David Webb spent three decades building Hong Kong's most complete market database — every price, every director filing, every buyback, every custody record. Before his death he released all of it. We rebuilt it, tested seven families of trading signals against it, and then set adversarial reviewers loose on every claim. This is what survived.
Only one family of positive signals survived verification — and it isn't insiders. It's the company's own wallet. Share buybacks, read the right way, are the strongest and most tradable thing in the dataset.
Companies buying back small, token amounts of their own stock outperform peers by about +27% a year before costs — roughly +18% after. Heavy, sustained buying earns nothing: that's price support in a falling stock. The information is in the gesture, not the volume. The effect is strongest in genuinely liquid names (~HK$18M/day), which makes it the most executable finding here.
When a company does its first buyback after 12+ quiet months, the stock beats peers by +33% annualized over the following month — at a strictly realizable next-day entry, after the skeptics stripped out the un-capturable announcement gap. Insensitive to how you define the gap, positive in every one of 22 years. Thin (~8 names a day), so it's an entry trigger, not a book.
Most of the dataset's alpha is defensive: Hong Kong small-caps destroy value in recognisable, machine-readable ways, months before the damage lands. Borrow is scarce here, so these work best as never-own screens — with a small short book in the liquid names.
Companies that raised capital three or more times in three years lag peers by −19% a year — and by −34% among liquid names, where the effect even survives a cap-weighted benchmark. The treadmill is habitual: a 3+ raiser has a 51% chance of raising again within a year (vs 10% for non-raisers). Webb's core thesis, confirmed. The single most robust avoid signal in the data.
When a director pledges shares to an unqualified lender, the stock lags by −16%/yr for the next year and has a 49% chance of a −40% drawdown (matched peers: 30%). When the lender enforces the pledge, it's over: 77% crash within twelve months, average return −50%. If you own it, this is the exit bell.
When a stock's shares pile up fastest into its top-10 custody accounts over three months, it lags by −8 to −10%/yr — and unusually, the effect is strongest in liquid names (−22%/yr). Big ownership reshuffles of any kind are bad news; stability is the bullish signal. This is the machinery behind Hong Kong's pump-and-collapse microcaps.
Two or more board resignations within 63 days precede roughly −10%/yr of underperformance for six months and double the odds of a −50% collapse. These companies bleed rather than die — delisting risk isn't elevated, returns just rot. Holds in 20 of 22 years.
Companies that share a director with a stock that just crashed −60% underperform for the next six months. After the harshest controls, the pure contagion effect is −3.8%/yr overall — but −8.9%/yr among the largest names, at HK$10M/day turnover, which makes it the dataset's one genuinely shortable governance signal. Webb mapped these director webs by hand in 2017; the data says he was right.
Stocks whose top-10 custody accounts hold 45–85% of all shares sit in a hazard zone: twice the crash probability and −5 to −6%/yr of drag. Above 85% it's just big caps at custodians — safe. The naive "concentration = danger" rule fails; the shape is a hump, and the middle of it is where blow-ups live. Best used as an index-exclusion screen.
Honest nulls are worth as much as the positives — these four ideas are widely believed, fully testable in this data, and dead. Knowing them saves real money.
The classic US result does not transfer. Across 41,000 open-market director purchases, the edge is zero. Several directors buying together? Still zero. Buying the dip after a −30% fall? They do worse — falling knives, caught. In Hong Kong the information is in the sells and in buybacks, never the buys.
t ≈ 0 · 41,410 eventsSFC weekly short positions — levels, changes, squeezes — predict nothing. The most liquid, best-powered test in the whole study, and the cleanest null. Week-old public data is already in the price.
t = 0.2 · 693 weekly reportsCompanies whose directors hold 6+ board seats do underperform — but only because they're small. Control for size and the effect vanishes entirely. Webb's own "Project Vampire" doesn't survive its own database.
size artifact · t = −0.9 controlledRights issues priced below half the market price look like death spirals — but all the damage is on announcement day. Afterwards, returns are a skewed lottery: a few multi-baggers, no reliable edge either way.
post-announcement t = −0.8In Hong Kong's small and mid caps, alpha is mostly about avoiding governance failure — and the failures announce themselves in filings months early. Companies quietly returning cash are telling the truth. Companies serially raising it, pledging it, reshuffling it between custody accounts, or losing directors in pairs are also telling the truth. The market prices neither fast enough.
Starting a buyback is information. Sustaining a huge one is desperation.
Raisers raise again — 51% within a year. The behaviour, not the event, is the signal.
A director borrowing against his own shares has told you exactly how this ends.
Big moves in who owns or governs a stock — any direction — precede losses.
Every claim above went through a two-stage gauntlet: an independent researcher agent produced it, then a separate adversarial agent tried to kill it. All 20 headline computations reproduced exactly; the effects shown are the corrected numbers after attack, not the flattering headlines.
Webb's 30 GB MySQL archive stream-parsed into an analytical database; the price series verified to be corporate-action-adjusted total returns (splits, rights, dividends — checked against Tencent and HSBC events to the cent).
Coverage, freshness, benchmark sanity, and join integrity checked before any signal work. Usable return history: Feb 2004 – Dec 2025.
Seven independent agents, one signal family each — insider dealings, buybacks, custody flows, concentration, short interest, dilution, director networks — under pre-registered hypotheses and mandatory lags.
Each finding handed to a skeptic with orders to refute: re-run the code, hunt for lookahead, split the sample, drop the top contributors, price the trading costs, discount for multiple testing. Several headlines were cut in half. Two died into nulls. What's left is what held.
t measures how unlikely a result is to be luck — roughly, the effect measured in standard errors. Above ±2 is conventionally significant; above ±5, luck is a non-explanation. The weakest surviving signal here is t = −2.8; the strongest, t = 8.3.