Any screener can find a winner.
This one tries to kill it first.
Point-in-time universe with 2,040 delisted names still in it. A matched control group on every test. Transaction costs subtracted. Standard errors clustered by date. What survives all four is worth your time — and most things don't.
Free to run. Holdings and export are on Pro.
Three of these four numbers argue against us. They are here because a research tool that only ever agrees with you is not a research tool.
The four ways a backtest lies to you
1 · It only met the survivors
Test on today's index members and every one of them survived to today — that is why they are on the list. Here the universe is rebuilt as of the test date. Between 10% and 23% of positions in a historical run are companies that no longer exist.
2 · It had nothing to be compared with
"Up 1.78%, t = 4.55" means nothing until you know what comparable names did over the same window. Add the control group and that particular result disappeared completely. Every test here ships with one.
3 · It traded for free
Spread and commission are subtracted, and the result is shown as the subtraction it is: gross → net → annual cost → break-even. Two numbers far apart on a screen force you to do the arithmetic in your head, and nobody does.
4 · It counted one day as a hundred trades
Two hundred positions opened on the same morning are one bet on that morning, not two hundred independent observations. Errors are clustered by date, which is what makes the t-statistic mean what it claims to mean.
And then the AI reads the result
The measured base rates are not a separate report — they are handed to the assistant as part of the evidence pack. When a setup has a measured base rate, the AI is required to anchor its direction to it and quote the sample size and t-statistic. When it doesn't, it has to say so.
Tested on SHEL, it returned up / high confidence citing +1.75% at t = 13.14. Tested on AAPL, where no base rate exists, it returnedlow confidence and the sentence "no measured base rate exists". That second answer is the one that makes the first one worth having.
Also in the screener
- Every instrument with prices — 14,909 of them
- Fundamental, technical and flow filters in the same query
- A clickable equity curve: click a month, see the names behind it
- 14 event families, 1.7M occurrences
- Export the exact result set you filtered, on Pro
Questions people actually ask
What makes a backtest here different?
Four things, and each one exists because it changed a result we had already believed: the universe is rebuilt as it stood on the test date (delisted names included), every group is compared against a matched control group, transaction costs are subtracted, and standard errors are clustered by date so that one wild day cannot masquerade as a hundred independent observations.
Have your own signals survived that?
Mostly not, and we publish it. Volume shocks and range breakouts underperform their controls at t = −7.84. A +1.78% edge with t = 4.55 was wiped out entirely once the control group was added. Post-earnings drift did survive: +0.64% excess, t = 2.83. Of the presets we ship, one beats SPY net of costs.
Why publish results that make your own product look weak?
Because a research tool whose every test comes back positive is a broken research tool, and the people we want as customers know that. If we only showed the winners you would be right to assume the machinery does not work.
What horizon do the edges live at?
Between 21 and 63 trading days. At a one-day horizon only 2 of 25 measured families clear t = 3 — which is worth knowing before you build anything that trades daily on an event.
Are the signals absolute or relative?
Relative, and the difference is not academic. Event studies measure excess return against a benchmark, so a positive result says "this did better than comparable names", not "this went up". In absolute terms every group we measured returned less than the benchmark over the window. Treating a relative edge as an absolute forecast is how a good study becomes a bad position.
How many event families are measured?
There are 14 families and 1.7M recorded occurrences behind them. Each family also carries its own control group, which is why the count of families excludes the controls — counting them would double the number we claim.
The measured signals, when they fire
Which instruments had an event with a measured base rate behind it, and the number that backs it. No market calls, no newsletter filler.
Run a test that is allowed to fail.
No card, no demo call, no sales email. The terminal is public — an account raises the limits, and the AI runs on credits.
Free plan for good. Pro is $69.99/mo, or $4.99 for a 7-day trial.See what each plan includes →