Why you shouldn't let an AI agent run your portfolio

The pitch is seductive: an emotionless AI that watches the market around the clock and trades without fear or greed. When researchers actually tested that idea over twenty years, the AI lost to doing almost nothing.

“AI can’t feel fear, so it should invest better than we do” is one of the most intuitive ideas in modern finance. It’s also, so far, wrong, and the way it’s wrong tells you exactly where AI belongs in your process, and where it doesn’t.

The most rigorous test to date

Most “AI beats the market” claims come from short windows and a handful of cherry-picked stocks. In 2025, a team built FINSABER, a framework to test LLM-driven trading strategies the way you’d test any real one: across a long horizon and a wide universe. Two decades. More than a hundred symbols.

The failure had a shape. The AI agents were overly conservative in bull markets, sitting out rallies and underperforming a passive benchmark, then overly aggressive in bear markets, trading into the downturns and taking heavy losses. Read that again: the machine we hired precisely because it wouldn’t panic or chase… panicked and chased. It reproduced the two most expensive mistakes human investors make.

Why smarter models don’t fix it

The instinct is to blame the model and wait for a bigger one. But the problem isn’t intelligence, it’s that markets are non-stationary. An AI learns the patterns in its training data, and the one thing you can count on in markets is that the next regime won’t match the last one. A war, a pandemic, a rate shock, a bubble in a sector that didn’t exist a decade ago, none of it is in the history the model learned from. No amount of extra parameters teaches a system about a future that hasn’t happened yet.

That’s also why the AI’s backtests looked so good and its live results didn’t. A model flexible enough to fit twenty years of noise will always look brilliant in a backtest, and the flexibility that fit the noise is exactly what breaks when the future arrives. It’s the oldest trap in quantitative investing, in a new and more convincing costume.

So what is AI actually good for?

Not deciding. Reviewing. The thing large language models are genuinely good at is reading, taking something you wrote and telling you where it’s weak. Point that at a trading rule you designed and it becomes useful in a way autonomous trading never was: it can flag a threshold you overfit to the last five years, a rule that only ever fires in one market regime, two “different” positions that are really the same bet, an exit condition you never defined.

The difference is accountability. An AI agent takes the decision away from you and (the evidence says) makes it worse. An AI critic hands the decision back, with your blind spots circled. One replaces your judgment; the other sharpens it.

That’s the line Cetagon is built on. You write the rules, with a clear head. The AI critiques them (overfitting, regime sensitivity, correlation gaps) and you decide what to change. The machine never touches the trigger. Write a rule and see the critique →