Four ways to measure whether Singapore property prices went up, and why they disagree by 18 points

Ask how much Singapore property has gone up and you will get a number. Ask two people and you may get two numbers eighteen points apart, both computed correctly from the same transactions.

Here are four, from our own record of private condominium and apartment resales between 2017 and 2025.

Method	2017 to 2025
Median price per square foot	+45.9%
The same, holding the 2017 mix constant	+47.2%
The same, holding the 2025 mix constant	+51.5%
Repeat sales, the same units sold twice	+33.5%

The highest and the lowest are 18 percentage points apart, over nine years, on one market. That is not a rounding difference or a data problem. Each method is answering a slightly different question, and the answers separate because the questions do.

This series takes the methods one at a time: what each one measures, how it is built, where it is the right tool, and what it cannot see. Last updated 9 September 2026.

What each method is actually asking

A median price asks what the middle transaction cost. It is the only measure that tells a buyer what to bring to the market, and it is the one most often quoted as though it measured prices. It cannot separate a change in prices from a change in what happened to sell.

A median price per square foot asks the same question after normalising for size. That fixes one confounder and leaves the rest. Our own notes show it moving far faster than total price on the same transactions: suburban resale two-bedrooms rose 38.9% in price and 59.3% per square foot over the same nine years, because the homes were shrinking.

A repeat-sales index asks what happened to the same home, by matching a unit's sale against its own earlier sale. Mix cannot fool it, because the two legs are the same property. It pays for that in three ways, and this series states all three with numbers.

A hedonic index, which is what URA and HDB both publish, asks what a property with identical characteristics would have cost in each period, estimating the contribution of age, size, floor and location and holding them fixed. It is the most sophisticated of the four and it controls only for the attributes in the model.

Why they disagree here

The obvious explanation is that the mix of what sold changed, and on this data the obvious explanation accounts for almost none of it.

Hold the composition of the market fixed, by region, tenure, bedroom count and building age, and median psf growth moves from 45.9% to 47.2%. That is the whole mix effect: 1.4 points out of a gap of 13.8 to repeat sales. Composition did move against the headline, because the market traded toward the outside central region, toward leasehold and toward older stock, all cheaper by the foot. It just moved a tenth as much as it would need to in order to explain anything.

There is a second lesson in the same calculation. Holding the mix constant is itself two different answers. Freeze the 2017 basket and you get 47.2%; freeze the 2025 basket and you get 51.5%. Both are standard, both are correct, and they are four points apart, because a basket from the start of a period and one from the end are not the same basket. A method that exists to remove ambiguity has ambiguity inside it.

And the fourth method is not looking at quite the same homes: a repeat-sales index can only see a unit that sold twice, and half of those first sales were from a developer rather than an owner.

So the gap is not composition. Twelve of the thirteen points sit inside the methods themselves. A repeat-sales index carries the ageing of every unit between its two sales, cannot see a renovation done in between, and only observes homes that sold twice. Each note in this series takes one of those apart.

The methods this series covers

Each note goes up as it is written, and this page links them as they land.

• Averages, medians and the fix between them. Why a mean is worse than a median for this, and what stratifying gets you before you reach for anything harder.
• Median price per square foot. What normalising for size fixes, and the comparisons it still reverses.
• Repeat sales. Holding the unit constant, and the three prices you pay for it.
• Property price indices, explained. How URA and HDB build a hedonic index, why the two are not comparable, and why a median can fall in a quarter an index says rose.
• The methods Singapore does not use. Sale price appraisal ratio, which works from assessed values, and asking-price indices, which lead transactions.

If you measure this for a living, or are learning to

This series is written by a practitioner working from transaction records, not by a research department, and there are questions in it we cannot settle from caveat data alone.

The gap above is the clearest of them. Composition accounts for about a tenth of it, and we can name four candidate causes for the rest: ageing embedded in the repeat-sales construction, renovation between the two sales, quality drift inside our comparison cells, and the selection of which homes sell twice at all. We cannot yet say how much each contributes.

There are others. Singapore publishes an annual value for every property, which makes a sale price appraisal ratio index technically possible here, and nobody builds one. Nobody has published a repeat-sales index for the HDB market, where the data would support it. And the question of how to compare growth across the public and private markets, measured by two differently constructed indices, has no good answer we have found.

If you work on housing price measurement, in a university, an agency or a research team, and you think one of these has been solved, or solved better elsewhere, we would like to hear it.

And if you are a student or just learning your way around this, the same invitation holds. Every method in this series is explained from the beginning, the questions above are open rather than rhetorical, and a question that makes us explain something better is worth as much to us as an answer. Write to hello@realila.com either way. We will publish what we learn and say where it came from.

What this series is not

It does not argue that one method is correct. Each of the four is the right instrument for a question, and the wrong one for the others, and the failure mode we see most often is a number computed by one method being used to answer another method's question.

It also does not publish a Realila index. Every figure here comes from the same URA caveat record the rest of the site is built on, computed by whichever method the note is about, so a reader can see what each method does to the same underlying transactions.

Method and data notes. Sale data to 1 September 2026.

Figures on this page cover private condominiums and apartments, resale transactions, 2017 to 2025, from URA caveat data. Median price per square foot is the caveat's own price per square foot, medianed. The mix-adjusted figures hold the composition of region, tenure, bedroom count and building age fixed, one at the 2017 basket and one at the 2025 basket, and reweight the other year to it. The repeat-sales figure is computed over units sold twice, matched on unit identity. Executive condominiums are excluded from all four, since their subsidised first sale is not comparable to an open-market one.

The three price-per-square-foot figures are computed on the same lattice of 95 cells present in both years, so they are comparable cell for cell: 10,935 resales in 2017 and 12,050 in 2025. The repeat-sales figure is not on that lattice and cannot be. It matches a unit's sale against its own earlier sale, and for over half of the 100,896 pairs that earlier sale was the developer's, not a resale. Restricting it to resale-to-resale pairs inside the lattice would discard 62% of the pairs and break the method to make a sentence tidy. So three of the four methods see the same slice of the market and the fourth sees a different one, which is part of why they disagree.

Sale data to 1 September 2026.