MODERATE badge is the point worth lingering on. This is the model marking down its own confidence and naming the reason: the flat is small, the building is old, and some of the supporting evidence comes from the district next door. A bank AVM serving the same property might return one number and disclose none of that. Because QPV surfaces it, a buyer or broker knows the band is wide for a reason and can decide whether to commission a surveyor follow up, hunt for fresher same-building comparables, or treat the range as their negotiation corridor. A narrower band would have looked more impressive and been less true.
A three step checklist for buyers and brokers.
When you receive any AVM valuation, whether from a bank, an estate agent, or an independent tool, run it through three questions before relying on the number.
One: where is the confidence band? If the report shows only a point estimate, treat it as a guideline, not a valuation. Every credible AVM publishes a range. The width of that range carries more information than the centre point. A HKD 10M valuation with a plus or minus 5% band tells you something completely different from a HKD 10M valuation with a plus or minus 25% band.
Two: how many comparables, and how similar are they? A valuation built on 20 same district transactions at 0.95 similarity is a different object from a valuation built on 5 transactions including a cross district proxy. Ask for the comparable list. A responsible AVM publishes it. If the provider declines, the confidence in the number drops by definition.
Three: how recent is the freshest data? Hong Kong residential prices fell roughly 24% from the September 2021 peak to mid 2025, before reversing in the past 12 months. A valuation built on comparables from before the most recent rate cycle is operating on a different market. The freshest comparable date should be on the report front page, not buried in an appendix.
For more on how to compare an AVM valuation to a chartered surveyor report, see the AVM vs surveyor decision framework. For the deeper definition of an AVM and how Hong Kong bank tools differ from independent ones, see AVM Hong Kong: a complete guide.
The regulatory direction is clear.
Three regulators have already moved. The CFPB AVM Final Rule in the United States took effect on 01 October 2025, jointly issued by the OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA. The APRA APG 223 in Australia has required risk tiered AVM governance with mandatory back testing since 2014. The MISMO Common Confidence Score Standard, published August 2025, is the cross vendor standardisation that lets banks compare AVM accuracy claims on like for like terms.
The HKMA has not yet issued an equivalent rule. The most recent module touching property valuation is CR-G-7 Collateral and Guarantees V.2, issued 11 April 2025, with full compliance expected by 10 April 2027. It incorporates Basel III property valuation standards and aligns with Financial Stability Board guidance on mortgage underwriting, but does not require published AVM accuracy disclosure. The direction of regulatory travel suggests this will change in the next regulatory cycle. When it does, the AVM providers already publishing methodology and confidence will be compliant by default; those operating as black boxes will face a retrofit.
For brokers, family offices, and property owners who need a defensible valuation today, the practical implication is simple: choose tools that already publish what the regulator is heading toward. The cost is identical. The trust is not.
What people ask about AVM confidence.
What is an AVM confidence interval?
A confidence interval is the price range around an AVM point estimate that reflects the model uncertainty. A HKD 4.11M valuation with an interval of about plus or minus 25% means the model expects the true market value to fall between HKD 3.10M and HKD 5.13M for typical similar transactions. The width of the interval is set by the quality and quantity of comparable transactions, not by a flat percentage.
What is a good AVM confidence interval for Hong Kong property?
For liquid Hong Kong residential districts with 20 or more recent comparable transactions, a half width of plus or minus 10% to plus or minus 15% is typical for a high confidence valuation. Thin data districts, luxury properties, and village houses naturally have wider intervals (plus or minus 20% or more). Anyone publishing a flat percentage regardless of the property is not doing confidence reporting.
Why do Hong Kong bank AVMs not publish accuracy?
Hong Kong does not yet have an AVM specific regulation requiring disclosure. The HKMA Supervisory Policy Manual sets a general bank valuation framework but does not mandate the published methodology and accuracy disclosure that the US, Australia, and global standards bodies (MISMO Common Confidence Score Standard, August 2025) now require for regulated AVMs.
What is the difference between MdAPE and PP10?
MdAPE (Median Absolute Percentage Error) is the median value of how far model valuations are from actual sale prices, expressed as a percentage. PP10 is the share of model valuations that land within plus or minus 10% of actual sale price. MdAPE tells you the typical accuracy. PP10 tells you how reliable the model is on most transactions. Together they describe the shape of the error distribution.
How is QPV confidence calculated?
QPV derives confidence from three factors: the similarity score between the target property and its comparables (zero to one), the basis risk between the target district and any proxy district used for thin data cases, and the sample size of available comparables. The model outputs a HIGH, MODERATE, or LOW confidence badge based on these inputs, plus an explicit price range. The full computation is published at the methodology page.
See QPV methodology in detail.
Walk through the inputs, weights, and confidence framework, or read the broader AVM Hong Kong guide next.