Bank risk teams and chartered surveyors ask the same question first: how accurate is the model? US AVM providers publish a number. Hong Kong bank AVMs do not. This is what the metrics actually mean, what global providers report, and what QPV will publish on Hong Kong residential transactions.
By the QPV Founder. Published 18/05/2026.
One number: MdAPE. The median absolute percentage error between the AVM estimate and the eventual sale price. A model with MdAPE of 3 percent means the typical valuation is 3 percent off the sale price. Below 5 percent is institutional grade.
One share: PP10. The percentage of valuations that land within plus or minus 10 percent of the sale price. A model with PP10 of 75 percent means three quarters of its estimates are within 10 percent of the truth. Above 70 percent is the institutional threshold (also written PPE10 in some literature; same metric).
One confidence interval: FSD. The forecast standard deviation, reported per valuation rather than across the model. FSD of 0.10 means there is approximately a 68 percent probability the actual value falls within plus or minus 10 percent of the estimate. US regulators expect AVMs to report it so that low-confidence valuations can be routed to a chartered surveyor.
These three numbers, together, describe whether a model is safe for a bank to use, sharp enough for a broker to argue with, and conservative enough for a buyer to trust. The rest of this article is what each one is, who publishes what, why Hong Kong publishes nothing, and what QPV does about it.
| Metric | What it measures | Institutional benchmark |
|---|---|---|
| MdAPE | Median absolute percentage error of AVM estimate vs sale price | Below 5 percent |
| PP10 (PPE10) | Share of valuations within plus or minus 10 percent of sale price | Above 70 percent |
| PP15 | Share within plus or minus 15 percent | Above 85 percent (used where comparables are sparse) |
| FSD | Forecast Standard Deviation per valuation; FSD of 0.10 implies ~68% probability the truth is within plus or minus 10 percent | FSD below 0.10 = high confidence; above 0.20 = escalate |
| Hit rate | Share of requested properties the model can return an estimate for | Above 90 percent for mature US AVMs |
| COD | Coefficient of Dispersion (IAAO mass-appraisal uniformity) | 5 to 10 for newer or similar residential; 5 to 15 for heterogeneous residential |
| PRD | Price-Related Differential (IAAO bias check; do cheap homes get over-valued vs expensive ones) | 0.98 to 1.03 |
The two metrics that travel furthest are MdAPE and PP10. MdAPE is consumer-readable ("our typical valuation is 3 percent off the eventual sale price"). PP10 is bank-readable ("seventy-five out of every hundred valuations land within ten percent of the truth"). Both are needed: MdAPE without PP10 hides the tails; PP10 without MdAPE hides the median. COD and PRD belong to mass-appraisal regulators (the International Association of Assessing Officers) and are required reading for anyone building toward institutional credibility, but consumers and brokers rarely encounter them.
This is what major AVM providers publish on their own websites, as of May 2026. All figures are vendor-published; some are independently tested quarterly, some are not. Quantarium, Veros, Clear Capital and ICE / Black Knight publish per-account testing rather than aggregate figures and are not in the table below.
| Provider | Coverage | Published accuracy |
|---|---|---|
| Zillow Zestimate (US) | ~104 million US homes | MdAPE 1.74% (on-market) / 7.20% (off-market), as published on zillow.com. Consumer tool, not certified for lending. |
| HouseCanary (US B2B) | 120 million US properties | MdAPE 2.8%. Independently tested quarterly. |
| ATTOM (US, AI-powered AVM, launched 05/05/2026) | 158 million US properties | MdAPE 2.9% / PP10 80%+ (new model). Prior model: MdAPE 4.3% / PP10 70%. |
| Cotality Australia (formerly CoreLogic, rebranded 24/03/2025) | 96 percent of AU residential | ~90 percent of estimates within 15 percent of sale price (PP15), per Cotality Australia. |
| HSBC HK e-Valuation | Not disclosed | Not disclosed. |
| Hang Seng Bank e-Valuation | Not disclosed | Not disclosed. |
| Multiple Mortgage instant valuation (HK) | Not disclosed | Not disclosed. |
| QPV (HK, this article) | V1 dataset: 7,096 HK residential transactions across 35 plus districts | Methodology published now; backtested MdAPE and PP10 in the V2 refresh, 2026. See Section 6. |
The pattern is consistent across markets where AVMs are mature: providers publish a number, third parties test it, lenders use the disclosure to size their reliance. The gap in Hong Kong is not the absence of valuation technology (HK banks use automated and surveyor-sourced valuation tools); it is a disclosure gap.
Three regulatory anchors define the global direction of AVM disclosure.
United States: Interagency AVM Rule, effective 1 October 2025. Joint rulemaking by the Office of the Comptroller of the Currency, Federal Reserve, FDIC, NCUA, CFPB and FHFA imposes five quality-control standards on any mortgage originator or secondary-market issuer using an AVM to value a property securing a consumer principal dwelling. The five standards: ensure a high level of confidence in estimates, protect against data manipulation, avoid conflicts of interest, require random sample testing and reviews, and comply with applicable non-discrimination laws. The rule is principles-based, not numeric-threshold-based; it shifts the burden of proving AVM quality onto the institution using the model.
United States: MISMO AVM Common Confidence Score Standard, published August 2025. The Mortgage Industry Standards Maintenance Organization (the standards body for the US mortgage industry) standardised a 0 to 100 confidence score anchored on the PP10 rule. The standard exists so that confidence numbers from different AVM vendors can be compared on the same basis. It is the operational complement to the Interagency Rule.
Australia: APRA Prudential Practice Guide APG 223. The Australian Prudential Regulation Authority allows AVMs as one of several acceptable mortgage valuation methods, alongside desktop assessments, kerb-side assessments and full inspections. An authorised deposit-taking institution using an AVM must analyse the model strengths and weaknesses, back-test on random samples, monitor on an ongoing basis, and maintain audit arrangements with the AVM provider. The core principle is risk-tiered: as the loan-to-value ratio rises, the rigour of the required valuation rises with it.
Hong Kong: no AVM-specific rule as of May 2026. The HKMA Supervisory Policy Manual governs property mortgage valuation at the framework level. Module CR-G-7 (Collateral and Guarantees, latest revision 11 April 2025) and module CR-G-13 (Counterparty Credit Risk Management) shape collateral practice but do not name AVMs, prescribe accuracy thresholds, or require disclosure of model performance. The HKMA October 2024 macroprudential standardisation set a flat 70 percent LTV cap on all residential property; AVM methodology was not part of that package. Hong Kong has 149 licensed banks plus 16 restricted-licence banks and 11 deposit-taking companies, supervised by the HKMA. None of them publishes an AVM accuracy disclosure.
The direction of travel is one-way. US and Australian regulators have taken positions. Hong Kong has not. Banks, brokers and buyers operating in Hong Kong today rely on AVM outputs that no public benchmark verifies.
Bank risk teams. The right question to ask any AVM vendor (internal or external) is not "what is your MdAPE" but "what is your MdAPE, on what test set, vs what ground truth, last refreshed when". The credible vendors answer all four. Pair MdAPE with PP10 to see the distribution, not just the median; pair both with FSD per valuation so you can route low-confidence properties to a HKIS-registered chartered surveyor. The US Interagency Rule five quality-control standards are a useful operational checklist even where no equivalent HK rule exists.
Mortgage brokers. When a bank valuation comes in below the agreed transaction price, the question is whether the bank number is genuinely low or fair. An independent AVM with a published MdAPE and PP10 gives the borrower a second data point. If the independent AVM lands meaningfully higher with a tight confidence band (FSD below 0.10), the broker has grounds to ask the bank to commission a fresh chartered surveyor valuation or to shop the deal to a competitor lender. Without a published accuracy benchmark from the second AVM, the conversation is one opinion against another.
Property owners and buyers. Ask whether the tool you are using publishes an accuracy benchmark and how recently it was computed. A consumer AVM that publishes nothing is a starting point for ballpark only, not a number to bring to a negotiation. A tool that publishes MdAPE and PP10 with a date and a sample size is usable in a conversation with your bank or your broker. The Hong Kong bank AVMs you encounter through your bank online portal do not publish either.
HKIS-registered chartered surveyors. AVM accuracy disclosures do not replace what you do; they raise the bar on the AVM market that surrounds you. A well-disclosed AVM that gets the easy ninety percent of cases right, and is explicit about its uncertainty on the hard ten percent, returns the surveyor profession to where its judgment is genuinely required: non-standard properties, legal proceedings, strategic transactions, and any case where the AVM confidence range is too wide to support the decision.
QPV will publish its accuracy benchmark against the Land Registry recorded sale price as the ground truth: MdAPE, PP10 and COD on a defined Hong Kong mass-market residential slice. The framework, the per-valuation confidence outputs, and the comparable transparency are live today. The headline accuracy numbers follow in the V2 refresh later in 2026, once the walk-forward backtest harness is complete, so that every figure is computed against recorded sale prices with no look-ahead leakage. The V1 dataset behind the model already covers 7,096 Hong Kong residential transactions across 35 plus districts, with 914 properties valued end to end.
QPV benchmark, status
MdAPE, PP10 and COD: publishing in the V2 refresh, 2026. Computed against Land Registry sale prices, with a date on every figure and no look-ahead leakage.
Three things will define the QPV benchmark when it publishes.
QPV is built benchmark-first: the methodology and the confidence framework are published before the headline accuracy number, not after, so the number arrives with its method already in the open. The figures will tighten as the dataset grows, as Land Registry integration matures, and as more Hong Kong districts move from sparse-comparable to dense-comparable coverage. The full methodology behind every QPV valuation is published at quantpv.com/methodology.
A published MdAPE and PP10 answer "how often is the model close on average". They do not answer four other questions that matter just as much.
Coverage by property type. Aggregate MdAPE can hide that a model is excellent on Taikoo Shing flats and weak on Sai Kung village houses. Always ask what the test set looks like. A model whose test set is dominated by one estate type is excellent at that type and unproven elsewhere.
Time-to-recompute. A model with MdAPE of 3 percent measured eighteen months ago is not the same model today. Markets move, comparables age, and machine-learning models drift if they are not retrained. Published accuracy needs a date and a refresh commitment.
Distinguishing easy cases from hard cases. MdAPE is a summary across all cases. FSD per valuation is what tells the user, on this property, today, how much to trust the number. A model that publishes MdAPE without per-valuation confidence is hiding which estimates are safe and which are not.
What the model was trained to optimise. A model trained to minimise MdAPE on past sale prices is good at predicting past sale prices. That is not the same as predicting what a bank will value the property at, what a court would accept as fair-market value, or what a HKIS surveyor would write down. The metric is real; the use case has to match the training target.
QPV publishes the benchmark in Section 6 with these limits stated up front. Any AVM provider, in Hong Kong or anywhere else, that publishes a single accuracy number with no caveats is asking the reader to take more on trust than the number supports.
PP10 is the share of AVM valuations that fall within plus or minus 10 percent of the eventual sale price. It is the industry shorthand for how often an AVM lands close to the right number. PPE10 is the same metric written differently; PP10 is the form used by MISMO, US federal regulators and US AVM testing firms. A mature institutional AVM typically reports PP10 in the 70 to 90 percent range; below 70 percent the model is generally not considered fit for credit decisions.
MdAPE is the median absolute percentage error: the median absolute percentage gap between each AVM estimate and the actual sale price, summarised in one number. PP10 is the share of valuations within a fixed tolerance (10 percent). MdAPE answers how close the typical valuation is. PP10 answers how often the AVM is close enough to be usable. The institutional benchmark is MdAPE below 5 percent and PP10 above 70 percent.
No. As of May 2026, HSBC, Hang Seng, Bank of China, Standard Chartered, Citi, and consumer AVM services such as Multiple Mortgage publish neither their methodology nor any aggregate accuracy number (MdAPE, PP10, FSD, COD, or otherwise). Hong Kong has 149 licensed banks under the HKMA framework; none publishes an AVM accuracy disclosure comparable to what Zillow, Cotality or HouseCanary publish in the US. The HKMA does not currently require it.
FSD is the forecast standard deviation, a per-valuation confidence measure expressed as a percentage. An FSD of 0.10 means there is approximately a 68 percent probability the actual value falls within plus or minus 10 percent of the AVM estimate. US AVM providers Cotality, Veros, HouseCanary and Clear Capital all publish FSD on every estimate. The US Interagency AVM Rule, effective 1 October 2025, expects regulated AVMs to report a confidence measure of this kind so that lenders can route low-confidence properties to a chartered surveyor.
A standardised 0 to 100 confidence scale published by the Mortgage Industry Standards Maintenance Organization in August 2025, anchored on the PP10 rule (plus or minus 10 percent of market value) so confidence numbers from different AVM vendors can be compared on the same basis. The standard supports compliance with the US Interagency Rule effective 1 October 2025.
Historically, HK mortgage valuation has been dominated by HKIS-registered chartered surveyors performing physical or desktop inspections under HKMA Supervisory Policy Manual guidance. AVMs are used by individual banks as an internal first-pass screen but have not been the subject of a formal accuracy or disclosure rule. The result is a vacuum: AVMs are used, but no public benchmark exists for how good they are. QPV publishes its methodology and will publish its accuracy benchmark to fill that gap.
Request access to the platform, or read how each QPV valuation is constructed and tested.