How I Made $700,000 In One Hour... Kinda

It's a Monday evening and I'm staring at my screen, watching numbers cascade through a backtest that would have turned $100,000 into $823,000 over twenty years. The S&P500 would have given me $650,000. That's a $173,000 difference. In about an hour of tinkering with ChatGPT.

Before you close this tab thinking I've gone full crypto-bro, hear me out. This isn't about get-rich-quick schemes or the latest meme stock. This is about what happened when I took a non-deterministic tool like an LLM and somehow got it to produce surprisingly accurate, repeatable financial analysis - while becoming a thought partner that asked questions I didn't know I needed to answer.

Over the years, I've followed quants like Meb Faber and Cliff Asness, fascinated by systematic approaches to investing. But getting institutional-grade backtesting tools? That's typically over $1,000 a year for a consumer-grade tool with basic market data. Sophisticated tool across multiple markets including intra-day data? A five-figure annual subscription. What if AI could democratise this kind of analysis?

So yes, the headline is tongue-in-cheek. But the implications are real. Here's how I went down this particular rabbit hole, what I discovered about the current state of AI financial analysis, and why ChatGPT unexpectedly beat every other model I tested.

Quick disclaimer: This is for educational purposes only. Past performance, future results, all that jazz. I've ignored transaction costs, slippage, and made simplifying assumptions that would make a real quant weep (for example, I've been very fast and loose about whether the Stooq prices were adjusted for dividends, splits etc). This is an AI exploration, not investment advice.

With that, let's dive in.

The Deceptively Simple Starting Point

Remember my blog about the world's best prompts (you can find it here: Best Prompts) ? A colleague had turned the advisory board concept into a panel of investing legends (think Buffett, Dalio, et al). That sparked an idea: could I get AI to build and backtest portfolios that beat the market at lower risk?

My first roadblock came immediately. Each LLM struggled to source historical data (Yahoo Finance has locked down UK access). I ended up at Stooq, a Polish site providing OHLCV data.

I tested Perplexity (strong finance features), Gemini, Claude, and GPT-5 (and Grok after I finished the blog). To cut to the punchline: they all got confused at various points except ChatGPT. Claude put up an impressive fight - more on that later - but GPT-5 won this practical test.

My opening prompt was embarrassingly simple:

"You are an expert financial analyst and advisor. I want you to run a 20 year backtest of holding and reinvesting in the SPY ETF. Please output key performance metrics and an overall account value chart assuming a 100k initial investment."

The results appeared very quickly. But what happened next surprised me - ChatGPT immediately warned about the 50%+ drawdowns during 2008. This wasn't just calculation; it was analysis.

Through various iterations, we arrived at a portfolio: a combination of QQQ (Nasdaq), SPY (S&P500), TLT (bonds) and GLD (gold). A simplified version of Ray Dalio's All Weather portfolio.

The results? Impressive. But what came next took it to the next level.

When AI Pushes Your Thinking

After delivering basic metrics, I asked ChatGPT to run Monte Carlo simulations so we could see what happens if history didn’t quite play out the same way in the future. It didn't just comply - it exceeded expectations.

Then, unprompted, it asked: "Do you want me to extend this by showing percentile bands across the entire equity curve path?"

This wasn't a bot following instructions. This was a collaborator pushing my analysis forward.

What emerged was beautiful: 10,000 possible portfolio paths with the actual historical path overlaid. I wasn't just seeing terminal values - I was seeing the entire probability space.

As MIT researchers found, 50% of AI performance gains come from the model, but the other 50% comes from how we interact with it (read more about that here: MIT Research Blog) ChatGPT didn't just analyse my portfolio; it pushed me to elevate my thinking.

The Precision Problem (And Solution)

But there was a flaw. Using Stooq's data meant the LLM had to manage data quality issues - filling gaps, handling anomalies. Across different runs and LLMs, I was getting 20-30 basis points variance in returns (more outside of the top two performers). Fine for sketching ideas, problematic for comparing strategies.

So I created what I call "analytical sovereignty" - maintaining control while AI augments. I asked ChatGPT to generate a YAML (YAML Ain't Markup Language - quite LLM friendly) file capturing our entire process. Here’s the opening section:

Now I had repeatability - this approach produced the same results to two decimal places every time. Time to give Claude another shot.

Claude's Surprise Rush To Try For The Win

In fairness, I dropped the market data and YAML into Claude and turned back to ChatGPT. When I returned, Claude hadn't just performed the analytics. It had built a full web app.

Functional, beautiful (well apart from the usual Tailwind purple familiar to AI-enabled developers; as an aside it was amusing to see the Tailwind CSS founder, Adam Wathan’s recent public apology for the indigo default).

The app came with file uploads and polished analysis. Classic Claude - always overdelivering (this was Opus 4.1 - Sonnet 4 may have gone the more traditional route).

Your Turn: The Backtest Challenge

Here's my challenge to you: take any investment strategy you're considering - could be as simple as a two-fund portfolio or as complex as a factor-tilted multi-asset allocation. But don't just run the numbers.

The Risk Interrogation Framework:

1. Start with basic performance metrics using market data of your choice

2. Ask for Monte Carlo simulations (10,000 runs minimum)

3. Request percentile band visualisations over time

4. Compare against simpler alternatives

5. Most importantly: ask ChatGPT what questions you should be asking

That last point is crucial. The best use of AI in financial analysis isn't getting answers - it's discovering better questions.

(Note 1: in the time taken to prepare the blog, ChatGPT and Claude were the only LLMs to deliver consistency and also worth noting, whilst I didn't fully verify the calculations - this is a blog rather than my day job after all - ChatGPT and Claude aligned on the outputs and I used the Portfolio Visualiser tool, for the sake of expediency, noted in the resources to check they were directionally correct

Note 2: I strongly believe in the value of independent financial advice and wealth managers - and I also believe in empowered decision making. I would always recommend getting advice and I wouldn't suggest this 4-asset portfolio in real world investing!).

The Bottom Line

My experimentation revealed two things. First, the value of having done this analysis manually before - that experience guided my AI interactions. Second, the power of AI as a domain exploration tool.

We're entering an era where AI doesn't just calculate - it collaborates. Success won't come from having the best models, but from knowing how to dance with AI in ways that amplify human judgment rather than replace it.

My portfolio analysis showed simpler might be better - the 4-ETF structure outperformed complex alternatives. But I never would have discovered this without ChatGPT pushing me to ask better questions.

As I have repeated before, the real risk in this kind of exercise isn't market volatility or model error. It's becoming so dependent on AI that we forget how to think critically about risk ourselves.

Until next time, you'll find me as the sun sets, talking to ChatGPT about tail risks while making sure I still remember how to calculate them myself...

Resources and Further Reading

Technical Tools:

• Portfolio Visualizer great free tool

  • https://www.portfoliovisualizer.com

• Monte Carlo Simulation

  • Outline of the method: monte carlo method

    
    

Essential Context:

• "When Genius Failed" great book on what happens when you over-rely on models

  • Lowenstein, R. (2000). When genius failed: The rise and fall of Long-Term Capital Management. Random House.

• "Fooled by Randomness" my favourite book on "risk" in its broadest sense

  • Taleb, N. N. (2005). Fooled by randomness: The hidden role of chance in life and in the markets (2nd ed.). Random House.

Practical Applications:

• Risk Parity vs. Traditional Allocation:

  • Bridgewater Associates. (2010). The All Weather Story. https://www.bridgewater.com/research-and-insights/the-all-weather-story