← All articles

I asked an AI to beat the market. It told me the truth instead.

By Stanley — Founder, NovaDyne · August 2026 · 7 min read

A few weeks ago I gave an AI a simple instruction: find me the best money-making trading bot.

I run an AI-first software company, I have the infrastructure, and like everyone else I've watched a thousand videos of influencers with screenshots of green candles telling me the market is a solved problem — you just need their pattern, their indicator, their $99 course.

So we did what nobody on social media ever does. We tested it. Properly.

The experiment

Over several weeks, my AI research agent and I ran what amounts to an industrial-scale audit of retail trading wisdom:

• 13 pre-registered studies covering about 120 trading strategies — every rule was written down, hashed, and sealed before we looked at a single result, the same way clinical trials work, so we couldn't fool ourselves after the fact.
• 46 million daily price bars across 21,936 US stocks over 28 years — including the dead ones. (Most backtests you see online quietly exclude every company that went bankrupt. Ours didn't.)
• Tens of thousands of SEC filings — insider purchases, activist stakes, buyback announcements — pulled and tested the same way.
• And when the results came back ugly, we audited our own machinery like an adversary, planting fake signals to prove the system could find an edge if one existed. It recovered a planted 0.50% edge to within 0.005%. It detected a known-real market anomaly across 932,899 events with overwhelming statistical confidence. The instrument works. That's what makes the readings believable.

What we found

Everything the influencers sell, measured honestly, falls into two categories: worthless or worse.

• Chart patterns? Cup-and-handle: statistically indistinguishable from buying on random days. Inverse head-and-shoulders: significantly worse than random — the breakout you're taught to buy is systematically the wrong moment.
• Breakout strategies in general: across 300,000+ simulated trades, worse than a coin flip, under three completely independent testing methods.
• Candlesticks, moving-average systems, momentum entries: null after null after null.
• Crypto "arbitrage"? The real version (funding carry) currently pays about 2% a year — before exchange risk. The version in your DMs is a scam.

Then we ran the question everyone actually cares about: what would $100,000 have become over the last ten years?

• A bot trading activist stock filings: $33,000. It lost two-thirds of the money.
• A bot trading buyback announcements: $247,000.
• Our very best signal — corporate insiders buying their own beaten-down stock, the one pattern in about 120 that shows genuine statistical life: $173,000.
• Doing absolutely nothing but holding the S&P 500: just over $400,000.

Read that again. The most sophisticated thing we built — sealed methodology, audited code, millions of data points — was beaten more than 2-to-1 by doing nothing.

The uncomfortable conclusion

Warren Buffett has been saying this for fifty years: buy broad, hold long, ignore the noise. He even put a million dollars on it — his famous 2008 bet that a plain S&P 500 index fund would beat a hand-picked collection of hedge funds over a decade. The index won. It wasn't close.

Social media tells you the opposite, every single day, because the opposite is sellable. Nobody goes viral saying "hold an index fund for thirty years." The influencer's edge isn't in the market — it's in your attention. Their winning screenshots survive; their losses don't get posted. There is a name for judging strategies by the survivors you can see: it's the exact statistical error our study design existed to prevent.

Our weeks of compute independently rediscovered what the world's most patient investor knew all along. The real edges that exist in public market data are about 0.3% per trade — almost exactly the size of the cost of making the trade. The market isn't a slot machine with a secret pattern. It's an aggregation engine that has already eaten every pattern simple enough to fit in a YouTube thumbnail.

The part that actually excites me

Here's the thing: the AI didn't fail. It succeeded at something much rarer than promising me riches.

It told me no. A hundred and twenty times. With receipts.

When its own first test run looked too good, it caught the statistical flaw itself and stopped the study. When I pushed it to check its work, it attacked its own code and quantified exactly how much it could and couldn't trust its readings. And where questions genuinely can't be answered from past data, it didn't guess — it set up eight sealed, tamper-evident forward experiments that are quietly collecting evidence right now, to be judged in years, not clicks.

That, to me, is what AI is actually for. Not oracles. Not confident hallucinations. Ascertainment — running the full scientific method at machine speed: hypothesis, pre-registration, test, adversarial audit, honest verdict. What would take a university research group years, we did in weeks — and the most valuable output was the disciplined no that protects you from every seductive, expensive yes being sold on your feed.

An AI that flatters you is a toy. An AI that can prove you wrong is a tool.

Why I'm telling you this

Because this is exactly how we build software at NovaDyne. AI-first doesn't mean "we sprinkle ChatGPT on it." It means every claim gets tested before it ships, every pipeline audits itself, and the system is designed to surface truth you didn't ask for — not just the answer you wanted to hear. Whether the question is "does this trading strategy work" or "will this feature actually move your revenue," guessing is expensive and knowing is cheap, if you build the machinery for knowing.

If you want that machinery working for your business, that's what we do.

And if you want the investment lesson for free, it fits in one line, verified by 46 million data points:

Buy the index. Hold it. Unfollow the guy with the Lamborghini.

This is a research story, not financial advice. Markets involve risk; past performance — as we just spent 46 million bars proving — guarantees nothing.