185 · Testing an outside strategy (hedge-fund friend's 5-criterion combo)
Date: 2026-07-17
We get asked, often, why the live universe is European and not American. This is the honest long-form answer. A hedge-fund friend proposed a concrete recipe built for the US market, we backtested it point-in-time on our own data, and it failed. We publish the losing windows, not just the winning ones.
The thesis under test
The proposal, verbatim in spirit: "insiders sell for a thousand reasons but only buy when they think the stock is undervalued; the alpha lives in illiquid micro-caps where institutional algos will not go." Five criteria:
- Small caps under 500M USD (smaller is better).
- Materiality: buy over 1M USD AND raising the insider's own position by over 10%.
- Info asymmetry: biotech plus gold/mining ("the best sectors").
- Cannibal trait: a company shrinking its net share count 2-3% a year (buybacks).
- After-market: enter early the next morning; free-market buys only, ignoring option exercises and tax-driven transactions.
What we could and could not test (stated up front)
We do not overclaim. Of the five criteria, only three are cleanly testable with the data model we have:
| Criterion | Testable | Why |
|---|---|---|
| Small cap under 500M | Partial | Market cap is a current snapshot, not point-in-time. Using it is a survivorship-optimistic proxy: names that crashed into the bucket after the trade get swept in, and companies that delisted to zero are gone. The real result is worse than shown. |
| Amount over 1M | Yes | Filing amount, currency-normalised. |
| Position over +10% | No | We do not store insider holdings history. |
| Biotech / health | Yes | Sector tag on the company. |
| Gold / mining | No | We currently tag zero gold/mining companies. Cannot test. |
| Cannibal (buybacks) | No | No share-count history to measure a 2-3% annual shrink. |
| After-market intraday | No | Daily end-of-day data only. Our standard next-session entry already is the friend's "enter next morning". |
| Free-market only | Yes | We already exclude grants, option exercises, tax withholding, gifts and derivatives. |
So the combo we actually tested is: small-cap AND material AND biotech AND free-market, run separately per region. Two of the friend's "best sectors" and two of five criteria are structurally untestable, and we say so rather than pretend.
How the test was run
Point-in-time construction (no look-ahead): each month we build an equal-weight book of every name matching the screen, hold for the standard window, and net out a realistic round-trip cost. Returns are winsorised to tame single-name blow-ups. We report an annualised Sharpe, a bootstrap 95% confidence interval, a maximum drawdown, and a deflated Sharpe that penalises for the number of variants tried (the Bailey and Lopez de Prado correction). Nothing here is cherry-picked to the best window.
The finding worth publishing: the biotech tilt is regionally inverted
Cut the raw insider-buy pool into health versus non-health, per region. Annualised Sharpe, whole 2022 to 2026Q1 window:
| Pool | US | Europe |
|---|---|---|
| Health (biotech) | -0.53 (drawdown -86%) | +0.14 (drawdown -73%) |
| Non-health | +0.23 | -0.88 |
The biotech tilt does opposite things on the two tapes. In the US it destroys the signal; in Europe it helps on a relative basis. This is the same US signal inversion we have documented for years, now showing up in the sector dimension too. It is scientifically clean and it is one more reason we do not trade the US tape.
It is NOT, to be equally honest, a tradeable European edge on its own: the European biotech number has a confidence interval that straddles zero, a negative deflated Sharpe, a -73% drawdown from sector concentration, and it leans on a single sub-window. It is a relative tilt inside a losing pool, not standalone alpha. We are not selling it as one.
Stacking the friend's filters (the full combo)
Adding each criterion to the US pool makes it worse, not better:
| US step | Sharpe (2022 to 2026Q1) |
|---|---|
| Free-market pool | 0.00 |
| plus biotech | -0.53 |
| plus small-cap under 500M | -1.30 (drawdown -98%, near-total ruin) |
| plus materiality over 1M (full combo) | negative mean return, only 38 trades in 4 years |
The US small-cap biotech intersection is the single worst cell in the entire dataset, and remember it is already the survivorship-optimistic bound, so reality is worse. In Europe the same filters rescue a losing pool to roughly flat and the full combo turns nominally positive, but on 26 trades across 17 active months with a confidence interval spanning zero. That is noise, not an edge.
Verdict: fail
The friend's strategy does not survive a point-in-time test and does not beat our live European selection on any window. Its core configuration, US micro-cap biotech open-market buys, is the worst-performing intersection we have ever measured. His premise ("illiquid micro-cap is where the algos are not, so that is where the alpha is") is backwards in the data: those names are where retail-followed insider buys bleed the most, through illiquidity, biotech binary risk, dilution and delistings.
The one genuinely new result, the biotech regional inversion, is interesting and consistent with everything else we have found about the US tape. But it is not a shippable edge, and we do not dress it up as one.
This is the fourth outside "conviction filter" we have tested and retired (alongside a small-cap tilt, a conviction/position-size overlay, and a fundamental-quality screen). Stacking intuitive filters on the insider signal does not manufacture alpha; it concentrates the book into small, high-variance, negative-drift corners.
Caveats
- The small-cap and sector filters use current market cap and sector tags, a known survivorship and look-ahead bias that flatters the small-cap results. The true US small-cap biotech number is worse than the one shown, which only strengthens the fail.
- Winsorisation caps the worst blow-ups, so the raw US combo understates the tail damage.
- The European combo's nominally positive mean rests on 26 trades. Do not read it as a signal. We do not.
Why this is on the public site
Because the honest answer to "why not the US?" is not a slogan, it is a backtest. The live proof for the strategy we actually run is on the performance page. This note is the counter-example: a plausible, well-argued, US-built strategy that we tested fairly and that failed. Publishing it is the point.