Since 2023, every quarter I have done something that must look rather tedious from the outside: opening every Notifiable Transaction circular published by Hong Kong Exchanges, and — from valuation reports that often run to hundreds of pages — copying out two unremarkable-looking numbers that can nonetheless swing an entire deal’s valuation: the discount for lack of marketability (DLOM) and the control premium.
For those years, it was all done by hand. A single circular is as thick as a book, and the valuation report hides deep in an appendix. To be sure which discount the valuer actually adopted, and from which study, I often had to read back and forth several times. Three years in, I have developed something close to an instinct for these figures — and an equally clear sense that one pair of hands, one person, can only ever cover so much.
To my surprise, this rather niche piece of work has quietly gathered a readership. More and more valuation peers now cite my findings in their reports and discussions and I have even had accountants from the overseas Big Four email us for the detail behind a particular figure. That feedback has been enormously encouraging — and it convinced me that making this record longer, more complete and more reliable was worth real effort.
This quarter — the second of 2026 — marks a turning point. For the first time, the figures were “read” out by an artificial-intelligence system I built myself. I no longer transcribe line by line, but sit behind it, reviewing and gatekeeping. At the same time, I have pushed the study all the way back to 2016, assembling a full ten-year longitudinal record. As far as I know, tracking Hong Kong’s discounts and premiums this systematically, at this scale and over this span, is a first for the city.
And there is a larger hope behind all of it. Valuation is a matter of professional judgement and within reason, differences are allowed. But precisely for that reason, a mature market needs someone standing to one side — observing and comparing those judgements over the long run, independently and systematically — to keep watch over valuation quality. I hope this ten-year study, and the AI behind it, can become an independent yardstick for the profession and that my team and I are willing, and able, to play that gatekeeper’s role — not to replace the valuer, but to give the market one more independent point of reference and check.
Two numbers, buried deep in the report
First, why these two numbers.
When a company sets out to buy or sell a business, its board needs an independent valuation to support whether the price is “fair and reasonable.” In arriving at a value, the valuer often starts from the share prices of comparable listed companies, then makes two key adjustments: if the target is private and its shares cannot be sold at will, a discount for lack of marketability is applied or if the buyer is taking control and can steer the company’s decisions, a control premium is added.
These two numbers look like technical detail, but their impact is large. A fifty-per-cent discount can halve the value of a stake and a thirty-per-cent premium can lift a price markedly. They are, in other words, often the most elastic — and the most in need of gatekeeping — figures in the whole report. And yet they sit buried in its depths, scattered across sections, tables and footnotes, rarely laid side by side for comparison. That is exactly what set me off, three years ago.
Why teaching a machine to read is so hard
Surely, you might think, you just pull the numbers out of the file and hand them to a computer? Only in the doing do you learn it is not that simple.
Hong Kong’s valuation reports follow no standard format. The same discount for lack of marketability is written a dozen different ways by different firms: some spell it out in prose, some drop a bare percentage into a table cell, some write it as a formula, and some rename it altogether as a “discount for lack of liquidity.” Sometimes it appears on page ninety or sometimes you must turn past page one thousand. Trickier still, one document may carry several similar-looking numbers — some the valuer actually adopted, others merely sensitivity analysis, reference ranges, or citations from the literature — and it is easy to mistake one for another.
This, in short, is not a job that “find and replace” can finish. It is a job that must be read and understood. To judge which figure was truly adopted, the machine has to follow the valuer’s logic: how much of the equity is changing hands? Which valuation method was used? Did that discount actually flow through to the concluded value? These are precisely the things conventional programs are worst at — and precisely what today’s large language models are beginning to handle.
Not keyword search, but reading comprehension
Let me give a concrete example.
One of this quarter’s circulars involved the acquisition of a forty-per-cent minority stake. At first the system found no discount at all — but its own reasoning flagged the contradiction: if a minority stake is valued by the market approach, there ought to be a discount for its lack of marketability. “There should be one, yet none is found” — that signal sent me to open the report myself. The valuer had indeed applied a discount, only phrased in a rather artful way: the report read, “we consider 8.7% is sufficient to reflect the lack of marketability” of the forty-per-cent equity interest. In that sentence the number sits far from the word “DLOM,” with none of the usual “a discount of X%” trigger — and any system matching on keywords would sail straight past it.
That small episode captures the difference between reading comprehension and keyword search. What the machine must do is not to hunt mechanically for words, but — like an experienced analyst — to follow the valuer’s train of thought and judge each number’s role and destination. This is the hardest, most painstaking part of the whole project.
The three lines of defence that keep the machine honest
When people hear “artificial intelligence,” their first worry is usually: will it just make things up? A fair concern. Large language models sometimes state, with great confidence, things that are simply not there — and in research that lives or dies on accuracy, that is fatal. So I built the system three lines of defence.
First, everything must have a source. Every number the system reports must be traceable, word for word, to a sentence in the original. If that sentence cannot be found, the number is treated as non-existent — no paraphrasing, no invention. Second, several voices must check one another. Rather than trust a single model, I have several independent models read the same report and compare notes. A figure is adopted only when they agree — several modest minds outdoing one lone oracle. Third, the maths must verify. Any adopted discount or premium has to reconcile within the valuation’s own “value bridge”: from market value, through the adjustment, to the concluded value, the numbers must walk through step by step. Only once all three gates are passed does it fall to me, the human supervisor, to sign it off.
It is worth adding that the whole system runs on a single laptop of mine: the data never leaves the machine, never touches the cloud, and the run can be reproduced, untouched, every quarter. What was once a quarterly study done by hand, hard to replicate, has become an instrument that keeps running and can be audited at any time.
Why it’s worth doing: the value of a gatekeeper
Some may ask: is all this effort worth it?
My answer is yes. Valuation has always been a matter of professional judgement, and within a reasonable range it admits of difference but that is exactly why someone needs to record and compare those judgements over the long run, systematically, so the market can see a rough set of coordinates. Over ten years, where have Hong Kong’s discounts and premiums actually landed? Which are the rare, extreme cases? Do different firms have their own “habits”? Scattered impressions cannot answer these questions and only a record long and complete enough can.
This ten-year study covers more than 470 adopted discounts and premiums, close to 300 circulars, and nearly 200 valuation firms. Its purpose is not to “grade” any single valuation, but to hold up a mirror to a profession whose very job is to gatekeeping valuation quality. And the role of AI is to widen that mirror: from a few dozen reports a quarter by hand, to a decade and thousands of documents — something no pair of hands could ever do.
This is only the beginning
From hand to machine, this step took three years. This quarter is only the start.
In the pieces to come, I will share what the ten-year study has found: the levels and trends of discounts and premiums, the extreme cases that stay in the mind the intriguing differences between Hong Kong and mainland valuers and what all of it means for regulation and professional gatekeeping. And on 17 August, I will present — under the title “Teaching AI to Read Valuations, and Gatekeeping Valuation Quality in Hong Kong” — the full picture of this ten-year study for the first time.
Can a machine replace the valuer’s judgement? I do not think it can, and I do not think it should. But can it read, record and compare tens of thousands of documents for us, so that human judgement can be spent where it is most needed? This quarter’s results have already given an early answer. Stay tuned — more to follow.
(Part one of the series “Gatekeeping Valuation Quality in Hong Kong.” To be continued.)