planted

Can a machine tell a real pattern from random noise?

Most market-prediction software can't — and that's exactly how it fools people. planted is a 10-second demo that proves it.

An open piece of my machine-learning research on financial markets. — Tim Gordon

---

Show a fortune teller pure TV static and they'll confidently "see" a pattern in it. A surprising amount of trading software does the same thing — and only the losses, much later, give it away.

The a-ha

I gave two pattern-finders 2,000 days of pure random noise. There is no pattern. I made sure. Here's how often each one "found" one anyway:

One shrugged and said "nothing here." The other found a pattern in every single day of pure noise — the kind of method that looks like a genius in a backtest and goes broke the moment it meets the real world.

Finding patterns is easy. Knowing when there aren't any is the whole game.

Run it yourself — about ten seconds:

git clone https://github.com/timgordontg/planted && cd planted
python3 -m planted demo

No setup, no dependencies. Pure Python.

It's not just made-up data — here's the real S&P 500

Fair question: sure, it works on your invented markets — but what about real money?

python3 -m planted spx     # the same two finders, on 12 years of real S&P 500

I ran both finders on every trading day of the S&P 500 from 2014 to 2026 — then on the same data shuffled into meaningless noise, to see if they could tell the difference:

"I've seen this day before"…	…on the real S&P 500	…on the shuffle
the honest finder	2% of days	5% of days
the tourist	100% of days	100% of days

The tourist calls almost every day a repeat of an earlier one — in the real market and in pure noise. It can't tell them apart. The honest finder barely fires on either: a handful of days, the same low rate you'd get by chance. By its test, day-to-day S&P moves don't repeat in a way you could trade — so it stays quiet, instead of selling you a pattern that isn't there.

(Nothing here is graded against an answer key — real markets don't come with one. The demo only asks whether a method claims structure it can't back up. That honest question is the whole point.)

Why I built it

I build machine-learning methods for financial markets. The models that make money stay private — but how I keep them honest shouldn't be a secret, and it's the part worth sharing. planted is a tiny world where I can hide a real pattern, or hide nothing at all, and prove whether a method is honest enough to know the difference.

That discipline — refusing to believe a pattern until it beats pure chance — is what separates a real edge from a story that falls apart on live data.

If you build, or hire for, ML that knows a signal from a story — I'd love to talk.

Go deeper

python3 -m planted bench    # the full scoreboard — now with real patterns mixed in

How it works (for the curious)

planted invents synthetic markets that behave like the real thing — fat tails, volatility clustering, the works — and secretly labels each one as has a pattern or pure noise. A method hunts for repeating structure; the honest one only counts a find if it is stronger than the same hunt run on a scrambled copy of the data. Every method is then scored on two things at once: did it find the real patterns, and did it stay quiet on the noise? The headline score multiplies the two — so you cannot win by crying "pattern!" at everything.

It even grades itself: python3 -m unittest discover -s tests runs checks that fail if the framework is fooling itself. Under a thousand lines of pure standard library — readable in one sitting.

Bring your own pattern-finder

planted isn't a closed demo — it's an open frame. Write your own finder in about twenty lines, drop it in, and it gets the same honest, ground-truthed score as everything else:

python3 examples/custom_method.py    # a starter finder — edit it, re-run, watch the score move

Change how it sees the market (the features function) and the score moves in front of you: a richer representation usually finds more real patterns, but has a harder time staying honest on the noise. I'm not handing you a number to trust — you run it and read it off yourself. That's the whole reason it's public.

License

MIT © Tim Gordon