My previous column described how I taught an artificial intelligence to read valuation reports. From this piece on, I turn to the findings.
In this profession, the question I am asked most is always the same one. How large should the discount for lack of marketability be? How large should the control premium be, before it counts as "reasonable"?
In the past, that question could only be answered from experience, from impression, from the dozen or so cases one happened to have seen. The more seasoned would cite the median of some American study. The more seasoned still would add, "it depends". Neither answer is wrong, but neither is an answer. Because in the Hong Kong market, there has never been a record long enough and complete enough for anyone to point to and say: over these ten years, this is what the market has actually done.
Once the ten years of data were compiled, the answer finally had a coordinate.
WHERE THE DATA COMES FROM
Let me set out the scale first. This study read through more than 4,000 Hong Kong "Notifiable Transaction" circulars. Of these, 704 contained a full business or equity valuation report, and 297 adopted at least one adjustment in the valuation conclusion, for a total of 477 adopted data points: 315 discounts for lack of marketability (DLOM), 146 control premiums, and 16 discounts for lack of control. Every figure is supported word for word in the original text, and verified against the arithmetic of the valuation itself, by the method I described in the previous piece.
Having said that, I must add one honest paragraph. These statistics were produced by an artificial intelligence, with my team and I performing only sample checks, and I cannot guarantee that every figure is one hundred per cent accurate. This is a limitation of using AI, and I will not dodge it. But two things can be said. First, every figure has a source in the original text and has been cross-checked several times, so even if there are individual omissions or misreadings, the statistical result will not be far from the truth. Second, were the same analysis done entirely by hand, reading several thousand documents each running to hundreds of pages, how could that guarantee one hundred per cent accuracy either? What matters has never been to claim to be error-free, but to lay the method, the sources and the limits out in the open, so that every figure can be verified and questioned. That is precisely the standard this study is willing to commit to.
One trade-off must be made clear first. This study counts only "adopted" figures. The percentages that appear in a valuation report far outnumber these 477. Some are sensitivity analyses, some are reference ranges, and some are merely citations of the literature. To count them all would double the number and distort the statistics. We kept only the one figure that truly walked into the valuation conclusion and reconciled on the value bridge. Better too few than too many, because once the coordinate is distorted, there is no gatekeeping to speak of.
While we are here, one interesting fact. Discounts and premiums are not mutually exclusive. A single circular can adopt both, first adding a premium and then applying a discount. Over the ten years, 119 circulars carried both.
DISCOUNTS: A NARROW NORMAL BAND, A STARTLING RANGE
Take discounts first. Across the 315 adopted DLOMs, the median is 18.1% and the mean is 19.9%. Half of the data falls between 15.7% and 20.9%. In other words, the market's "normal band" is in fact quite narrow, roughly one sixth to one fifth.
The single most crowded bucket is 15% to 20%: 125 of the 315, meaning two in every five. Only 16 sit below 10%, and only 14 above 40%. Both tails are rare animals.
But the range itself is worth pausing on. The lowest is 2%, the highest 56.8%, a 28-fold difference. The same term, two extremes. The former barely discounts the value of the asset, the latter more than halves it. This is the line I often repeat: the average stays the same year after year, the range is worlds apart.
PREMIUMS: A HIGHER CENTRE, A WIDER SPREAD
Now the premiums. A reminder first. A control premium is a premium, the opposite direction to a discount. A discount lowers value, a premium raises it, and on the valuation bridge the two, one added and one subtracted, each do their job.
Across the 146 adopted control premiums, the median is 24.8% and the mean is 23.5%, a centre clearly higher than the discounts. The main body falls between 20% and 30%: 89 of the 146. The full range runs from 5% to 50%.
To sum up the market's collective judgment over ten years in one line: "control" is worth about a quarter, and "cannot be sold at will" is worth about a fifth. Neither is a small number.
Let me put it through the simplest arithmetic. Suppose the reference value derived from listed comparable companies is HK$100 million. The buyer takes a controlling stake and adds a 25% premium, and it becomes HK$125 million. The target is a private company whose shares are hard to realise, so a further 18% discount is applied, leaving about HK$102.5 million. Between one addition and one subtraction, the difference runs into the tens of millions. And these two percentages often occupy just one line each, deep in the report. This is why I have always stressed that the two must be treated with "equal weight, recorded separately". They run in opposite directions, and their effect is equally large.
THE TEN-YEAR TREND: THE BIGGEST SURPRISE IS STABILITY
Laying the ten years out year by year, what surprised me most was that these two lines are almost flat.
The annual average DLOM has hovered around 18% to 22% throughout the decade. The annual average control premium has moved roughly between 20% and 27%, sitting at 23% to 25% in most years. We all remember what Hong Kong went through in these ten years: the market's sharp rises and falls, the revisions of the rules, the turnover of industries. Yet the centres of these two figures have barely stirred.
My original assumption was that the trend would follow market conditions, with discounts narrowing when the market is buoyant and widening when it is depressed. The data does not support that assumption. Whatever the year, the annual average holds stubbornly within the same band. It seems that when valuers select these two figures, what they refer to is not the temperature of the market at the time, but the same body of studies and the same set of conventions. That finding alone is worth the profession's reflection.
This gives us two messages. First, the market does have a consensus. A median that holds for ten years shows that valuers, as a group, in fact have a clear sense of the "normal level", which is a mark of professional maturity. Second, an individual case can sit far from the centre, and the average will hide it completely. Look only at the average and you would think all is orderly. Open up the distribution and you will see those 2% and 56.8% figures. What is stable is the group, what is unstable is the individual, and a transaction is always made one individual case at a time.
WHY THE GAP IS STILL SO WIDE
The same adjustment, how can it run from 2% to 56.8%?
Part of the reason is method. Some valuers cite American restricted stock studies, some use option pricing models, and some refer to Mainland M&A databases, and different methods naturally give different levels. Part of the reason is the case itself. The equity proportion, the industry, the size of the company, the difficulty of realisation, each has its influence. Reasonable differences ought to exist in the first place.
But "reasonable differences ought to exist" and "any number is acceptable" are two different things. The real question has never been whether departing from the median is wrong, but rather: is the reason for the departure written down? Which study is cited? Is it the median of that study, or deliberately its high or low end? Do the figures before and after the adjustment reconcile in the valuation conclusion?
With a coordinate, these questions can, for the first time, be asked precisely. Imagine a board receiving a valuation report that carries a 35% discount. In the past, most could only accept it as given, because there was nothing to compare against. Now one can point to the data and ask: the market median is 18.1%, half of all cases fall between 16% and 21%, so why does this one need to be nearly double?
If it can be answered, that is professional judgment, and it deserves respect. If it cannot, that is what gatekeeping is for.
For the directors and management of listed companies, I would suggest keeping the following questions to hand. First, which study or dataset does this discount or premium cite? Second, is it the median, or one end of the range, and if the latter, what is the reason? Third, is this percentage truly reflected in the final valuation conclusion, and does it reconcile? Fourth, placed against the ten-year market distribution, where does it fall? Four questions, requiring no valuation expertise at all, yet enough to expose a slipshod report.
This is exactly where the value of this data lies. It is not to replace professional judgment, but to place that judgment in the open, to be compared.
THE INSTRUMENT KEEPS RUNNING
Let me note the latest quarter in passing. In the second quarter of 2026, the quarter mentioned in the previous piece as the first to be read from end to end by an artificial intelligence, 22 circulars carried a business or equity valuation report, of which 14 adopted a discount and 6 a premium. The discounts ran from 8.7% to 42.9%, and the premiums from 18.1% to 31.4%, their centres almost exactly matching the ten-year normal band.
This is what it means to turn the research into an "instrument". Research of the past was a photograph, out of date the moment it was taken. The research now is a machine that keeps recording. Each quarter's new circulars are read and filed back into place, and the coordinate grows together with the market. The eleventh year, the twelfth year, will attach themselves automatically. For a subject whose very life is long-term comparison, there is no property more important than this.
(Part two of the series “Gatekeeping Valuation Quality in Hong Kong.” To be continued.)