The model is locked before each month's resale prices become public. Then it is graded against every one of them, misses included.
Loading the track record…
How this works
Locked in advance. Around the 1st of each month, that month's valuation
model is frozen and its SHA-256 fingerprint is published in the table above, before the
month's resale prices appear in the public record. The fingerprint cannot be changed
afterwards, and this page is independently snapshotted by the Internet Archive.
Scored on every sale. HDB resales register with a delay of weeks to
months. As each sale registered in a locked month becomes public, it is priced with the
frozen model exactly as the live site would have priced it at the time, and compared with
what the buyer actually paid.
Nothing is hand-picked. Every registered sale of the month is scored.
The few that cannot be scored are counted with their reasons in each month's details, so
the denominator is always visible.
Accruing, then final. A month's score updates daily while its
registrations arrive, and is locked for good about three months later. Final scores are
never revised, even if the public record is amended afterwards.
Misses are expected. The "likely range" is an 80% range: roughly 2 in
10 sales should fall outside it. If far fewer missed, the ranges would be dishonestly
wide; far more, dishonestly narrow.
One thing to be clear about: this is not a forecast of future deals. A resale registers weeks
after the deal was agreed, so the test is not "we saw the future" but something more
useful: the estimate shown on the site was committed before the outcome was public, and
here is how it did. The "old method" column scores the approach this site used before the
model (the typical price of comparable sales in the same block or street over the
previous 12 months) on the same sales, where it can price them at all.
About the estimates
The estimates come from the same model behind the fair-value card on the
map: a statistical model trained on five years of registered resale
transactions, refreshed daily, with ranges calibrated against its own measured errors.
It cannot see renovation, view, or urgency, which is exactly why the range exists. A
statistical guide, not a valuation.
The same model also powers a monthly
resale price index that compares like with like,
for Singapore and every HDB town.