Twenty-one rivals, one bar
The test bench
Every morning, next to the pick this site publishes, twenty-one other strategies each name one stock of their own. Ten read the same hundred-name shortlist as the product but decide differently — one deliberately takes the name the models refused, one takes the least extended of the names they approved, two read what the options market paid for each name at yesterday's close, and one sells the best-scored name short instead of buying it. Ten run their own screen across the whole market — two of them only on companies that reported earnings since the last close, one buying the beat and one selling the miss, and one selling the pre-market spike. One works top-down from sectors. None of them is shown to anyone before the public clock, and none of them is for sale. They exist to answer a question in public: is there a simpler, or different, way to choose that does better than ours?
Why this is not a leaderboard
The tempting design is to run thousands of variants and keep the winners. We replayed that idea on a hundred past sessions before building this page: the twenty best-looking weightings on the sessions used to choose them made about +0.2% a day; on the sessions held aside they made nothing that could be told from chance. A leaderboard crowns noise, and it crowns a new champion every quarter.
So this bench works the other way round. Each strategy's hypothesis and exact rule were written down on 2026-09-07, before its first session, and do not change. A strategy that needs fixing is a new strategy, under a new name, starting from zero. And the verdict is the same one the site applies to its own scoring weights: at least 60 sessions, and a mean return against the S&P 500 steady enough to reach t = 3.0. With 22 candidates, the luckiest one has about one chance in ten of touching 2.6 standard deviations by luck alone and just under three in a hundred of reaching three, so the bar still holds. Below it, the strategy is not retained — and that is published, right here, next to the ones that pass.
One more rule, decided before the first session: the verdict is read only at fixed checkpoints — after 60, 120, 250, 500 sessions — and the label is frozen between two readings. A verdict recomputed every evening would be a test that looks every day, and something that looks every day for a year will see a lucky streak cross the line sooner or later. The means and t-statistics in the table are live; the verdict is not, and it is written on each strategy's own page, not in the table. “Not retained” at a reading means exactly that: the strategy stays on the bench until the next one.
The rows below are in the order the strategies were registered, not by performance. Sorting them by return would be the leaderboard again, one click away. The published pick is counted here from the day the bench started, on exactly the sessions the other strategies faced — its full record since launch is on the track record page.
Sessions on the bench
27
S&P 500, same sessions
-0.06%
The shortlist, same sessions
-0.33%
Whole market, same sessions
-0.42%
The twenty-two strategies
| Strategy | Today | Sessions | Mean return | vs S&P 500 | t | Up days |
|---|---|---|---|---|---|---|
| The published pickthe product | — | 24(2 unpriced) | -0.19% | -0.15% | -0.25 | 46% |
| Unanimous jury | — | 23(3 unpriced) | -0.61% | -0.55% | -0.88 | 43% |
| Shortlist, numbers only | — | 24(2 unpriced) | -4.91% | -4.89% | -1.36 | 38% |
| Shortlist, news only | — | 17(+10 passed) | -1.03% | -0.94% | -1.28 | 29% |
| Gap continuation | — | 22(3 unpriced, +2 passed) | -0.47% | -0.39% | -0.78 | 36% |
| Pre-market leader | — | 22(4 unpriced) | -0.36% | -0.27% | -0.27 | 45% |
| Calm trend | — | 19(7 unpriced) | +0.03% | +0.06% | 0.58 | 42% |
| Oversold rebound | — | 24(1 unpriced, +1 passed) | -0.44% | -0.42% | -0.33 | 42% |
| Relative strength | — | 23(3 unpriced, +1 passed) | -1.37% | -1.30% | -0.98 | 35% |
| Sector rotation | — | 12(4 unpriced, +10 passed) | -1.18% | -1.16% | -1.08 | 17% |
| Fresh news, quiet price | — | 17(+10 passed) | +1.34% | +1.43% | 1.66 | 53% |
| Vetoed by the jury | — | 24(2 unpriced) | -3.75% | -3.73% | -1.01 | 42% |
| Pre-market fade | — | 25(1 unpriced) | -0.66% | -0.62% | -0.98 | 44% |
| Single strongest conviction | — | 22(4 unpriced) | -0.79% | -0.77% | -0.71 | 36% |
| First green day | — | 22(4 unpriced) | -0.34% | -0.28% | -0.43 | 45% |
| Jury direction, options amplitude registered 2026-09-22 | — | 11(4 unpriced, +1 passed) | -1.42% | -1.37% | -2.76 | 27% |
| Where the calls are dear registered 2026-09-22 | — | 11(4 unpriced, +1 passed) | -0.85% | -0.89% | -0.69 | 36% |
| Approved, least extended registered 2026-09-23 | — | 11(4 unpriced) | -0.89% | -0.83% | -0.56 | 45% |
| Reported since the close registered 2026-09-25 | — | 4(1 unpriced, +9 passed) | -3.82% | -3.55% | -2.01 | 0% |
| Sell the best-scored namesells short registered 2026-09-25 | — | 11(2 unpriced) | -5.01% | -5.04% | -0.73 | 45% |
| Sell the pre-market spikesells short registered 2026-09-25 | — | 10(1 unpriced, +3 passed) | +2.02% | +2.00% | 0.84 | 60% |
| Sell the earnings misssells short registered 2026-09-25 | — | 0(1 unpriced, +13 passed) | — | — | — | — |
What is being tested, and what is not
Nothing on this page costs an extra model call. The shortlist strategies reread the verdicts the three models already gave on the product's hundred finalists; the others read only prices and volumes, including the morning's pre-market trade on the liquid part of the market. Two strategies, registered on 22 September 2026, add one more source: yesterday's option chain on each finalist, read at the money by inverting Black-Scholes on the quoted prices. Not to predict — the formula gives every stock a hair under fifty-fifty on one session, by construction — but to read what the options market paid for: how far it expects the stock to move, and whether calls were dearer than puts. That is deliberate: the question is whether a different way of deciding beats the product, not whether ten times the budget does.
The three figures at the top are there because “beats the S&P 500” is not the right question. The mean of the product's own hundred finalists is what the shortlist alone is worth: a strategy that reads the shortlist and does not beat it has added nothing to the screen. The mean of every tradable stock is what the screen itself is worth: a strategy that does not beat it carries no information at all. Each row should be read against both, not only against the index.
Two things a hundred replayed sessions could not test are here precisely because of that. Nothing in our history records the news of past mornings, so a news-only strategy could only ever be tried live. And pre-market prices are served for sixty days at most, so the hypothesis that the moderate pre-market band carries information — the very largest gaps demonstrably lose — has to earn its verdict one morning at a time.
A strategy that passes the bar does not automatically become the product. It becomes an argument, with sixty measured sessions behind it, and the decision to change what this site publishes is then taken in the open on the learning page, as every change to the scoring has been.