Why a 22-Hour AI Trading Win Proves Nothing

Updated 2026-09-30

A popular video says an AI doubled a $100 trading account in 22 hours. It's a fun watch. It also proves almost nothing, and the reasons are simple math.

We know because we ran the slow version. AI research agents logged more than 6,000 tests of trading ideas on ten years of US stock data. Two ideas survived. Neither looks anything like doubling an account in a day.

The break-even bar: 55.6% before fees

Many of these videos use short up-or-down bets. You guess whether a price goes up or down over a short window. A right guess pays back 1.8 times your stake. A wrong guess loses it.

That payout sets a bar. To break even you have to be right 1 ÷ 1.8 = 55.6% of the time. Add a small fee, say 2.5 cents on every dollar you bet, and the bar rises to about 57%.

So "it won more than it lost" isn't enough. At 1.8x it has to win well over half the time just to stand still.

Try 34,000 rulebooks, keep the best five, and luck wins

A common shortcut: generate tens of thousands of variations of a trading rule, test them all on past prices, and keep the top few.

Here's the problem. Picture 34,000 rulebooks with no skill at all. Each one is a coin flip. Give each one 100 bets. The best five will win about 68 to 70 of their 100 bets. Every one of them is pure luck.

Holding back a third of the data as a second check helps, but less than people think when the winners were picked from thousands. Test those five lucky coins on 50 hidden bets each and about two will still come out ahead. About one will even clear the 55.6% bar. Nothing but chance. And each time you peek at the hidden data and go back to tweak, it stops being hidden.

Bet size: the part that decides whether you survive

Now suppose the edge is real. You are right 57% of the time on a bet that pays 1.8x (ignore fees for a moment). How much should each bet be?

The classic answer is the Kelly formula: bet size = p − q ÷ b, where p is your win rate, q is your loss rate and b is what you win per dollar. Here that's 0.57 − 0.43 ÷ 0.8 ≈ 3.25% of the account per bet.

Bet 40% of the account instead, about 12 times that, and the swings take over:

• Each win multiplies the account by 1.32. Each loss multiplies it by 0.6.
• Six wins in a row turn $100 into about $530.
• Five losses in a row leave less than $8.
• Over many bets the typical account shrinks, even with the real 57% edge, because each loss takes away more than each win adds.

With fees included, 57% is roughly break-even, so the right bet size there is close to zero.

Ten bets in a day is noise

A pure coin-flipper wins 7 or more of 10 bets about 17% of the time. Someone with a real 57% edge still ends about 44% of 10-bet days flat or down. One day can't tell you which of the two you're watching.

Real edges need far more than a day. One of the patterns below needs about 570 trading nights, more than two years, before its record alone can be told apart from luck.

What we did instead: 6,000+ tests, 12 finalists, 2 survivors

Our Edge Lab is the slow, boring version of the same question. AI research agents tested trading ideas on US stocks from 2016 to 2026, and every test was logged, failures included.

1. More than 6,000 logged tests: calendar patterns, trading volume, sectors, options activity, company reports, insider buying, short interest, overnight moves and more.
2. The years 2023–2026 locked away. Ideas were built and judged on 2016–2022 only.
3. 12 finalists, each tested on the locked years exactly once.
4. A correction for having 12 finalists. The more ideas you test, the higher the bar each one has to clear.
5. Re-priced at the official auction prices these trades actually get at the open and the close. This check could only lower a result, never raise it.
6. 2 survived.

The two survivors, as backtests (simulated, no real money, after an allowance for trading costs; see our risk disclosure):

• Cash-backed earnings at report time. When a company files its quarterly or annual report, buy the firms whose earnings are most backed by cash actually coming in, and bet against those whose earnings are least backed by cash. Close both four trading days later. On the locked years: about +0.49% per trade, and it passed the correction (p ≈ 0.04).
• An overnight model. Each afternoon a statistical model picks stocks to buy, and others to bet against, for one night: in at the closing auction, out at the next morning's opening auction. About +0.09% per night on the locked years. It barely passed the correction (p ≈ 0.05).

What didn't make it matters as much. A simpler overnight rule, buy last year's strongest stocks at the close and sell at the next open, looked fine on ordinary daily price data. Re-priced at the real auction prices, it fell below our pass mark. In one case a +15% overnight "win" on the daily data was 0.0% at the official opening auction. A third finalist, buying after company insiders report big purchases, held up on its own but not after the correction. The other eight failed outright.

Even the survivors are small, slow edges: a fraction of a percent per trade. None of it is proven for the future; the locked years are one test, not a promise. The Edge Bot now trades these patterns on a paper (practice money) account, scored at the official auction prices. Its page says how strong the evidence is for each one, including the simpler overnight rule, whose evidence is now the weakest.

Where to watch AI-written rulebooks trade, in public

On this site about 20 trading bots trade $100,000 paper (practice money) accounts in public, on real market prices. Most follow rulebooks written by AI models. Each is compared with the S&P 500 over its own dates, and every closed trade is listed.

The leader since start is Patience · Grok, a bot that follows a rulebook Grok writes. As of the September 29, 2026 close it was up +43.9% on paper since July 28, while the S&P 500 did +3.4% over the same dates. That is before trading costs; see the risk disclosure.

We don't trust that number either. By August 17, three weeks in, it was already up about 43%. In the six weeks since, it has barely moved. See how it did it, trade by trade. A good three weeks can be luck. The only fair test is a long record, next to the market over the same dates, with every trade shown.

Follow a bot — one email a day

Paper trading (practice money, real prices). Not financial advice. Risk.

See the live paper-trading scoreboard — free — stocks

These are paper trades — simulated money, real market prices — published as a record of what happened, not as advice and not as a prediction. Nothing here is a recommendation or a forecast, and no figure on this page describes money anyone earned or could have earned.