The four roles are not four flavours of the same thing. Each one gives up something specific to buy something else, and the pool's job is to hold the trade-offs in balance rather than to maximise any single line.

At ฿50–75M and roughly five units, the mix is genuinely constrained — five units cannot express a subtle allocation. These are the three realistic shapes, and subscribers see which one a close is built on before they commit.

A foreigner — a natural person or a juristic person — may own a condominium unit in freehold, registered in its own name, provided the total floor area held by foreigners in the building does not exceed 49 % of saleable area. The remaining 51 % stays Thai-owned. The rule has stood since 1979.

One mechanical condition comes with it: the buyer must evidence an inward remittance of foreign currency into Thailand of at least the purchase price. That is the FET form, covered in Tax & exit.

The deciding line is anti-avoidance exposure. Singapore's Section 10L, in force since 2024, can tax gains on disposal of foreign assets received in Singapore where the structure lacks economic substance — precisely the profile of a passive holding vehicle. Hong Kong has no equivalent that bites as directly on this use case.

Malaysia's treatment of foreign gains is workable, but recognition among European subscribers is weaker and the treaty position needs case-by-case verification. It offers no net advantage over Hong Kong for this structure.

The stated objective is to avoid a regulated structure — no management licence, no authorisation, and none of the fixed costs that come with them. That is achievable, but the boundary is precise and does not depend on what the arrangement is called.

Three variables determine it: the number of subscribers, the minimum ticket, and how subscribers are solicited — public solicitation flips the arrangement regardless of participant count. And all three are assessed in each subscriber's jurisdiction, not the company's.

Two points are worth stating plainly because they are counter-intuitive. First, SBT and stamp duty are mutually exclusive — a transaction bears one or the other, never both, and for a company it will almost always be SBT.

Second, the SBT exemption after five years of ownership, available to an individual with registered domicile, does not apply to a juristic person. That is a structural cost of holding through a company: roughly 5–6 % of price at exit, carried in the model rather than discovered later.

When foreign currency is converted to baht for the purchase, the Thai bank issues a Foreign Exchange Transaction form (FET, formerly Thor Tor 3) and reports it to the central bank above a USD 50,000 equivalent.

It is used twice. At purchase, it conditions registration in the foreign quota. At sale, it conditions repatriation: the bank requires the original FET, the Land Office sale agreement, and evidence the withholding tax was paid. Repatriating up to the amount originally brought in is procedurally clean; repatriating the gain above it requires additional documentation but is not blocked.

Role 4 rests on buying units discounted because they are dated, renovating, then letting or reselling higher. If renovation spend is capitalisable into the cost base, it reduces the taxable gain on disposal and role 4 becomes materially more attractive. If not, its weighting comes down.

The treatment is being confirmed with Thai tax counsel on two points: whether the spend is capitalisable against the disposal base, and whether it is deductible at the operating level against rental income. Until that is settled, role 4 is weighted on its commercial merit alone — the measurable gap between as-is and renovated prices in the same building — with no tax benefit assumed.

Any tax outcome in a subscriber's own country depends on their residence and on how a distribution is characterised in their hands. That is a question for their own adviser, and nothing in this dossier should be read as answering it.

A district where one developer lists 200 identical units will have its median crushed by that single building. The statistic then describes one seller's commercial strategy, not the market. Median per building first, then median of buildings: one building, one vote.

A valuable side effect: this neutralises age, floor, view and orientation, which vary enormously within a building and far less between buildings in the same district. It is what makes a defensible figure possible even though year of construction is recorded nowhere.

The yield quoted everywhere else is a ratio: district median rent over district median price. Where a district mostly sells large units and mostly rents studios, that ratio describes no real property. Here rent and price come from the same building. Where the pairing is too thin, the figure is flagged rather than smoothed.

Hard to copy: it requires collecting both markets simultaneously, with building-name normalisation good enough for the two sides to find each other. Most players cover one side only.

A listing that disappears and reappears under a new ID looks like a sale followed by a relisting. Most often it is an automatic repost to climb back up the results. Counting disappearances as transactions massively overstates absorption. Tracking runs on a cohort — building × bedrooms × size band × type — rather than an ID. If the cohort's stock is unchanged, nothing was absorbed.

The same bounds once existed in three copies that ignored each other. The measured result: properties excluded from one table still counted in the medians, and an "opportunities" list led with rentals miscategorised as sales, therefore at −100 % of expected price — a source artefact presented as the best deal on the market.

Bounds are now defined in one place, and an agent verifies every cycle that the application and the database say exactly the same thing. Out-of-bounds property is not a deal, it is data to exclude — and it stays visible: statistics are filtered, nothing is deleted.

A query once identified "1,399 exact duplicate live listings", 8.7 % of stock. Deduplication was ready to be written. Ten rows were inspected first: the IDs were consecutive at the source and the listings were simultaneously live. They were not duplicates — they were distinct units in a new building, posted in bulk by one agency, at the same price because they are identical.

Deduplicating would have erased real supply — precisely what seller pressure is supposed to measure. Nothing is merged or deleted automatically; ambiguous cases go to human review and influence no statistic until cleared. The general lesson runs through the whole system: an aggregate count does not tell you what it counted. "1,399 duplicates" was a correct query and a false conclusion.

First observation 2026-06-21. No claim of trend, seasonality or cycle is supportable at this depth. What is solid today is the photograph of the market and the method; the series is a documented promise, not a deliverable.

The reverse is worth stating in both directions: a competitor starting tomorrow gets tomorrow's snapshot, never the history. Delisted listings cannot be collected retroactively. A price series does not backfill — it accrues at one day per day.

The database gives an average duration between first observation and delisting. That figure must not be presented as time-on-market: first observation is not publication, it is when collection happened to cross the listing. On a daily cycle the measure is bounded by our own rhythm, not the market's.

One source's declared publication date was tested: median gap against first observation was −16 days — the listing seen sixteen days before its declared publication. The field actually measures being bumped to the top of results, so it measures agents' diligence rather than the market. It was discarded.

One source changed its page structure and the collector kept running without error while returning zero listings, for several days. That is structural — a third-party site does not give notice. What changed is that the break is now detected the same day and escalated. The risk is not removed, it is made visible quickly.

Collection also runs on a single machine that must be powered on at scheduled times. A hardware failure interrupts the series, and an interrupted series does not backfill. A full archive is replicated locally, which protects the history already acquired — but not the continuity of collection.