We built a screener for low-float stocks gapping up on heavy volume, wrote down the edge we believed in before we could see whether it worked, and ran it forward for seven weeks. It didn't just fail. It failed in the opposite direction to our prediction — and the early result that made us optimistic turned out to be noise.
128 picks across seven sessions were logged after the market had already opened, because our scheduler drifted past the bell. Five of those sessions fall inside this experiment's window — about 96 graded picks here. They are now excluded everywhere.
The corrected in-window mean moves from −2.64% to −3.65%. The verdict below does not change — the failure is slightly larger than reported, not smaller. We could not show that the late data flattered us either: the entire mean difference is one trade. The full correction and post-mortem →
309 post-registration trades · −3.0 percentage points per trade versus the honest alternative · 95% confidence interval [−4.4, −1.5] · holds after correcting for the fact that we were repeatedly betting on the same 16 names.
The screen finds stocks that are small, thinly floated, and jumping on unusual volume — the kind of setup an entire industry of alert services and courses is built around. Our own record showed the picks tended to spike early and fade later, so we formed a specific belief: the edge isn't in picking these, it's in getting out of them.
So on 23 June 2026 we froze a rule. Buy at the opening price. Rest a limit order 10% above it for a five-day hold. If it fills, you're out with roughly +8% after costs. If it never fills, you exit at the fifth day's close. The bar it had to clear was simple and set in advance: beat just selling at the end of the first day.
Every backtest we had said this was good. In the sample we'd already seen, it produced a 63% win rate and a +8% median return. That is exactly the kind of number that gets put on a sales page.
Publishing a strategy after you know it worked is not evidence of anything, so the whole project is built to make that impossible.
The rule went into a public file with a date on it. Only trades that happened after that date could ever count. Every pick is written to an append-only log before the market opens, with its price, and the file is sealed into a hash chain so that any later edit to a past entry would be provable. (That timing is now enforced by a hard gate. For seven sessions it was not, and we published the correction rather than quietly restating the numbers.) The daily price paths used for grading are captured once, at grading time, and never re-fetched. And the whole thing runs itself on a schedule — nobody decides which days to include.
If you can only check a record when it's flattering, it isn't a record.
This is the part worth sitting with. Here is the same number, measured three times as the sample grew:
| Trades graded | Edge vs. the simple alternative | What we'd have concluded |
|---|---|---|
| 30 | +1.7 pp | "It's working." |
| 200 | −2.1 pp | "…that's concerning." |
| 309 | −3.0 pp | The rule loses, significantly. |
At 30 trades the strategy looked like it worked. Had we been selling something, that is exactly the moment the marketing would have gone out — a positive number, a plausible story, and a small sample nobody would have asked about. The sign then inverted and stabilised negative as the evidence accumulated.
That is not bad luck. It is what a coin flip looks like when you stop flipping too early, and it is the single most common way retail strategies get sold honestly by people who genuinely believe in them.
Final measurement on 309 post-registration trades: the rule averaged −5.9% per trade against a baseline of −2.9% — a gap of −3.0 percentage points, with a 95% confidence interval of [−4.4, −1.5] that excludes zero. It doesn't merely fail to beat selling on day one; it loses to it, reliably.
We also checked it the harder way. Because the screen kept picking the same 16 names, 309 trades are nowhere near 309 independent pieces of evidence. Re-running the statistics treating each company as the unit rather than each trade, the interval is [−6.0, −0.6] — still negative, still excluding zero. Only 5 of the 16 names favoured the rule at all.
And the four selection filters we'd registered alongside it — skip the cheapest names, skip the biggest floats, skip the most extended gaps, skip the hottest tiers — all came back flat, every confidence interval straddling zero.
Across the eleven rules we had running, six produced a positive-looking point estimate. If none of them worked at all, you would expect about 5.5 to look positive by pure chance. Two results cleared statistical significance — and both of them were negative.
That is what "no effect" looks like from the inside. Any project running eleven rules and reporting only the best-looking one could have shown you a winner here.
While auditing this result we found that one of our analysis scripts was reporting the baseline exit as +8.0% when the true figure is −2.9%. The cause was mundane and instructive: for a minority of older picks the script re-downloaded historical prices, and those come back adjusted for stock splits while the entry price we'd stored was not. One reverse split injected a four-figure return and dragged the averages up.
It only inflated the rules whose exit price came from a market price — which is why it looked plausible rather than obviously broken. We removed the re-download entirely and added a check that every trade must reproduce its stored return before it's allowed into the study. After the fix, not a single one of the 23 exit rules we'd tested beats simply selling at the first day's close.
We're telling you this because a track record that has never published a correction is not a careful record — it's an unaudited one.
This one generalises well beyond us, so we wrote it up on its own: the full write-up of the split-adjustment trap — why the error is always positive in cheap stocks, why it survives review, and the checks that catch it.
Having established that the score predicts nothing about direction, we asked what it does predict. The answer is magnitude: it reliably ranks how violently a stock is about to move, while carrying essentially no information about which way. Stocks in the top fifth of our score have roughly a 48% chance of drawing down 20% or more within five days, against 20% for the bottom fifth.
That relationship replicates on a completely separate universe of stocks, survives correcting for repeated names, survives a multiple-testing correction across 26 tests, gets stronger in the most recent data, and isn't explained by share price. It is also — and we want to be precise here — a rediscovery of volatility clustering, one of the best-documented regularities in markets, not something new. It is not an edge and it is not tradeable on its own. Knowing a stock will move a lot tells you nothing about which direction to bet.
It is now registered as a forward test in its own right, so it has to prove itself the same way everything else here does.
Across 524 graded picks: a 30% win rate, about −2.9% per trade on a same-day hold and −8% to −10% over five days, after a modest cost assumption. No exit rule fixed it. No filter fixed it. Delaying the entry didn't fix it either — we measured the drift entering at every point across the following week and it is negative everywhere, around −0.9% per session.
The raw drift is under 1% per session in either direction. A realistic round-trip cost on stocks this thin is larger than that. The signal was smaller than the friction — which is why nothing worked long or short.
One rule we tested — take profits at +5% — produced a 67% win rate and an expectancy of −4.4% per trade. Tight targets buy you a pleasant-looking win rate by capping the few trades that would have paid for all the others. If a strategy's win rate rises while its expectancy falls, you are being sold the wrong number.
Every apparently profitable variation we found in this data turned out to be one stock that ran +663%. Remove it and each one collapses to a loss. This is the cheapest, most useful check you can run on anybody's backtest, including your own.
None of the above requires trusting us. The complete pick log and every graded outcome are downloadable as raw CSVs from the track record. The frozen rules with their registration dates are in HYPOTHESES.md. The full verdict, including the bug write-up, is in the audit log. The method explains how grading works and where it falls short.
If you recompute any of these numbers and get something different, we want to know.
It says this screen, on these stocks, over this seven-week window, in this market regime, lost money and could not be rescued by any exit or filter we registered. It does not prove that no edge exists in this class of stock, and we haven't tested anyone else's version of it. That's what the rest of the experiments are for.
We turn the same machine on somebody else's strategy — one of the most widely published retail setups of the last twenty years, and one that is almost always sold on its win rate. Registered, running, verdict published either way.
A verdict lands every few weeks. We'll email you each one, free, the day it publishes — and nothing else. No signals, no picks, no offers. Verdicts stay free and public for everyone either way; this just means you don't have to remember to check back.
One confirmation email first — you're not on the list until you click it. Privacy
Educational and informational only — not investment advice, not a recommendation, and not a broker-dealer. ThePickLog is operated by AMD Ventures, LLC (Florida). Past results, including ours, do not predict future results. Every figure on this page is recomputable from the public files linked above.
Disclaimer · Privacy · All experiments