Three AIs. One stock. Every trading day.
Today's pick is sealed until 9:00 AM New York
56:04:50
Today's pick goes up on this page at 9:00 AM New York time, right after the 7:30 scan.
Before the opening bell, GPT, Claude and Perplexity score the US market and agree on one name. It goes out by free email and lands on this page at 9:00 AM New York time — thirty minutes before the market opens. Every past call stays online, and it costs nothing.
Does it actually work?
The honest answer is that the track record is too young to say. Below is the whole of it, and under that the measurement where we already know we come off badly — scored against two yardsticks, one of which beats us.
The track record so far
Every pick is measured open to close against the S&P 500 on the same session, low-conviction calls included, and no published call is ever removed. With 43 sessions on the board, treat all of it as noise: a run of good days is exactly what chance produces at this sample size.
- Sessions measured
- 43
- Closed up
- 47%
- Beat the S&P 500
- 44%
- Average session
- -0.25%
The part we already know is weak
This next measure does not grade the pick. It grades the screen that draws up the shortlist every morning — a hundred names, out of which the models then choose one. We publish it because it is the place where we currently come off worst.
137/820
of the biggest tradable gainers were already on our shortlist, over 41 sessions where all three counts can be compared
39.4/820
what a random shortlist of the same size would have caught over those same sessions
248/820
what simply shortlisting the most volatile stocks would have caught, same sessions and same basket size
Every evening we measure the day's move for every stock we track — 240,735 stock-days so far — then check how many of the session's twenty biggest relative gainers had made it onto our hundred-name shortlist that morning. Twenty chances per session. Beating pure luck is the easy test, and we clear it on the sessions graded so far.
The hard test is the one that matters, and it is the one we would run on somebody else. Big movers cluster: a stock that has been jumping around keeps jumping around, so any screen that leans on momentum and volume catches more of them than chance without predicting anything. Against a basket of the most volatile names from the preceding sessions — same size, same price and liquidity floors, no models involved — we catch 137 where it catches 248. On this measure our screening is currently behind a rule you could write in one line.
We publish it because it is true and because it points somewhere: it suggests our screen filters too hard before the models ever see a name. Note what it cannot tell you either way — neither yardstick shows direction, and finding the stocks that move is not the same as knowing which way they go. The grading page carries that number too, negative as it currently is.
Step 01
The whole market, every morning
Every listed US stock is scanned at 7:30 AM New York time, then narrowed to about a hundred finalists on volume, price action and fresh news — and those hundred are re-scored on the morning's actual pre-market trading.
Step 02
Three AIs vote, equally
Perplexity digs up the latest news, then GPT, Claude and Perplexity each score the finalists. No single model decides.
Step 03
Graded the same evening
After the close we measure the pick, the ninety-nine finalists we passed over, and every other stock we track. The record is public and permanent.
The evening shift
Every night, we ask what we should have done differently
Grading yourself only on the stock you recommended is how a blind spot survives for years: a filter that throws away a whole family of winners leaves no trace of it in what it kept. So after every close we record the move of every stock we track — not just the pick, not just the shortlist — and mark the morning's ranking against what the market actually did. One call a day becomes thousands of measured outcomes, and the questions worth asking — what did we miss, which model was right, where did the ranking point the wrong way — get answered with data instead of instinct.
That record is also what sets the scoring weights. They are re-examined after every close, against the whole accumulated history rather than the last good day, and they move only when a change survives being tested on sessions it was not chosen from. On a handful of sessions nothing can survive that test, so the answer is no and the weights stay where they are — which is the correct answer, and the one published each evening alongside the reasoning. The ledger meanwhile builds, in public, session by session, and unlike a track record it gets more informative whether the picks are right or wrong.
See the graded sessions