What the data says
Things we have measured on SuperAlerts and the pre-market runners, including the ideas that failed. Each one shows its sample size and whether it held up on data the idea was not built from.
What the data shows
Of overnight gaps go down
Across 936 SuperAlerts from 2022 to 2026, the next morning's gap was down about 7 times in 10. After-hours moves were down about 64% of the time.
We banded the gap against close-in-range, day move, day range, float, rotation and the AI Score. Every band sat on the base rate. A stock that closes on its highs gaps up 35% of the time, and one that closes on its lows gaps up 32%. That difference is noise.
Holding into the close hoping for a gap up is a bet that loses about 7 times in 10.
09:25 beats 19:55
We compared two fixed entries on the same alerts: the 19:55 after-hours price on the alert day, and the 09:25 pre-market price the next morning. Both were judged on the same five sessions.
The share that reached +20% was 46% for after-hours and 50% for pre-market in 2025 (n = 276), and 47% and 51% in 2026 (n = 341). The pre-market price was cheaper on about two rows in three.
This fits the overnight fade: the pre-market buyer usually buys after the drop, not before it.
Get back to breakeven next session
Of entries that were red at the entry-day close, 50% (2025) and 51% (2026) traded back to the entry price within one session. Within 20 sessions it was 69% and 72%.
The ones that never came back fell hard: the median is now −94% (2025) and −84% (2026). Recoveries mostly happen fast or not at all.
This describes what happened. It is not a rule to hold or sell.
The early dip on after-hours runners
On after-hours runners with an Overall score of 3 or more, a dip formed between 04:01 and 04:10 on 14 of 16. Buying that dip reached +20% within five sessions on 14 of 14.
This is in-sample and small. The rule came from one of the charts in the set, and the earliest pre-market prints are the thinnest of the day, so real fills could be worse. We are collecting new days before trusting it.
Two categories beat none
Each SuperAlert can carry categories such as Extreme Range (a day range of 200% or more), All Day Bid and Afternoon Ignition. They are defined from alert-day facts only, and they were named before they were measured.
Alerts with two or more categories reached +20% within five sessions 63% of the time, against 37% for alerts with none. The gap held in both 2025 and 2026. Three categories did no better than two.
Five sessions is the right window
The median alert reached 98% of its 20-session peak by session 5. Meanwhile the median drawdown deepened from −22.7% to −36.1% over the extra fifteen sessions.
That is why every SuperAlerts result on this site is measured over five sessions.
Tested and failed
Good news does not pick winners
We scored every headline from 0 to 10 (a definitive deal scores high, an offering scores low) without seeing any outcomes, froze the scorer, then tested it.
Its AUC was 0.467 on 2025 and 0.466 on the held-out 2026 alerts. 0.5 is a coin flip. The strongest news did slightly worse, if anything.
Our basing buy alert lost to random
Replayed with no look-ahead over 172 alerts, the trigger reached +10% within six hours 11.0% of the time. Random entries on the same names in the same window reached it 18.2% of the time.
We publish failures like this because they are most of the work, and because they stop us selling something that does not work.
Buying the capitulation flush
A detector built from a real capitulation chart was tested on 699 alerts from 2022 to 2025. It did 10.6 points worse than buying at 10:30, and it failed in all four years.