Experiments · run on 29–30 August 2026
We built the “I gave an AI $1,000” trading machine. It made $0.
The viral version: hand an AI $1,000, let it read the news and trade, watch it reach $3,900 — and buy access to the system for a monthly fee. We built the machine it describes. Over a day and a half it checked the headlines 138 times and declined every trade.
Read this first. These were paper-trading experiments: live market data, simulated money, no real orders. The figures are from our own run logs and describe what happened in a short window. They are not a forecast, a strategy or advice. Disclosures
What we built
A news-driven trading loop, running against a broker's paper-trading account. Every five minutes it:
- Reads the latest headlines for a watchlist of ten stocks and five coins.
- Scores each story against a list of 22 kinds of market-moving event, with older news counting for less.
- Checks that money is following the story — unusual trading volume, moving in the same direction as the news. No volume, no trade.
- Sizes the position to risk 1% of the account, with hard limits: three positions at most, and trading halts for the day after a 2% loss.
- Writes down the exit before the entry — the stop and the target go in the journal before the order is sent.
- Logs the fill, the slippage and the fee separately, then reviews its own losing trades each night and adjusts its rules inside fixed bounds.
What happened
Nothing — and the log says why, line by line.
- In 138 cycles the machine never found a story that passed both tests.
- Once, a crypto headline scored high enough to trade. Trading volume was 4% of normal. The journal entry reads: no money behind the story. Skipped.
- Nine times, a stock story scored high enough — over a weekend, with the stock market closed. Skipped, nine times.
- The nightly review found no closed trades and changed no rules: not enough evidence.
The only orders it placed were two $40 test round-trips to prove the plumbing worked. They are logged with their slippage and fees, and are not counted as trades.
Why $0 is the interesting answer
A day and a half over a weekend is not a track record, and we are not presenting it as one. That is the point. Neither is a screenshot of an account balance.
A machine with real risk rules spends most of its life refusing to trade. Something that turns $1,000 into $3,900 quickly is either taking very large risks or showing you only the account that worked. Before believing — or paying a monthly fee for — any trading system, ask for what we are showing here: the full journal, including every trade it took and every trade it turned down, with fees and slippage on each line.
We will keep this page honest. If the machine runs long enough to produce a real record, the numbers go here — good or bad.
What to take from it
You do not need a bot to improve your results. The reliable gains are the boring ones: paying less in fees, and not paying tax you do not owe. Those are arithmetic, not prediction: what your retirement account's fees add up to, and what harvesting a loss saves before 31 December.
Method: paper-trading account, live news and price data, every decision journaled before it was acted on. Figures are counts from that journal for 29–30 August 2026. Published 4 October 2026.