The Thai Earnings Call Nobody Reads: 5,755 SET Opportunity Day Sessions, Transcribed and Searchable Across the Whole Market

There is a disclosure event in Thai equities that is public, free, quarterly, held by nearly every listed company, and read by almost nobody.

It is not a filing. It is a room.

Every quarter the Stock Exchange of Thailand hosts Opportunity Day — now run through the Exchange's Listed Company Presentation portal — where management presents results and then takes questions from analysts, live. The recording goes up publicly. Anyone can watch it.

And that is precisely the problem.

A filing is written. An Opp Day is answered.

The distinction is the whole argument, so it is worth being exact about it.

A Form 56-1 One Report is drafted. It goes through legal, through the audit committee, through investor relations. Every sentence in it is a sentence somebody chose to write, reviewed, and signed. That is what makes it reliable, and it is also what makes it incomplete: nothing in a filing is obliged to answer a question the company would rather not be asked.

An Opp Day is different in kind. An analyst asks about the margin compression in the second half. Management answers in the room, out loud, on the record, with no drafting cycle between the question and the answer. Sometimes the answer is a number that never appears in any document. Sometimes it is guidance given verbally and never printed. Sometimes the informative part is the hesitation — the question that gets a long preamble and no figure.

None of that reaches the filing. It is not misconduct; it is just the difference between a document and a conversation.

Why the format defeats ordinary research

Suppose you decide to do it properly and actually watch them. Three walls arrive in order.

The first is time. A session runs an hour or more. Coverage of thirty names is thirty-plus hours a quarter before you have thought about a single one of them.

The second is that video is not searchable. You cannot Ctrl+F a recording. If you want to know whether management discussed customer concentration, you scrub — and you scrub through a presentation deck being read aloud to reach the Q&A, which is where the informative part usually is.

The third wall is the one that actually matters, and almost nobody names it. Even if you had transcripts of all of them, the question you want to ask is not "what did this company say?" It is "what did every company in this sector say about this, in the same quarter, and who said something different from the rest?"

That is a cross-sectional question. No recording answers it. No single transcript answers it. Reading thirty transcripts in sequence does not answer it either, because by the time you reach the thirtieth you are no longer holding the first one in your head. It is the same structural gap that makes the One Report unreadable at market scale — except here the raw material is not even text yet.

What we built

Four years of Opp Day sessions, transcribed by AI, indexed beside the filings, and searchable with the same query that searches the Form 56-1.

Measured on 21 September 2026, against the rolling four-year window:

That is 99.9% of eligible sessions, and 95.1% of every session in the window. Both numbers are real and they are not the same number, which is a distinction the next section is entirely about.

When a passage comes back, it carries a deep link into the recording at the timestamp where it was spoken. You do not have to trust our transcript. You click, and you hear the person say it.

What the coverage number means, and what it does not

Most products would publish the 99.9% and stop. Here is the rest of it, because the gap between those two figures is where an honest reader's next question lives.

99.9% is a fill rate over eligible sessions. 95.1% is the same work measured against every session in the window. The 275-session difference is not a backlog and not a failure — it is a bucket our pipeline deliberately declines, and as of 21 September 2026 essentially all of it resolves to three structural causes:

• Property and infrastructure trusts. A REIT files a 56-REIT, not a Form 56-1 One Report. Our filing axes — financials, risk, auditor, governance, business — have no rows for those tickers at all, so a session with nothing to cross-reference against is skipped rather than served thin.
• Sessions that are not a company. The Exchange also hosts group seminars run by the analyst association, presented by sector rather than by issuer. There is no single issuer to attach them to.
• Recent listings, which have not filed enough yet to cross-reference against.

None of that bucket comes from audio quality, background noise, or the language spoken in the room. We know, because we went and looked at each one.

And it is not a verification rate

This matters more than the fill rate, so it gets its own heading.

Elsewhere on this site we publish a verified share for the filing axes — the proportion of findings matched back, word for word, to text in the original document, with the passage boxed on the page it sits on. Anything that fails is withheld and counted out loud.

Opp Day findings are not counted in that figure, and they never will be. The evidence for a spoken passage is not a box drawn on a page of a PDF; it is a link to the minute of the recording. Two different kinds of proof. We measure them separately, we label them separately, and we do not add them together. A product that blended a transcription fill rate into a document-verification rate would be inflating the number that its entire pitch rests on.

The four rules our code enforces — not our editorial policy, our code

An AI transcript of a financial conversation is a liability unless somebody has thought hard about what the machine is allowed to do with it. These are enforced in the serving layer, and a row that violates one does not reach a customer.

1. Do not alter. The text served to you is the raw speech. Internally the system keeps the search representation and the displayed text as two separate fields precisely so that nothing done to make a sentence searchable can leak into the sentence you read. A guard check asserts this on every index build.

2. Every row is labelled as an AI transcript. Not a footnote at the bottom of a page — a required field on the row itself. The excerpt-policy check rejects any row that does not carry the disclosure, so an unlabelled machine transcript cannot physically be served.

3. Numbers must appear on the slide. This is the strongest guard in the pipeline and the reason it exists: speech recognition on Thai financial vocabulary gets units wrong in ways that are catastrophic and invisible — "9,200 million baht" can come back as "9,200 baht" with perfect confidence. So a figure extracted from speech must also appear on the presentation deck published for that same session, and it is separately magnitude-checked against our own financials corpus. A number that fails either test does not get served.

4. No invented words, and no invented severity. Where the correction layer touches a term, the replacement has to exist in a closed vocabulary and stay close to what was actually said — the model is allowed to fix a garbled word, never to author a new one. And there is no severity score on any Opp Day row. Not a low one. The field is null, and a guard asserts it is null, because a machine-assigned "how bad is this" over a spoken sentence is exactly the kind of fabricated authority this product exists to argue against.

Where this is genuinely weak

Three things, stated plainly, because you will find them yourself within a day of using it.

Thai financial speech is hard for ASR, and we do not claim otherwise. Ticker names, accounting terms, and numeric units are the failure modes, in that order. Every design decision above — the timestamp link, the slide-deck gate on numbers, the mandatory AI label — exists because the transcript is a lead, not a citation. Treat a passage as "go listen to this minute", which is what the deep link is for.

The window is four rolling years, not ten. The ten-year history on this platform belongs to the filings corpus, which is a different body of documents with a different verification standard. Opp Day is a leading indicator — what management is saying now — and it rolls forward with the calendar rather than accumulating into an archive.

A session that is not cross-referenced is not served. See the trusts and the group seminars above. We would rather skip 275 sessions than serve 275 sessions with nothing behind them.

The questions this actually answers

The point is never one company. It is the cross-section:

• Which managements raised input-cost pressure this quarter, and which ones were asked about it and changed the subject?
• Who guided down verbally before it appeared in any written disclosure?
• Across a sector, who is describing the same demand environment in materially different language?
• When a theme first appears in the room, how many quarters pass before it reaches the filing?

That last one is only answerable because the two corpora sit in the same index. The room and the document, same query, side by side.

Try it, or put it in front of a desk

The free tier opens the two most recent Opp Day rounds with no card — enough to check the transcripts against the recordings yourself, which is exactly what we would want you to do first. Open it at terminal.boomleverage.com.

Paid packages widen the window: two years, four years, and the full archive on the desk licence, which also carries multiple seats, bulk export with the source passages attached, and API access. If you are deploying this across a research desk or an asset manager, the scoping conversation is at boomleverage.com/th/enterprise or straight to contact@boomleverage.com.

Disclaimer: This article is written for education and to explain how to read disclosures that listed companies make publicly. It is not investment advice, it recommends no security, and it guarantees no outcome. · Opportunity Day recordings and presentation decks are published by the Stock Exchange of Thailand and its listed companies; we link to the original recording at the timestamp and do not re-host it · Opp Day passages in our product are produced by AI speech recognition and are labelled as such on every row. They are a pointer to a moment in a public recording, not a certified transcript, and they are deliberately excluded from the word-for-word verified share we publish for document-based findings · No severity score, risk score, or ranking is assigned to anything said in an Opp Day session, and none of the relationships described here has been backtested against market outcomes · All figures in this article were measured against the index as of 21 September 2026 and change as new sessions enter the archive · Classification and extraction are the output of AI systems, not the opinion of a licensed professional · Investing carries risk; past results do not guarantee future ones