Accuracy & method
All accuracy figures are walk-forward tests: each quarter the model is trained only on the previous 5 years of sales, then values every real registered sale in that quarter and is compared with the actual price. The model never sees the sales it is tested on.
Overall (tested quarterly since 2020Q1)
| Method | Coverage | MdAPE | ±5% | ±10% | ±20% | COD | PRD | PRB |
|---|---|---|---|---|---|---|---|---|
| HomeDash Estimate (HD-AVM 2.2) | 100% | 4.66% | 52.8% | 78.7% | 93.6% | 7.4 | 1.023 | -0.020 |
| Traditional hedonic regression | 100% | 11.34% | 24.1% | 45.1% | 73.8% | 17.4 | 1.087 | -0.055 |
| Same-building $/sq ft × index | 90% | 6.37% | 40.9% | 68.5% | 90.3% | 9.3 | 1.023 | -0.015 |
| Same flat’s last sale × index | 57% | 9.72% | 28.7% | 51.1% | 76.0% | 14.7 | 0.977 | 0.097 |
Head-to-head: sales with both building and same-flat history
| Method | Coverage | MdAPE | ±5% | ±10% | ±20% | COD | PRD | PRB |
|---|---|---|---|---|---|---|---|---|
| HomeDash Estimate (HD-AVM 2.2) | 100% | 4.32% | 55.9% | 82.3% | 96.1% | 6.4 | 1.013 | -0.018 |
| Traditional hedonic regression | 100% | 10.64% | 25.2% | 47.5% | 77.8% | 14.4 | 1.030 | 0.001 |
| Same-building $/sq ft × index | 100% | 6.21% | 41.5% | 70.2% | 91.8% | 8.7 | 1.013 | -0.002 |
| Same flat’s last sale × index | 100% | 9.60% | 29.0% | 51.5% | 76.6% | 14.3 | 0.981 | 0.091 |
MdAPE: median absolute percentage error. ±10%: share of estimates within 10% of the sale price. COD / PRD / PRB are IAAO mass-appraisal standards: lower COD = more uniform (residential ≤10–15); PRD 0.98–1.03 and PRB within ±0.05 mean no systematic bias between cheap and expensive flats.
Weekly live accuracy
Before each update, the model then in service values the sales registered after its training cut-off (sales it has never seen).
The first weekly check appears after the next update.
By amount of building data (sales in the same building, last 5 years)
| Building sales | Cases | MdAPE | ±10% | ±20% |
|---|---|---|---|---|
| 0 | 38,929 | 7.30% | 61.8% | 82.4% |
| 1-4 | 24,563 | 7.67% | 60.6% | 85.2% |
| 5-19 | 79,909 | 5.61% | 72.5% | 92.2% |
| 20+ | 252,466 | 3.97% | 84.9% | 96.5% |
80% range actual coverage: 78.4% (median width 21.6%).
By quarter
| Quarter | Sales | MdAPE | ±10% | Median estimate/price | COD |
|---|---|---|---|---|---|
| 2020Q1 | 10,520 | 4.58% | 76.9% | 1.013 | 8.0 |
| 2020Q2 | 16,737 | 3.77% | 86.8% | 0.998 | 5.7 |
| 2020Q3 | 16,315 | 3.59% | 87.1% | 1.011 | 5.3 |
| 2020Q4 | 17,428 | 3.61% | 84.6% | 1.006 | 5.7 |
| 2021Q1 | 17,788 | 3.43% | 88.2% | 1.000 | 5.2 |
| 2021Q2 | 20,683 | 3.63% | 85.0% | 0.994 | 6.1 |
| 2021Q3 | 19,681 | 4.00% | 84.4% | 0.999 | 6.0 |
| 2021Q4 | 14,976 | 4.47% | 79.8% | 0.999 | 6.6 |
| 2022Q1 | 11,206 | 4.82% | 81.2% | 1.016 | 6.5 |
| 2022Q2 | 13,800 | 4.16% | 85.3% | 0.990 | 5.9 |
| 2022Q3 | 11,320 | 4.92% | 77.2% | 1.014 | 7.1 |
| 2022Q4 | 7,996 | 5.82% | 73.6% | 1.024 | 7.5 |
| 2023Q1 | 14,655 | 5.75% | 75.0% | 0.967 | 9.1 |
| 2023Q2 | 13,385 | 4.83% | 79.5% | 1.014 | 7.5 |
| 2023Q3 | 9,220 | 5.35% | 74.1% | 1.020 | 8.7 |
| 2023Q4 | 7,372 | 6.26% | 69.2% | 1.010 | 10.5 |
| 2024Q1 | 9,642 | 5.70% | 71.8% | 1.005 | 9.0 |
| 2024Q2 | 17,861 | 5.76% | 71.0% | 1.005 | 8.4 |
| 2024Q3 | 10,334 | 5.75% | 73.0% | 1.014 | 8.8 |
| 2024Q4 | 14,562 | 5.15% | 76.4% | 0.996 | 7.7 |
| 2025Q1 | 11,725 | 4.97% | 77.0% | 0.998 | 7.7 |
| 2025Q2 | 16,309 | 4.71% | 77.4% | 0.995 | 7.4 |
| 2025Q3 | 16,549 | 4.82% | 76.1% | 0.995 | 7.7 |
| 2025Q4 | 18,904 | 5.75% | 69.6% | 1.006 | 10.6 |
| 2026Q1 | 19,385 | 4.85% | 77.9% | 0.993 | 7.3 |
| 2026Q2 | 21,824 | 5.13% | 76.4% | 0.986 | 8.0 |
| 2026Q3 | 14,436 | 5.37% | 72.6% | 0.997 | 8.4 |
| 2026Q4 | 1,254 | 4.15% | 80.8% | 1.003 | 6.8 |
Method
Data: transactions derived from Hong Kong Land Registry records — 2,297,533 after cleaning, latest registration 2026-10-09. Removed: no saleable area, nominal / related-party deals (half or double the same building’s yearly median) and duplicates.
Real-time index: each month, every sale is compared with its own building’s index-adjusted average over the past 5 years; the median change is shrunk towards the wider region. Only data up to that month is used.
Model: gradient-boosted trees (LightGBM). Key inputs are recent index-adjusted $/sq ft for the same building, stack, estate and district, plus the flat’s last sale, floor, relative floor, size, age and MTR proximity. Since HD-AVM 2.2 it also uses developers’ price lists published on the government’s first-hand sales website (the development’s recent list $/sq ft and the flat’s own list price), which improves new and recently completed flats. All inputs use only information before the valuation date.
Estimate range: an 80% range calibrated on the last four quarters of real errors, grouped by how much building data is available.
Listing price advice: 25,549 2025–2026 listings with price changes, collected from major Hong Kong online property portals, matched to registered sales. The asking premium is measured against the model estimate at listing time to learn “premium → final price” and “premium → chance of selling within 120 days”. Median sale price was 97.1% of the final ask and 93.2% of the first ask.
Expected bank valuation: 20,466 online bank valuations compared with our model to give a ratio range by price band.
Limits: only buildings with registered sales are covered; new developments and village houses are less accurate; transaction data to 2026-10-09, valuation basis 2026-10; listing data only includes listings with price changes, so selling chances are shown as relative values. Demo use only, not a formal valuation.