The Translation Gap: Why Vietnam's 1,500+ Listed Companies Need Native-Language AI Research

Vietnam's securities disclosure rules contain one sentence that decides what every English-language filing in this market is actually worth, and it is not in the fine print — it is in Article 5, the article that defines the language of disclosure (consolidated text, Điều 5):

The rule, verbatim (Điều 5, khoản 1): "Trường hợp có sự khác biệt hoặc có cách hiểu khác giữa thông tin bằng tiếng Việt và tiếng Anh thì thông tin bằng tiếng Việt được áp dụng."

Rendered into English — by me, which is itself the point: where there is a difference, or a different reading, between the information in Vietnamese and in English, the Vietnamese information applies.

Read that again with an analyst's hat on. The English filing exists. It is required. And the regulation that requires it says, in advance, that it loses to another document if the two ever diverge.

What I have to disclose before anything else: this piece is published by a company building a search engine for exactly this problem, and our Vietnamese coverage today is zero companies and zero filings. Nothing below is a demonstration of a product you can use in Vietnam this morning. It is the argument for why we are building it, with the state of the work stated where you can check it. Credibility in this business comes from saying what you do not have yet.

A market that arrives on a calendar, and coverage that does not

On Monday 21 September 2026, at the market open, Vietnam stops being a frontier market in the classification most index money follows. FTSE Russell confirmed the reclassification to Secondary Emerging in April 2026, with inclusion running in four tranches from that date through September 2027.

That is a rare thing: demand with a calendar attached. Passive flows do not need a view — they buy because the index says to, on the date it says, in the size it says. Every other pound of capital does need a view, and the research required to form one arrives on no schedule at all.

The market those flows are arriving into is not small, and it is not one venue:

Now ask the question a research director asks: how many of those have an English-language research note published on them in the last twelve months? Not the top thirty — those are covered, competitively, by people who have been in Ho Chi Minh City for fifteen years. The question is about the other fourteen hundred.

This is what information asymmetry looks like written down. It is not that the information is missing. Every one of those companies files. The information is published, complete, and unreachable — a different problem, with a different solution.

What the rule actually requires, and when

Most foreign investors get the shape of this wrong in one of two directions: either "the documents are all in Vietnamese" (no longer true) or "they publish in English now, so it is fine" (not what the rule says). The regulation is precise, so it is worth being precise back.

Circular 68/2024/TT-BTC — issued 18 September 2024, in force 2 November 2024 — amended the disclosure rules of Circular 96/2020/TT-BTC to require bilingual disclosure, and it did so on a phased roadmap rather than all at once:

Three consequences follow, and each one changes how a research process should be built.

1. The English version is not the document of record

That is the sentence at the top of this piece. It is not a translator's disclaimer buried in a footer; it is how the disclosure regime is written.

Consider what it means operationally. Every number you cite, every commitment you underwrite, every phrase you read as a change in management's tone — in English, you are reading a rendering the rule itself designates as losing to another document. In a market with a mature research culture, an analyst who built a thesis on a text explicitly designated as non-governing would be expected to disclose it as a limitation of the work.

2. The English archive is about a year deep. The Vietnamese archive is a decade deep.

This is the part that quietly destroys the most valuable kind of analysis.

Reading one filing tells you what a company says about itself today. Reading the same section across five years tells you what changed: when the hedging language appeared, when a risk described in one sentence grew into three paragraphs, when "temporary" stopped being used about a margin problem. That comparison is where the signal lives, and it is the reason a serious desk reads the Báo cáo thường niên and the Board of Management's discussion inside it rather than a summary of either.

Look at the roadmap table again. English periodic disclosure starts in 2025. English ad hoc disclosure starts in 2026 for the large caps and 2028 for everyone else. So the English record available for a five-year comparison is roughly one to two years deep, against a Vietnamese record that runs for a decade.

A rule from model-validation work: if your history is shorter than your comparison window, you do not have a weak signal — you have no observation. A backtest that silently truncates its own history is not conservative; it is wrong, and it is wrong in the direction that flatters you. The same applies to reading language across years: you cannot read a five-year arc in an archive that starts two years ago, no matter how good the reader is.

3. The obligation does not cover every company you can buy

The 2025 and 2026 dates apply to listed organisations and large-scale public companies. That is a defined set, and it is not the same set as "every company whose shares a foreign investor can own" — the rest of the public-company universe is phased in during 2027 and 2028.

A market where the top of the exchange discloses in two languages and the long tail discloses in one, for another two years, is precisely a market where mispricing sits in the long tail. The opportunity and the blind spot are the same companies. That is not a coincidence; it is the mechanism.

Why machine translation does not close the gap

The obvious response is to push everything through a translation engine and read the output. It fails for two reasons, and neither is about translation quality.

A translated sentence is no longer the company's sentence. The moment a research claim is backed by a machine's paraphrase of a filing rather than the filing, the claim stops being verifiable — and an unverifiable claim about a listed company is worth less than no claim, because it carries the shape of evidence without the substance. Anyone who has watched a language model produce a fluent, confident, entirely invented citation knows the failure mode.

Translation also answers the second question, not the first. It tells you what a document says once you already know which of 1,500 documents to open. The expensive question comes before that: which companies are talking about this? Keyword search cannot answer it across a language barrier, because you must already know the phrase management chose. Twelve companies describing the same input-cost squeeze will describe it twelve ways, and no keyword hits all twelve.

How to read a Vietnamese filing without waiting for the translation

This is the workflow, and it is deliberately one you can run yourself today, with no product involved.

1. Find the governing document, not the summary. Start from the company's disclosure page or the venue's disclosure portal, and take the Vietnamese Báo cáo thường niên. If an English version exists, download both — you will need the pair in step 5.
2. Establish which document is the record. For anything you will cite in writing, the Vietnamese is the operative text under Điều 5. Note the file, the date and the page you took it from before you read a word.
3. Go to the Board of Management's discussion, not the financial statements. The statements tell you what happened; management's discussion tells you what they say caused it and what they intend to do about it. That is the part with judgement in it, and the part a ratio cannot reproduce.
4. Pull the same section from the three or four prior years. Put the paragraphs side by side. You are not reading for facts here — you are reading for changes in how the same subject is described.
5. Diff the pair, where a pair exists. Where you have both languages for the same year, compare length and completeness section by section. What is shorter in the English version tells you where to spend your reading time in the Vietnamese one.
6. Cross-check the narrative against dated disclosure. Insider transactions, major-shareholder changes and board resolutions carry dates. Lay them against the months management describes as unremarkable, and note where the two accounts do not sit comfortably together.
7. Record page numbers for every line you will quote. If you cannot point at the page, you cannot defend the claim later — and in this market, "later" means after the position is on.

That workflow is slow. It is roughly a day per company, done properly, which is exactly why it does not scale to fourteen hundred of them, and exactly why it is worth automating rather than skipping.

Five traps on the way there

The risk-management half of this, from someone whose job for years was validating other people's models rather than building flattering ones.

1. Citing the English text as the filing. It is a translation of the filing. Where they differ, it is the losing text. Cite it if you must, but label it.
2. Treating machine translation output as a quote. A paraphrase in quotation marks is the single easiest way to lose an argument with a company's IR team, and it will be the first thing they check.
3. Reading absence of English as absence of disclosure. A company with nothing in English may have filed everything required of it. The gap is in your reading, not in their compliance — and until 2028, that is true by design for much of the market.
4. Comparing across years in the shorter archive. See the rule above. If the English record starts in 2025, an English-only five-year comparison is not conservative, it is fabricated.
5. Assuming the reclassification changes the documents. Index inclusion changes who buys. It does not translate a single annual report, shorten a single roadmap date, or make a single UPCoM company disclose in English before 2028. The money arrives on 21 September 2026; the disclosure regime arrives on its own timetable, and the two are not the same event.

What we are building, and where it actually stands

This is the problem Boom Leverage is pointing its engine at. Here is the honest ledger.

What exists. The search engine has been in production since 2026 in another market, where it serves 916 companies as 479,663 verified finding rows drawn from management's own discussion in their annual filings. It searches by meaning rather than keyword — you type the thesis, it returns the companies discussing it even when their wording is nothing like yours — under one rule that shapes everything else: every result is the company's own sentence, linked to the source document, and anything that cannot be verified word-for-word against the filing is withheld and counted rather than blended in.

What we are building here. The same engine, reading the Vietnamese source rather than its English summary: the Báo cáo thường niên and the Board of Management's discussion inside it, plus insider-transaction and major-shareholder disclosures, across HOSE, HNX and UPCoM. Reading the governing document is not a stylistic preference in this market. It is the only way to search the years the English record does not cover, and the only way to quote a company in the text the rule says wins.

Where it stands today: no Vietnamese filings are indexed. Coverage is zero.

The line not to cross: we could have written that sentence differently. A page that lists a number it does not yet serve, or an article describing a scan of 1,500 companies in the present tense, would read better and convert better this month. It would also be the first thing a customer checks — and a research tool caught overstating its own coverage has destroyed the only asset it sells. The Vietnam page states the same zero next to the prices, before anyone leaves an email.

What we will publish, and what we will not

Since this is the first piece written for this market, it is a reasonable place to put the standing rules in writing, so they can be held against us later.

• Coverage is stated per year, per venue. Not "Vietnam" — the specific companies and the specific fiscal years searchable on the day you read it.
• Nothing renders unless it is quoted. A finding is the company's sentence with its source, or it does not appear.
• We do not paint companies with risk labels or rank them. The tool surfaces what management wrote and shows you where. What it means is the analyst's call, and it stays the analyst's call.
• Corrections are published, not quietly edited. This piece already carries one: an earlier draft put ad hoc English disclosure at 1 January 2026 for the whole market. It is 2026 for large-scale listed organisations and 2028 for the rest, and that difference matters enough to say out loud.

The short version

Vietnam's reclassification takes effect on 21 September 2026 and brings a great deal of new attention to a market whose primary documents most of that attention cannot read. English disclosure is required, phased, and — by the regulator's own article — not the version that governs. The English archive is roughly a year deep where the analysis that matters needs five. That distance is the Translation Gap, and it is where the next twelve months of mistakes will be made.

Two ways to go further with this:

• If you run a desk: email contact@boomleverage.com with the companies and fiscal years you follow. Early access is scoped with the desks that will use it, and named coverage requests move up the queue.
• If you would rather see the machinery first: the engine is live in another market — open the terminal and run a thesis through it before you decide whether the Vietnamese version is worth waiting for.

To be explicit: this is research and educational writing about public disclosure, not investment advice, not a recommendation to buy or sell any security, and not a forecast of any price. Every claim above is either linked to its source or stated as our own position. Where we describe our own coverage, the number is what we serve on the day of writing — including the zero.