Can an AI Researcher Beat the Market? A 6-Month Experiment
Published on
The Experiment: AI-Managed High Growth Stock Trading
I've always been fascinated by the intersection of artificial intelligence and financial markets. The question that keeps coming up is simple but compelling: can an AI research tool, given access to the right data and the ability to act on its own recommendations, actually manage a profitable high-growth stock trading portfolio?
To find out, I've built a system to put this to the test - and I'm going to let it run for six months to see what happens.
The Setup: Trading Playground
The core of this experiment is a project I've been working on called Trading Playground. It's designed as a practical testbed for AI-driven trading decisions, connecting an AI deep research tool to a live trading execution environment.
Here's how the system works at a high level:
1. AI Research & Recommendations
Each day, the AI researcher kicks off by analysing the current state of the portfolio. It doesn't just look at what's in the portfolio - it casts a wider net, considering:
- Current holdings and their recent performance
- The broader market landscape - sector trends, macro-economic indicators, and market sentiment
- Emerging opportunities - stocks showing high growth potential based on current momentum, earnings reports, analyst sentiment, and technical patterns
- Risk factors - potential downsides, overexposure to specific sectors, and overall portfolio balance
The AI then produces a detailed research report and, crucially, a set of actionable trade recommendations. These aren't vague suggestions - they're specific: buy X shares of stock Y, sell Z shares of stock W, with clear reasoning behind each decision.
2. Structured Data Exchange
The recommendations are exported as structured JSON into the Trading Playground application. This is a key part of the architecture - by using a well-defined data format, the AI's recommendations can be cleanly consumed by the execution engine without ambiguity. The JSON includes the stock ticker, the action (buy/sell), the quantity, and the rationale.
3. Trade Execution
The Trading Playground app has basic market data feeds that provide current pricing information. When it receives the AI's recommendations, it evaluates each trade against the current market data and the portfolio's available cash position. If a trade is feasible - meaning the market conditions align and there's sufficient cash to execute - it places the trade. If not, it skips it and logs the reason.
This happens once per day, keeping the approach grounded in a measured, daily rebalancing strategy rather than frantic intra-day trading.
4. Performance Tracking & Feedback Loop
Here's where it gets interesting. The Playground app tracks the portfolio's performance on an hourly basis throughout the trading day. At the end of each day, it compiles the results - gains, losses, portfolio value, individual stock performance - and exports this data back to the AI researcher as structured JSON.
The AI researcher then uses this feedback to inform its next round of analysis. It can see what worked, what didn't, and adjust its strategy accordingly. This creates a continuous feedback loop:
AI Research Report → JSON Recommendations → Trade Execution
↑ ↓
Daily Results ← JSON Performance Data ← Hourly Tracking
Why This Approach?
There are plenty of algorithmic trading systems out there, but most rely on pre-programmed strategies - moving average crossovers, mean reversion, momentum indicators. They're rigid. What makes this experiment different is that the AI researcher isn't locked into a fixed strategy. It can:
- Adapt to changing market conditions in ways a static algorithm cannot
- Incorporate qualitative factors like news sentiment and earnings surprises alongside quantitative analysis
- Learn from its own mistakes through the daily feedback loop
- Consider the broader context rather than just looking at price charts
Of course, whether this flexibility actually translates into better returns is exactly what this experiment aims to find out.
What I'm Measuring
Over the six months, I'll be tracking:
- Total portfolio return vs. a benchmark index (likely the S&P 500)
- Win rate - what percentage of individual trades are profitable
- Maximum drawdown - the worst peak-to-trough decline
- Sharpe ratio - risk-adjusted return
- AI reasoning quality - are the research reports making sensible arguments, even when trades don't work out?
The Risks
I want to be upfront about the risks and limitations:
- AI hallucinations - the researcher might make confident-sounding but fundamentally flawed recommendations
- Market conditions - a sustained bear market could make any strategy look bad
- Execution limitations - the basic market data feeds may not capture all the nuances of real-time trading
- Overfitting to recent data - the AI might chase recent trends that don't persist
This is an experiment, not financial advice. I'm genuinely curious whether the AI can outperform a simple buy-and-hold strategy.
What's Next
I've set the system up, funded the portfolio with some Monopoly money, and the AI researcher has started its daily cycle of research, recommendation, and review. I'm going to leave this running for six months and then report the results back here on the blog.
Will the AI beat the market? Will it crash and burn spectacularly? Will it play it so safe that it barely moves? I honestly don't know - and that's what makes this experiment exciting.
Stay tuned for the results. I'll be back with a full breakdown of how the AI performed, what it got right, what it got wrong, and what I learned about AI-driven trading along the way.