Sixteen Towns Rose on the Medians. Twenty-Two Fell on Matched Blocks.

On 1 October, HDB published its flash estimate for the third quarter: the Resale Price Index at 202.4, down 0.2%, and in HDB's own words "the third consecutive quarter of decline in resale prices, following decreases of 0.1% and 0.3% in the 1st and 2nd Quarters of 2026 respectively."

Over the same stretch, the median price of an HDB resale flat went up. Not by much — $630,000 in the second half of 2025, $635,000 across April to September 2026, a gain of 0.79% — but up, and in the opposite direction to the index that is supposed to measure the same market.

Both numbers are correct. They disagree because a median is a statement about which flats sold and an index is an attempt at a statement about what flats are worth. There is a third way to settle it, and the full transaction register supports it: find the blocks that sold flats in both windows and ask what happened to the price of a near-identical unit. That estimator says −1.11%.

What a matched block is

Take every combination of block, street, town and flat type — 345 Clementi Avenue 4, four-room, say. Keep only the combinations that recorded at least one sale in the second half of 2025 and at least one in April to September 2026. That leaves 4,284 cells. Compute the median price in each window for each cell, take the per-cell change, and report the median of those changes.

It is not a hedonic index and it does not pretend to be. What it does is hold the address and the flat type fixed, which is most of what makes one flat worth more than another, and it is immune to the thing that moves a national median: a change in which towns and which flat types happen to transact.

Tighten the cell and the answer barely moves:

Four specifications, a range of 21 basis points, and fewer than 41% of cells rising in every one of them. The level of the decline is modest. Its direction is not in doubt within this method.

For four years it agreed with the medians. This year it does not.

This is the exhibit that matters, because the obvious objection to a matched estimator is that it is systematically biased — that matched blocks are older, or thinner, or otherwise not representative, and so they always read lower. If that were true, the gap would show up every year.

Run the identical construction on each of the preceding four years, same window shape, same six months against the following year's April to September:

In four consecutive years the two measures agreed on direction, and in three of those four the matched measure read higher than the median, not lower. The share of blocks rising sat between 74% and 86% throughout. This year it is 40.8%, and the sign splits for the first time.

So the estimator is not biased downward. Something changed in the market it is measuring.

Thirteen towns change sign

Per town, with towns under 60 sales in either window or under 20 matched cells excluded:

The medians put 16 of 25 towns up. Matched blocks put 22 of 25 down. Only Central Area, Marine Parade and Geylang are positive on the matched measure, and Central Area carries 37 cells on 84 and 96 sales — thin enough that its +28.86% naive figure should be read as a statement about which units transacted, nothing more.

The error has a shape. In the mature towns the naive median overstates — Clementi, Toa Payoh, Bedok, Tampines, Kallang/Whampoa all read positive on medians and negative on matched blocks. In the newer towns it understates: Sembawang, Jurong West, Yishun and Sengkang all look worse on the median than their matched blocks do. The mix error reverses direction with town maturity, which is why a national median cannot be corrected with a single adjustment.

The decline sits in short leases

Cut the same comparison by remaining lease at transaction:

The matched column is monotonic with no exceptions: the shorter the lease, the harder the fall, from −2.77% to −0.26%. The naive column is not monotonic at all, and three of its five bands carry the wrong sign, including the newest flats, where the median says +2.61% and matched blocks say −0.26%.

We have priced the level of the lease discount at 60 years before. This is the change in it over the past year, and it is the cleanest gradient in the data.

What moved the median

Two things, and neither is price.

Between the two windows the share of transactions that were five-room, executive or multi-generation flats rose from 29.07% to 30.51%. Median remaining lease at transaction fell from 71.6 years to 70.9. More big flats sold, so the middle of the distribution sat higher, while the flats themselves were marginally shorter-leased.

The simplest cross-check is price per square metre, which carries its own problems but is at least indifferent to flat-type mix in a way the price median is not. Across the same windows the median fell from $6,517 to $6,465 per square metre, −0.81%. Quarter on quarter, third quarter against second, the median price rose from $630,000 to $638,000 while median price per square metre fell from $6,505 to $6,436. The quarter that produced a "median up 1.3%" headline produced a price per square metre down 1.1%.

Where this breaks down

One honest failure, and it is in the robustness check rather than the main result.

Requiring more sales per cell shrinks the sample fast, and at the fourth step the sign reverses:

The result is stable from 4,284 cells down to 266. On the 67 cells that sold at least four flats in each window it flips positive. Those 67 cells are the very largest and most liquid blocks in the country, so one reading is that the biggest blocks genuinely held their prices while everything else slipped. The other is that 67 cells is too few to carry any weight. We cannot distinguish between those with this data, and anyone citing the −1.11% should know the figure does not survive that particular squeeze.

What to do with this

If you are selling, price against the last transactions in your own block and flat type, not against your town's median. In 16 of 25 towns the town median is currently pointing the wrong way relative to what matched blocks did, and in the mature towns it is pointing optimistically.

If you are buying and have been told the market has turned up, the index, the price per square metre and the matched blocks all say otherwise for the twelve months to September. The median is the one series that says up, and it says up because the sales mix shifted toward bigger flats.

If your flat has under 55 years left, this is where the softness is concentrated — −2.77% on matched blocks against −0.26% for flats over 85 years. That gradient has been the structural story for years; what is new is that it is now negative all the way along the lease curve rather than only at the short end.

If you are waiting for a bottom, note that volume is not behaving like a falling market: 13,599 transactions in April to September 2026 against 11,959 in the second half of 2025. Prices drifting down on rising volume is a different situation from prices falling on thin trade, and it is not the pattern that usually precedes a sharp move either way.

Method

Every figure was computed directly from the full register of 241,920 HDB resale transactions on data.gov.sg, January 2017 to October 2026, retrieved 5 October 2026. October 2026 held only 212 registrations at retrieval and is excluded from every figure, leaving 241,708 rows across complete months. The two comparison windows are July to December 2025 (n = 11,959) and April to September 2026 (n = 13,599); the gap between them is deliberate, so that the windows do not share a quarter. A matched cell is a unique combination of town, block, street name and flat type present in both windows; the per-cell statistic is the within-window median price, the reported figure is the median of per-cell ratios, and the tightened specifications add storey band and exact floor area in square metres as stated in the table. The back-series applies the identical construction to the four preceding year-pairs with the same window shape. Remaining lease is parsed from the register's own remaining_lease field and banded as shown; the lease table's matched column uses the same cell definition restricted to each band. Towns with fewer than 60 sales in either window or fewer than 20 matched cells are excluded, which drops only Lim Chu Kang. This is a matched-unit estimator, not a paired-household or repeat-sale one: two sales in one block and flat type are almost never the same flat, so the correct description is "near-identical units at the same address", and it controls for neither storey nor floor area nor flat model except where the table says so. The register carries no buyer field, no grant data and no CPF data, so nothing here speaks to who is buying, to net cash outlay, or to the effect of the 28 July 2026 removal of the 15-month wait-out period. The month field is registration date rather than transaction date, so quarter boundaries are registration boundaries and recent months revise upward as late registrations land; September 2026 is still filling. Floor area as published includes recess area and space-adding items, which adds noise to every per-square-metre figure. The Resale Price Index is quality-adjusted and is not reproducible from this file; the 202.4 flash estimate, the −0.2% change and the quoted sentence are from HDB's flash estimate release of 1 October 2026 and flash estimates are revised in the full release later in the month. All central tendencies are medians, never means. This is general information about the resale register, not advice on your own purchase or sale; confirm policy, grant eligibility, valuation and loan terms with HDB and CPF before deciding anything.