I built an AI that reads every Trump post the moment it goes up, decides whether it could move the stock market and which way, and trades Nasdaq futures if it is confident enough. Then I ran it over every post from 2019 to mid-2026, on tick data, with real costs.
The posts. Every Trump post from 2019 to June 2026: Twitter until his ban in January 2021, then Truth Social from February 2022. Reposts and image-only posts were removed, leaving 26,238 text posts. Posts made within five minutes of each other were grouped into one event, so a ten-post rant counts once.
The AI. Llama 3 8B, a free open-source model, running on my own computer. It wasn’t fine-tuned or trained on prices. It sees the words of the post and nothing else: no date, no price chart, no news.
The market. Nasdaq-100 futures, using every trade printed on the exchange, so each entry and exit is a real price at a real second.
An AI that remembers what happened after a post can “predict” it perfectly. So before any trade was run, the model was asked 20 questions about events after mid-2023: the 2024 election, the 2025 tariff rounds, the April 2025 tariff pause, Iran in 2025. It answered 0 of 20. It thinks Kevin McCarthy is still Speaker and that Trump’s term ended in 2021.
It answered 3 of 3 control questions from before its cutoff correctly (FTX, the 2020 election, the March 2023 bank failures), so it is a working model with a 2023 memory, not a broken one.
That leaves one handicap in place on purpose: in 2025 the AI didn’t know Trump was president again. A real trader would. Telling it would have meant telling it the future.
The prompt. On a 50-post spot check, the first prompt called 88% of posts market-moving, including “Senate Democrats should all resign” as stocks down. The design allowed one fix before any trading result existed. It added one sentence telling the AI that most posts don’t move markets. That sentence names no topics, so it doesn’t sneak in what we already knew moves prices. After it, 76% of posts came back as irrelevant. Nothing was changed after that.
A lookahead bug. When several posts were grouped into one event, the first version timed the trade from the first post, but the AI had read all of them. In 57 trades it had read words posted up to 15 minutes after it entered, including the single biggest winner. It was fixed so the clock starts at the last post, which cut the profit by about a quarter. Every number on this page is after the fix.
Run at temperature 0, so the same post always gets the same answer. Gold and oil calls were tested too. Neither made money at any speed, so this page is about the Nasdaq.
The same trades, entered at different delays after the post, June 2023 – June 2026, one contract, after costs. The AI needs half a second to two seconds to read a post, and Truth Social has no live feed, so 5 seconds is about the fastest a bot at home could manage.
| Entry after the post | Win rate | Per trade | Total |
|---|---|---|---|
| Instantly · 0 seconds | 52% | +$20.23 | +$6,615 |
| 1 second | 50% | +$17.25 | +$5,640 |
| 5 seconds · the bot | 51% | +$13.40 | +$4,382 |
| 30 seconds | 53% | +$8.94 | +$2,924 |
| 1 minute | 49% | +$4.43 | +$1,448 |
| 5 minutes | 46% | −$0.99 | −$324 |
From 30 seconds on, the result can’t be told apart from zero. The money is in the first few seconds. A person reading the post and clicking buy is too slow.
For any of this to work, his posts have to move prices. We compared how much the Nasdaq moved in each minute after a post with the same minute on nearby days when he didn’t post. Second term, single posts only.
How much the Nasdaq moves in the first minute after a post, against a normal minute. “Major” = the AI called it up or down at 7/10 or higher. The spike fades fast: 1.5× in minutes 2–5 and 1.2× after that.
Movement in the five minutes after a post, against the same day’s normal, all posts on each topic:
| Posts about trade & tariffs | Nasdaq moves |
|---|---|
| First term · 2019–2020 | 1.43× |
| Out of office · 2022–2024 (all posts) | 0.95× |
| Second term · 2025–2026 | 1.83× |
As a candidate, nothing he posted moved anything. As president, trade posts move the Nasdaq about twice as much as normal. Posts about war only mattered in the second term (Iran), when they moved oil 1.7× normal. Posts on anything else, about half of everything he posts, move it 1.02×: nothing.
Jul 2024 – Jun 2026, 40 micro Nasdaq contracts per trade, 5-second entry, after costs.
Share of the two-year profit by six-month period (23, 95, 84 and 95 trades). Posts mentioning Iran made $110,680 of the $133,920 in Jan – Jun 2026.
His first-term trade posts moved the Nasdaq too (1.43×), and the AI may even remember those years, which could only help it. It still lost. The profit doesn’t carry over to his other presidency.
| Year | Trades | Win rate | Per trade | Total |
|---|---|---|---|---|
| 2019 | 174 | 36% | +$0.33 | +$58 |
| 2020 | 135 | 51% | −$4.16 | −$561 |
| 2021 · until his Twitter ban | 3 | 67% | +$6.33 | +$19 |
| 2022 | 20 | 30% | −$1.98 | −$40 |
| 2023 | 22 | 41% | −$2.66 | −$58 |
| 2024 | 39 | 46% | +$2.76 | +$108 |
| 2025 | 179 | 47% | +$5.44 | +$974 |
| 2026 · Jan – Jun | 95 | 61% | +$35.24 | +$3,348 |
One micro Nasdaq contract per trade, 5-second entry, 15-minute hold, after costs. 2019–2022 fall before the AI’s memory cutoff, so it may remember them.
The ten best trades of all 667, by profit on one contract. Together they made …, more than the whole strategy made in total. Without them, it loses money.
Loading…
Profit for one micro Nasdaq contract, 5-second entry, 15-minute exit, after costs. The video used 40 contracts, so multiply by 40. Times are US Eastern. Tap a post to read all of it.
All 667 trades the bot took, each with the post that triggered it, the AI’s call and confidence, the entry and exit prices, and the result. Nothing removed.
His major posts really do move the market, and an off-the-shelf AI can read which way, as long as it gets in within seconds. But nearly all the profit came from a handful of posts, most of it in one six-month stretch, and the same AI lost money in his first term. That looks like one unusual period, not an edge that lasts.
The same brutal test — matched controls, real costs, one honest out-of-sample shot — run on what we sell: edges that survive it.
View our strategies