Nathan Smith isn’t your typical investor. For one, he’s 17. For another, he wasn’t trying to get rich, he was trying to test a theory. Can AI run its own portfolio?

What would happen if you gave a generative AI model full control over a real-money stock portfolio?

In June 2025, Nathan launched what he called a “real-money experiment” using just $100 and a free version of ChatGPT-4o, a large language model developed by OpenAI. The rules were strict: the AI could only invest in U.S.-listed micro-cap stocks (companies worth less than $300 million), had to work with whole-share positions (no fractional investing), and the portfolio would be judged over a six-month period, ending in December.

This wasn’t a paper-trading simulation. Nathan hooked the AI’s decisions into his brokerage account and manually executed the trades ChatGPT recommended. He treated the experiment as seriously as any professional backtest.

“I wanted to see if an LLM could generate real alpha,” Nathan wrote in the project’s first post. “Not in theory, but with actual money, under real-world constraints.”

The project quickly picked up attention online—not just because of the novelty, but because it was working.

Over the first ten weeks, the AI-powered portfolio soared over 32%, massively outperforming the S&P 500’s return of about 4.5% over the same period .

But how did it actually work? What exactly did the AI do—and how much of this was luck?

To the untrained eye, Nathan Smith’s trading setup might have looked like magic. But under the hood, it was a tightly scoped system powered by automation, discipline, and prompt engineering—not blind faith in AI.

The rules

Nathan set strict guardrails for the project:

• The AI could only pick U.S.-listed micro-cap stocks (market cap under $300 million).

• The portfolio was limited to $100 in total, and only whole-share purchases were allowed.

• The timeframe was fixed: 6 months, from June 23 to December 23, 2025.

• The AI had complete discretion over what to buy, when to sell, and how to manage risk, including stop losses and position sizing.

• Nathan acted purely as the executor. He provided daily market data to ChatGPT, implemented the trades it chose, and asked for updated instructions.

“I will only reach out to provide data and get decisions, nothing else,” he wrote. “You have full control… Your decisions must be based on deep, verifiable research.”

The build

Nathan used ChatGPT-4o, one of OpenAI’s publicly accessible models, and augmented it with a specific structure for how it received and responded to data. He created a consistent pipeline where each day’s closing prices, position sizes, and any pending news were sent into the model. Once a week, it was also allowed to use what Nathan called “Deep Research”—a prompt that told the AI to step back, review the portfolio from scratch, and make strategic decisions like a human fund manager might at a Monday investment meeting.

The AI would then:

1. Evaluate its current holdings.

2. Suggest trades.

3. Justify each move—whether tactical (short-term catalyst) or strategic (long-term value).

4. Set specific entry prices and stop-loss levels.

5. Document all changes in clear, reproducible steps.

Nathan coded basic tracking tools in Python to calculate returns, log trades, track benchmark comparisons (initially to the Russell 2000 and XBI, later to the S&P 500), and even calculate performance metrics like Sharpe Ratio, Sortino Ratio, Beta, and Alpha .

Keeping track

Early on, the experiment attracted some criticism for being unclear or anecdotal. Rather than brush it off, Nathan responded by making the entire process radically transparent.

He committed to:

• Publishing full logs of prompts and AI responses.

• Posting raw performance data.

• Adding risk-adjusted metrics.

• Documenting every single mistake he’d made in tracking or execution.

“I want this to be judged by the same standards as anyone else,” he said. “I take this seriously and want the work to stand on its own—not be excused because of my age.”

That mindset set this project apart: it wasn’t a hype machine or a meme bet. It was a methodical test of whether a publicly available AI could act like a real-life hedge fund analyst, making decisions under constraints, in a high-risk part of the market.

So what did it actually buy?

With the system in place, ChatGPT was given its mission: build the strongest possible stock portfolio using only micro-cap U.S. equities and a $100 budget.

Micro-caps are volatile, often illiquid stocks with market capitalisations under $300 million. They’re risky, sensitive to hype, and difficult to analyse—but for a nimble trader or an enterprising AI, they can offer massive upside.

Early picks and adjustments

In its first week of “deep research,” the AI settled on three small companies:

1. Myomo Inc. (MYO) – A medtech company developing wearable robotics to help people with neuromuscular disorders. Chosen for its Medicare-driven growth and improving revenue.

2. Candel Therapeutics (CADL) – A biotech firm working on cancer treatments, with positive Phase 3 results in the pipeline.

3. CPS Technologies (CPSH) – An industrial turnaround play with defense contracts and actual profitability, rare in micro-cap land .

Each decision was justified in writing, and orders were placed with limit prices and position sizes designed to balance the budget—roughly $30–$40 each.

“I think the AI won’t find totally hidden gems,” Nathan said early on, “but it’s making rational, verifiable decisions about companies with data behind them.”

As weeks went on, ChatGPT began to shift its strategy. It leaned more heavily into biotech names with binary catalysts—events like FDA approval decisions or clinical trial results that could make or break a stock overnight.

By week 9, the portfolio was up over 32%, led by a heavy bet on aTyr Pharma (ATYR), which alone made up 50% of the total portfolio. Other holdings included:

• Abeona Therapeutics (ABEO) – transitioning to commercial stage with a funded runway.

• Fortress Biotech (FBIO) – picked for its upcoming FDA decision.

• 4D Molecular Therapeutics (FDMT) – a tactical buy following positive trial data .

Each move included entry prices, stop-loss limits, and clear reasoning based on news flow, earnings catalysts, or technical setups. For instance, it exited a position in Axogen (AXGN) after an FDA decision was delayed, rebalancing into stocks with nearer-term events.

Risk Management and Metrics

Despite the risky nature of the stocks, the portfolio showed signs of structured risk control:

• Stop-losses were enforced and adjusted over time.

• Sharpe and Sortino Ratios (measures of risk-adjusted return) were calculated and posted.

• Nathan even measured Alpha and Beta vs the S&P 500, though he acknowledged the sample size was too small to draw real statistical conclusions .

Performance as of week 10:

• Total Return: +32.6%

• Max Drawdown: −7.1%

• Sharpe Ratio (annualised): 2.96

• Sortino Ratio (annualised): 5.26

• Alpha vs S&P500: 182.96%

These are exceptional numbers, particularly for a portfolio made up entirely of volatile micro-caps. Even professional funds rarely achieve this level of alpha in such a short timeframe.

AI's biggest weakness

Nathan did point out one emerging flaw. As the portfolio evolved, ChatGPT started to fixate on high-risk, low-probability biotech events—drifting from diversification toward a concentrated “lottery ticket” approach. While this had paid off so far, it made the strategy more fragile.

“With only four holdings, leaning too heavily into binary outcomes turns the portfolio into more of a lottery ticket than a repeatable strategy,” Nathan warned .

Still, the AI’s ability to justify trades, manage risk, and adapt to news was far more sophisticated than many expected from a public model working under tight constraints.

Shutting it down

After 10 weeks of trading, a 32% return, and hundreds of AI prompts, Nathan—still just 17 years old—decided to pull the plug on his experiment.

It wasn’t because the portfolio failed. In fact, it had dramatically outperformed the broader market, with results many professional traders would envy. But as the project grew, its original purpose was becoming less clear.

“It had become a bit of a meme portfolio,” he said. “People loved the returns, but I felt like I was becoming more of an entertainer than an experimenter.”

The project had started as a genuine test: Can an AI build and manage a real-money portfolio using logical reasoning and open financial data? But as his social media following grew—largely thanks to screenshots of 10% weekly gains and AI-generated stock picks—expectations began to shift.

Followers began requesting daily updates, asking ChatGPT for new tickers, and meme-trading some of the AI’s selections. One of the stocks even pumped over 10% intraday after it was posted publicly—an echo of the “SPAC stack” or WallStreetBets frenzy.

Nathan didn’t want to fuel that kind of speculation.

“It felt like we were slipping into hype and influencer territory,” he said. “That’s not what I built this for.”

Despite shutting it down, Nathan came away with real insights:

• AI can do a lot more than people expect, but it needs well-defined constraints and constant human oversight.

• Risk management is the hardest part, especially with volatile assets. ChatGPT handled this surprisingly well, using stop-losses and adaptive sizing.

• Garbage in = garbage out. The biggest determinant of success wasn’t the AI’s brain, but the clarity of the prompt and rules Nathan used to guide it.

He also recognised the AI’s natural bias: when left unchecked, it would often chase catalysts and fixate on short-term wins over long-term fundamentals.

“It could read earnings calls, find FDA dates, do sentiment checks—it was brilliant. But it wasn’t thinking in decades. It thought in weeks.”

Nathan isn’t done building.

He’s already working on a new project—a web-based app that uses AI to summarise earnings calls, pick up anomalies in trading volumes, and suggest potential breakout stocks before they trend.

He also wants to revisit the trading bot idea with a more robust stack:

• API-connected execution (rather than human input of orders)

• Real-time data ingestion

• Multiple LLMs working in tandem—one for sentiment, one for valuation, one for strategy

Eventually, he hopes to build a fully autonomous research assistant, not just a trading tool. Something that could read, digest, and recommend across 10,000 listed companies without bias—or hype.

For now, though, the experiment is paused. But the question lingers:

If a 17-year-old with $100, a laptop, and ChatGPT can do this… what happens when institutional capital tries the same?