You tested 200 strategies with ChatGPT. Your best one is probably luck.

It has never been easier to build a trading strategy. Ask a model for an idea, have it write the code, run the backtest, repeat. In an afternoon you can test hundreds of variations.

That is exactly the problem.

#Why more tests means worse results

Every strategy you try is a lottery ticket. Test enough and one will look spectacular on past data by chance alone. If you then pick the best one, you have selected for luck, and luck does not repeat.

Researchers call the risk the probability of backtest overfitting. The more variations you try, the higher it gets, and the more you should discount the winner.

#Warning signs

• The equity curve is suspiciously smooth.
• Performance collapses when you shift a parameter slightly.
• It works brilliantly in one year and poorly in others.
• You cannot explain why it should work in one sentence.

#A practical defence routine

1. Hold out data you never touch. Keep the last 20% to 30% of history locked away until the very end.
2. Walk forward. Optimise on one window, test on the next, roll forward, repeat.
3. Count your trials. Keep a log of every variation. The more trials, the higher the bar.
4. Stress the parameters. A good strategy works across a range of settings, not one magic number.
5. Add realistic costs. Fees, spread, slippage and funding kill many "edges".
6. Paper trade, then trade small. Live data is the only out-of-sample test that counts.

#Using AI responsibly here

Language models are great at generating ideas and boilerplate. They are terrible judges of their own results, and they will happily "improve" a strategy until it fits noise perfectly. Keep the evaluation rules fixed and outside the model.

This is education and personal experience, not financial advice. Crypto and leveraged trading can lose you money fast. We may earn from sponsors shown on this page; they never edit our conclusions. How we make money.