The Storey Premium Shrank 11%. In Tall Blocks It Did Not Move At All.

In September we measured what a high floor is actually worth in the HDB resale market, and found that comparing flats inside the same block collapses a raw 96% gap down to about 0.6% per storey. That answered the level. It did not answer whether the level has moved.

It has. Holding the block, the flat type and the calendar year fixed, one additional storey added 0.7042% to the resale price across 2017 to 2019 and 0.6246% across 2024 to 2026 — a difference of 8 basis points per floor, or 11.3% less than it used to buy, on a t-statistic of −9.5.

The same arithmetic run in dollars points the other way: a floor cost $2,691 at 2017's median price and $3,838 at 2026's. Both statements are true, and the gap between them is the finding. What a buyer pays for height has risen 43%. What height delivers, relative to the flat below it, has fallen.

What is being compared

The estimator is deliberately narrow. Every transaction is grouped by block, street, town, flat type and year, and only groups with at least three sales contribute. Within each group, price is regressed on storey after both are demeaned, so the only variation used is the same flat type, in the same block, in the same calendar year, at different heights. Everything that makes one block worth more than another — town, MRT distance, lease, estate age, model, the market level in that year — is absorbed by the group.

Storey is coded as the midpoint of the register's three-floor band. The bands are uniform at 01 TO 03 through 49 TO 51 across the entire 2017-onward series, 17 of them, with no change of banding regime, which is what makes a comparison across nine years clean.

This is not a price-level claim and not an index. It measures one thing: the slope of price against height, inside a block, and whether that slope has changed.

Year by year:

The peak is 2019 and the trough is 2025. Peak to now is 0.7481% to 0.6092%, an 18.6% relative fall. The series does not drift — it steps down in 2020-22 and then sits flat for four years.

It is not composition

The obvious objection is that this is a changing mix of blocks rather than a changing price of height. Punggol, Sengkang and the other tall post-2000 estates have been maturing into the resale market throughout the period, and tall blocks have flatter storey premiums. If that were the whole story, restricting the comparison to blocks present in both periods would kill the result.

It does the opposite. On the 12,164 block-and-flat-type cells that transacted in both windows, the premium falls from +0.7036% to +0.6074% — a compression of 9.6 basis points, larger than the 8.0 on the full sample, at t = −9.69.

The same block, the same flat type, nine years apart, pays less for height than it used to.

Nine specifications

Eight of the nine agree on sign and sit between 6 and 10 basis points. The ninth is the exception that defines the result.

The premium is front-loaded

Height is not priced linearly, and it never was. Splitting by where in the block the flat sits:

Getting off the bottom three floors is worth roughly twice as much per storey as climbing within the teens, in both periods. Most of what buyers pay for is escaping the ground, not reaching the top — and most of the compression is in that same low band, where the 2024-26 premium is 18 basis points thinner than it was.

Where it does not happen

Grouping blocks by the highest storey observed in them:

The compression is monotone in block height and reaches zero at the top. Tall blocks started with the flattest premium — a little over half a percent per floor against nearly nine-tenths in low-rise — and kept it unchanged.

So the honest headline is narrower than the national number: the storey premium compressed in low-rise and mid-rise HDB stock and did not move in the tall stock. In a 40-storey block, height is priced in 2026 exactly as it was in 2017.

Twenty-two towns of twenty-five

Of the 25 towns with at least 200 qualifying transactions in each window, 22 compressed and three widened.

The largest compressions: Serangoon −0.2424pp, Pasir Ris −0.2000, Punggol −0.1836, Yishun −0.1693, Hougang −0.1656, Jurong East −0.1588, Woodlands −0.1551.

The three that widened: Central Area +0.0756pp, Bukit Batok +0.0697, Sembawang +0.0529. Bishan, at −0.0013, is flat to three decimal places.

Central Area widening is consistent with the block-height result rather than against it: it has the flattest premium of any town in both periods (+0.3992% rising to +0.4749%) and is the most high-rise-dominated, so it belongs with the tall blocks, not with Serangoon.

By flat type, with block and year still fixed:

Three-room flats lost twice as much of their storey premium as four-room flats did. Two-room is a null result on 2,161 transactions.

Where this breaks down

Three failures, and the third is the most important.

The result dies in heavily traded blocks. Raising the minimum sales per block-cell shrinks the sample and shrinks the effect, and past 12 it is gone:

Sample sizes collapse — 1,664 and 2,630 transactions at a minimum of 12, against 37,123 and 41,580 at the baseline — so some of this is lost power. But the point estimate itself decays steadily from −0.084 to zero well before the standard errors blow up, which is not what pure noise looks like. The defensible reading is that this is a broad-market phenomenon in thinly and moderately traded blocks and is absent in the most liquid ones. Anyone who samples only high-turnover blocks will not find it.

The dollar figure does not survive the balanced panel. On the full sample, cash per floor rises from $2,691 in 2017 to $3,838 in 2026, +42.6%. On the 12,164 balanced cells it rises to $3,508 by 2022 and then stops: $3,505 in 2023, $3,578 in 2024, $3,518 in 2025, $3,644 in 2026 — +3.9% over four years, while the same cells' median price rose 20.5% from $502,000 to $605,000. The full-sample 42.6% is substantially composition, newer and pricier estates entering the resale pool. The claim that holds is "flat in cash terms since 2022, and therefore falling in real and relative terms", not +42.6%.

The quarter specification is the weakest of the nine. Tightening the group to block × type × quarter cuts the usable sample to 10,858 and 12,827 transactions and the t-statistic to −4.71. It still agrees on sign and magnitude, but it is the one row a sceptic should press on first.

Two method limits on top of those. Storey is a band midpoint, so the true floor is unobserved within three floors; this attenuates the slope toward zero in both periods, which leaves the difference broadly intact but makes every level here a lower bound. And the standard errors are conventional rather than block-clustered — clustering would widen them, and at t = −9.5 the baseline tolerates a generous inflation factor, but the quarter specification and the rows at a minimum of eight or more would become fragile.

What to do with this

If you are paying up for a high floor, the premium you are being quoted is thinner than it was in 2019 in every kind of block except a tall one. In low-rise and mid-rise stock the gap between a third-floor and a tenth-floor unit of the same type is a smaller share of price than it used to be, and the cash amount has been static since 2022.

If you are selling a low-floor flat, the compression works for you, and it is concentrated exactly where you are: the floors 1-9 band lost 18 basis points per storey, more than any other. The discount for being near the ground is narrower than the 2017-19 comparables in your own block would imply.

If your block is 30 storeys or taller, none of this applies. Height is priced there as it was nine years ago, at about half a percent per floor, and the national trend is not a reason to re-rate a high-floor unit in a tall block either way.

If you are using a rule of thumb, 0.6% per floor is the current national figure and it is an average over a sharply non-linear curve: nearly 1% per floor over the bottom nine, under half a percent above the thirteenth. Applying one number across a 40-storey block will overprice the top and underprice the bottom.

Method

Every figure was computed from the full register of 242,032 HDB resale transactions on data.gov.sg, January 2017 to October 2026, retrieved 6 October 2026 via the dataset's datastore_search endpoint and verified against the published row total. October 2026 held only 326 registrations at retrieval and is excluded from every figure, leaving 241,706 rows across complete months; September 2026 had 2,246 registrations, at the 2026 monthly mean, and is retained. The estimator is a within-group ordinary least squares slope of log resale price on storey, where the group is the unique combination of block, street name, town, flat type and year, and only groups with at least three transactions contribute; reported coefficients are multiplied by 100 and read as percent of price per storey. Storey is the midpoint of the register's storey_range band, coded as the band's lower bound plus one; the register uses 17 uniform three-floor bands from 01 TO 03 to 49 TO 51 with no regime change across the period. Standard errors are conventional within-group errors computed from residual variance over the within-group sum of squared deviations in storey, and are not clustered by block; clustered errors would be wider. The two comparison windows are 2017-2019 and 2024-2026, where 2026 is nine months. The balanced panel restricts both windows to the 12,164 block-and-flat-type cells observed in each. Dollar figures apply each year's estimated slope to that year's median transacted resale price and are a presentation of the same coefficient, not a separate estimate. Block height is the maximum storey-band midpoint observed in that block and flat type across the whole period, so it is a floor on true building height and is unobserved for blocks whose top floors never transacted. Town, flat-type, storey-band and threshold cuts re-estimate the same specification on the stated subsample; the storey-band cuts overlap deliberately so that each band contains enough within-block height variation to estimate a slope. Towns with fewer than 200 qualifying transactions in either window are excluded, which leaves 25. The register carries no data on renovation, orientation, facing, lift-landing, corner position or unit layout, and those are absorbed only to the extent they are constant within a block and flat type; nothing here separates the value of height from the value of whatever correlates with height inside a block. The month field is registration date rather than transaction date, and recent months revise upward as late registrations land. Floor area as published includes recess area and space-adding items. The September 2026 piece this builds on, A Storey Is Worth 0.6%. The Raw Numbers Say 96%., estimates the level on a narrower four-room sample; this piece estimates the change over time across all flat types and the two are not directly comparable coefficients. All central tendencies are medians, never means. This is general information about the resale register, not advice on any individual purchase or sale; confirm valuation, grant eligibility and loan terms with HDB and CPF before deciding anything.