Investing & Dividends Misleading — the headline number is real but unrepresentativ

By the Income Reality Check Editorial Team · Updated 2026-09-11 · 7 min read

GPT-6 Astra as a 24/7 stock trader: the 8% claim, checked

Verdict: Misleading — the headline number is real but unrepresentative. One winning month from an AI trader tells you almost nothing about the next twelve.

Nate Herk’s video “I Turned GPT-6 Astra Into a 24/7 Stock Trader” opens with a hook that does most of the work: he gave Claude $10,000 of real money to trade for a month and finished “beating the S&P by a little over 8%.” Now he’s running the same experiment with GPT-6 Astra, and the tutorial walks you through wiring an AI agent to a brokerage so it trades on a schedule while you sleep. The setup is real and the tools exist. Whether that 8% means anything is the whole question.

What the video actually claims

To his credit, Herk front-loads a disclaimer. “This is not financial advice,” he says, and he actively discourages handing an agent $10,000 cold. He recommends paper trading first, getting comfortable with a strategy, then automating pieces of it. That’s a more responsible framing than most of this genre.

The headline result is still the sales engine, though. A prior month-long run with Claude beat the benchmark by roughly 8%, and the promise of the new video is replication: you too can turn GPT-6 Astra into a trader that runs on its own. The method is specific. He uses Codex (OpenAI’s coding agent) connected to Alpaca, a commission-free brokerage with an API, and sets up six scheduled “wake-ups” per trading day — read the news at 7:45 a.m. Central, look for a qualifying trade at 9:30, review positions at 11:00, manage them at 1:00, start closing at 2:15, confirm flat before the bell. It’s a day-trading loop, not long-term investing.

Most of the tutorial is plumbing. Because each agent run is stateless, Herk builds a “continuity” system — a progress log and handoff notes so each wake-up reads what the last one did before acting. He pays Alpaca $99/month for real-time market data, connects ClickUp for notifications, and points all six routines at one shared thread. It’s a genuinely clever automation. That’s a different thing from a genuinely profitable one.

Does one 8% month prove anything?

No. And this is where the framing gets misleading, even with the disclaimers attached.

Beating the S&P 500 by 8% over a single month is well inside the range of pure luck. Markets are noisy over short windows, and a portfolio holding a handful of concentrated positions will routinely swing several percent away from the index in either direction for reasons that have nothing to do with skill. One month is a sample size of one. You cannot separate a smart strategy from a coin that happened to land heads until you’ve watched it flip many, many more times — ideally across different market conditions, including a bad one.

The broader research on humans doing exactly this loop is not kind. A study of roughly 360,000 day traders in Taiwan found 84.3% lost money, with a median return of about -8.7%. Regulators echo it: FINRA data has shown around 72% of day traders finish the year in the red, and the North American Securities Administrators Association (NASAA) has warned for years that most day-trading accounts lose money and that the activity “is not appropriate for anyone of limited resources.” NerdWallet cites research finding the average 20-day return on day-traded stocks was -4.7% — losses, on average, before you subtract data fees and taxes.

Academic tests of AI stock-picking land in a similar place once you leave the backtest. LLMs can look impressive on historical data — one University of Chicago study reported meaningful alpha from GPT-4 with chain-of-thought reasoning — but most real-money and out-of-sample experiments struggle to beat plain buy-and-hold once trading costs are included. GPT-6 Astra being smarter than its predecessors doesn’t repeal that. A better reasoner still can’t reliably predict the next hour of a stock price, because that’s not a reasoning problem — it’s a problem of information nobody has yet.

What the method actually requires

Set aside performance for a second and just count the costs, because the video treats them as trivia.

Then there’s the part the tutorial genuinely can’t automate: judgment about whether the strategy works. Herk had Astra invent the strategy by “fanning out 10 sub-agents” to do research — which means an AI wrote the plan, an AI executes the plan, and an AI reports on the plan. Nobody in that loop has an independent read on whether the approach has an edge. When the agent loses money, it will still produce a confident handoff note explaining why. That’s not a safeguard. It’s a narrator.

You’d also want to know the tax side. In the U.S., short-term trading gains are taxed as ordinary income, and frequent trading creates a paperwork trail (wash-sale rules, quarterly estimates) that a hands-off “24/7 trader” pitch quietly ignores. U.K., Australian, and Indian readers face their own capital-gains regimes — automation doesn’t change what you owe.

Who actually wins this game

Follow the incentives. Herk makes money whether or not the bot does — from YouTube views (this video had over 249,000), from his free school community that captures your email, and from being early and credible on a topic people are searching for. That’s a legitimate business. It’s just a different business from trading.

Among people who actually run bots like this, the consistent winners are a narrow group: quant-literate operators who backtest across years of data, size positions tiny, and treat losing strategies as disposable. The overwhelming majority of retail day traders — the audience most likely to copy a YouTube tutorial — sit in that 70–85% who lose. Wrapping the same loop in a smarter language model doesn’t move you from the losing group to the winning one. It just lets you lose faster and with better-formatted logs.

What you’d realistically earn

Here’s the honest range. Over one month, an AI trader might beat the S&P by 8% — or trail it by 8%, or blow a chunk of the account on a single bad session. All three are ordinary outcomes for a concentrated, high-frequency strategy over 20 trading days.

Over a year, the base rates take over. If 72% of day traders lose money and the median short-term return is negative, the most probable result for a beginner copying this setup is a loss, plus $99/month in data fees, plus AI subscription costs, plus the tax and time overhead. The realistic expectation isn’t “beat the market on autopilot.” It’s “underperform a boring index fund while paying for the privilege.” A note worth sitting with: the pattern day trader rule that once required a $25,000 minimum was eliminated by the SEC in 2026 (SEC filing; NerdWallet), so the guardrail that used to keep small accounts out of rapid-fire trading is largely gone.

If you want to see where the money actually flows in AI side hustles, our looks at proven ways to make money with AI with no experience and 16 stocks to buy now are more grounded starting points than automating a day-trading loop.

Who this is (and isn’t) for

This makes sense for exactly one profile: someone who already trades, already has a tested strategy with a real edge, and wants AI to handle the mechanical monitoring so they stop pocket-watching charts. Herk describes that person well — and for them, the automation is a productivity tool, not a money machine.

It does not make sense for a beginner with $10,000, no trading experience, and a hope that a smart model will figure it out. If you can only risk money you can’t afford to lose, this isn’t a side hustle — it’s a leveraged bet dressed up as engineering. NerdWallet’s own guidance is to risk no more than 5–10% of a portfolio on active trading, and to day trade only with money you can lose.

What to remember

The tutorial is technically sound and honestly disclaimed, and the 8% figure is probably real. It’s also a single month from a single run, which is the weakest possible evidence that a trading strategy works. Automating the loop doesn’t change the odds that have held across decades of day-trading data — it just removes the human who might have hesitated. If you already have an edge, use AI to save time. If you don’t, no model, however new, will manufacture one for you.