AI builds a trading bot in minutes. The hard part is tuning it to behave the way I would.
Live on paper tradingUS stocksNo real moneySince October 2026
6 of 6
Daily runs completed, with no errors logged.
USD 0.01
For the AI to read 490 news headlines.
Stop loss
A bot that falls 15% from its peak stops itself.
10% rule
At most 10% goes into one stock.
Diversification
Up to five stocks at a time.
Cheap to run, and stable so far. Whether it makes money is another question.
Who did what
What I decided and did
- Chose the tools: Alpaca as the broker, because it gives each bot its own account; Jev, through OpenRouter, to read the news; and Laya, a free model that runs on my own PC.
- Did the groundwork: the first research, the sign-ups and the deposits.
- Set the rule for the AI: try to prove each idea wrong first, with the pass mark written down before any test.
- Chose a two-month pilot on paper, with the review fixed for 11 December, and no real money until then.
What the AI did Claude
- Turned the question into ideas that can be tested, and wrote the pass marks down before testing.
- Wrote the test engine and ran about 140 tests on 98 years of US stock prices.
- Had a second AI agent redo the calculations on its own, and check the bot's code before launch.
- Built the bot, its safety limits and the daily report.
The research
Three well-known ideas, tested on 98 years of US stocks.
US stock prices from 1928 to 2025, costs included. The pass mark was set before the first test: at least 95% certainty that the advantage is real.
- Follow the trendHold stocks only while prices are above their average of the last 10 months. Otherwise, hold cash.
It cut the worst fall in half, −51% instead of −84%, but it didn't earn more. In the 2010s it cost about 7 points a year.
- Step back when the market is nervousHold fewer stocks when prices swing a lot, and more when they are calm.
The closest to a real advantage, but a small one, and it came mostly from a few crises, such as 1929 and 2008.
- Buy the dipBuy after a sharp short fall, and sell after the bounce.
It only worked before costs. It traded so often that the costs ate everything.
How sure can we be that the advantage is real?
Follow the trend62%
Step back when the market is nervous91%
Buy the dipalmost 0%
None of them reached the bar. The closest, stepping back when the market is nervous, became the pilot's base bot, not a proven edge yet.
And over the 98 years, simply holding the whole market still ended with the most money.
The pilot
Since 12 October, three bots trade US stocks on paper.
Each bot has its own paper account with 5,000 USD. The AI reads the news; fixed rules decide what to buy and when to sell.
A · The base
The nervous-market rule, on the S&P 500.
−1.54%1 order
C · News read by AI
Jev reads each headline about 15 large US companies. Clearly good news: buy. Bad news, or three days later: sell.
−0.41%15 orders
D · News read by a word list
The same rules, but a fixed list of words decides instead of the AI.
−0.49%15 orders
B · News read by Laya
Waiting: Laya runs on my PC and needs more disk space.
–not started
Returns as of 19 October 2026, after 6 trading days.
Does the AI read the news better than a list of words? It is only useful if it does.
Also studied
Bitcoin: safer, not better.
Before the pilot, we checked whether Bitcoin would be a better market. The best idea lost far less in the crashes but ended with much less money, and the certainty stopped at 72%. Bitcoin stayed out of the pilot.
Limitations
- Not a blind testThese ideas were found by other researchers in the same data. The only new evidence is what the pilot produces from now on.
- One marketOnly US stocks, the best stock market of the century. The result may not hold elsewhere.
- A short pilotA few weeks can't tell skill from luck. The pilot mainly tests the machinery.
- The AI can't be tested on the pastA model like Jev already knows what happened, so it can only be judged going forward.
- A home PCThe bot needs my PC on at 16:00. Hibernation once almost cost a day.
- Human errorAn API key was visible in a screenshot and had to be revoked.
- Not a productThis is not financial advice. Selling something like it would need regulatory approval.
Next
Let the pilot run to its review on 11 December, then continue, adjust or stop, as written in advance.
A first test with real money would only come after a clean pilot, with the loss capped at 200 CHF. It is proposed, not agreed.