Does AI trading actually work?

Is AI trading profitable?

For most people who try it, no, and the reasons are ordinary rather than mysterious. Costs, slippage and taxes remove a real edge before they remove an imaginary one. Crowding erodes anything simple enough to be found quickly by many people at once. And the strategy that reached live trading was usually chosen because it looked best across many attempts, which is exactly the selection that produces results that do not repeat.

You will find confident numbers on both sides of this question. Be suspicious of all of them, including the pessimistic ones. Published win rates and failure rates for retail algorithmic trading are almost never derived from a sample anybody can inspect. We decline to quote a figure here for the same reason we would not quote one to you in the product.

Backtested and hypothetical results are not indicative of future performance. Nothing on this page is investment advice.

Why do AI trading backtests look so good?

Because a backtest of the winner is not a test. If you try one strategy and it works, that is weak evidence. If you try ten thousand and report the best, you have measured the luckiest draw from ten thousand noisy ones, and it will look spectacular whether or not anything real is there.

AI makes this worse in a specific way: it lowers the cost of a trial to nearly zero. A person testing by hand might try forty variants and remember roughly how many. A model can try tens of thousands in an afternoon and remember none. The number of attempts is the single most important input to whether a result means anything, and it is the number most systems never record.

This is not an argument against using AI for research. It is an argument that the count has to be kept, and that the bar the winner must clear has to rise with it.

What would it take to know whether an AI strategy works?

The question becomes answerable once you stop looking at the equity curve and start asking how the strategy was selected. Four checks do most of the work, and all four are standard in quantitative finance rather than anything proprietary.

• Deflated performance — discount the result by how many variants were tried, how long the sample is, and how skewed and fat-tailed the returns are.
• Probability of backtest overfitting — estimate how often the in-sample winner lands below median on data held out of the choice.
• Out-of-sample testing — evaluate on data that played no part in any decision, not merely on the last slice of the same file.
• Walk-forward validation — refit and re-evaluate as the sample rolls forward, the way the strategy would actually have been run.

None of these tells you a strategy will make money. Each of them tells you how much weight the backtest can carry. That is a smaller claim, and it is the only honest one available.

Can an AI trade for me?

Mechanically, yes. Brokers expose APIs, and connecting a model to one is not the difficult part. The difficult part is that automation multiplies whatever you point it at. If the strategy is a coincidence, automating it converts a coincidence into a schedule of losses, faster and with less friction than doing it by hand.

So the useful order of questions is: is this edge real, then how much would it survive costs, then and only then should anything be automated. Handing execution to a system you have not adjudicated is not delegation, it is amplification.

Lagias does not place orders, connect to a broker, or manage money, and it is not built to. It researches and grades strategies. Every decision, including the decision to do nothing, stays with you.

Is ChatGPT useful for stock analysis?

For reading and summarizing, yes — filings, transcripts, filings-adjacent noise, competitor mentions, the tedious parts of research. That use is real and undersold.

For judging whether a strategy is sound, no, and not because of the specific model. A general-purpose assistant asked to evaluate a strategy it just proposed will produce a fluent evaluation, and the fluency is uncorrelated with whether the evaluation is right. It has no access to how many variants were considered, no held-out data, and no incentive to disappoint you. Ask it to grade its own work and you learn how it writes, not whether the strategy holds.

So is any of this worth doing?

Yes, if you change what you expect from it. Used as a machine for generating candidates and killing them quickly, AI is genuinely useful: it can explore a space of ideas far faster than a person, and most of what it produces should die. Used as an oracle that names winners, it is worse than useless, because it is persuasive.

The uncomfortable part is that a well-run research process mostly produces rejections. A tool whose most common output is "this does not hold up" is behaving correctly. That is a hard thing to sell and an easy thing to verify, which is roughly the opposite of how this category usually works.

Common questions