Measured, not claimed
How accurate is it?
Every valuation tool in Singapore will give you a number. None of them will tell you how often that number is right. Independent comparisons have found the major portals disagreeing by tens of thousands of dollars on the same property, with no way for you to judge which to believe. So here is ours, measured and published, including the parts that make us look worse.
Headline
3.8%
Median error across all property types
79%
Land within 10% of the price achieved
1,600
Blind valuations tested
By property type
| Segment | Tested | Median error | Within 10% | Within 20% | Bias |
|---|---|---|---|---|---|
| All property | 1,600 | 3.8% | 79% | 93% | -0.4% |
| HDB flats | 800 | 3.0% | 91% | 99% | -0.0% |
| Private residentialCondominiums, apartments, executive condominiums and landed | 800 | 4.9% | 68% | 86% | -1.1% |
Private residential is materially harder than HDB and our figures say so. HDB stock is near-identical block to block and trades densely, so comparable evidence is abundant. Private developments differ in ways that thin evidence cannot always price. We publish the weaker number because a page that reported only the flattering one would not be worth reading.
By the confidence we stated
| Stated confidence | Tested | Median error | Within 10% | Within 20% | Bias |
|---|---|---|---|---|---|
| High | 1,588 | 3.8% | 80% | 93% | -0.4% |
| Medium | 12 | 12.9% | 42% | 58% | +1.4% |
| Low | Not enough observations to report | ||||
This is the table that matters most. A confidence label is only worth printing if it predicts the error, and here it does: valuations we call High are several times more accurate than those we call Medium. When we tell you we are unsure, we are right to be.
Method
- · Each sale in the test was valued as of the month before it closed. The engine saw no comparable dated on or after the sale it was predicting, and no price index that already reflected what happened next.
- · The test covers the 6 months to 2026-09, sampling 800 sales per segment with a fixed seed, so any figure here can be reproduced exactly.
- · 1,600 valuations were attempted and 0 were declined for want of evidence. A refusal is counted separately, not scored: declining to answer is the correct behaviour when the comparables are not there, and folding those into the error rate would reward a model for guessing.
- · Scoring a model against sales it was built on measures memory rather than accuracy. We verify the cutoff on every run, both directly and by confirming that the same sales score better when the engine is allowed to see the future.
What these numbers do not say
- · The benchmark is the price actually transacted, which is not the same as market value. A sale can be above or below it for reasons no model can see, and that counts against us here.
- · Singapore is an easier market to value than most. Caveats are lodged for every sale, the geography is small and much of the stock is standardised. Comparing this figure with an overseas tool flatters us.
- · A median error means half of all valuations are further out than the number shown. Use the range and the confidence band, not the point estimate alone.
- · This is an Estimated Market Value for guidance. It is not a formal valuation and must not be relied on for a mortgage, probate or any legal purpose, which require a licensed valuer.
Measured 2026-09-13 against data through 2026-09. Re-measured automatically after each nightly data refresh. Try a valuation.