← All articles
· By Bjørn Niklas Sjøstrøm

I let an AI write trading rules for the Oslo Stock Exchange. It lost money in six of eight years.

An AI got the prices of four energy stocks and wrote trading rules. On historical data the rules made 16 to 47 % a year. On years it had not seen, they lost money in six of eight.

NOK 25,000 became 16,500.

Build your own AI bot, let it day trade, and let your money work while you sleep.

This past year I have seen it everywhere, on Instagram and in plenty of other channels, and it is always sold as both simple and smart.

I had my doubts. Earlier I built a trading system where a language model made the decisions, and it had one trait I could never get rid of. It always wanted to do something. It traded too often. It bought at the top, and it handled price swings badly.

So I looked into it properly, using the Oslo Stock Exchange as the test ground and data from 2013 to today.

What I tested

An AI got the prices of four energy stocks, plus oil, gas and currency, and had to write a trading rule from them. I then tested that rule on a year the model had not seen. I did this eight times, one year at a time. Names and dates were removed, so the model could not remember what happened.

16–47 %annual return on the years the AI had seen
6 of 8years with losses on the years it had not seen
NOK 16,500left of NOK 25,000 after eight years
+251 %for simply holding Equinor over the same period

On paper the rules looked brilliant. In reality they lost.

I also tested insider buying, short positions, quarterly reports and overnight oil price moves. None of them made money after costs for someone trading short term.

Why it goes wrong

Three reasons.

AI finds patterns in noise too. Give a model enough historical prices and it will almost always find a rule that would have made money, but that is not an edge, meaning something the market has not already priced in. It is fitting to the past. You only see it when the rule meets a year it does not know.

Costs take the rest. Every trade costs commission, spread and a little price impact. At Nordnet's active trader pricing that is around 0.15 % round trip in Equinor, which sounds small until you trade every day for a whole year. Without a real edge, the chance of being up after one year is under 6 %.

With smaller amounts, or in smaller stocks, it is close to zero.

The news is priced before you can trade. On days when oil had risen overnight, Equinor rose 0.24 % on average from open to close. That looks like an edge. But the whole exchange rose on those same days, and measured against the rest of the market, Equinor actually fell 0.10 %.

The oil move was already in the opening price.

The research says the same

Researchers in Brazil followed everyone who started day trading equity futures between 2013 and 2015. Of those who kept at it for at least 300 days, 97 % lost money.

Only 8 of 1,551 earned more than a bank teller.

In Taiwan, day traders as a group lost money every single year for 15 years. A few are skilled, but less than 1 % made money year after year. The people you are up against are professionals with faster systems and lower costs than you.

What AI is actually good at

The one thing that held up in my tests was avoiding. Companies whose quarterly reports were badly received kept falling over the following months, and that held in both periods I tested. An AI can read the reports and help you stay away from the worst.

But even that did not beat a plain index fund after costs. A fund tracking the Oslo Stock Exchange main index returned 11.5 % a year from 2013 to 2026, and was up after one year in about four out of five periods.

What I am left with

AI makes it easier to build a bot. It does not make it easier to find an edge, and the edge does not get easier to find because everyone now has the same tools.

If you want to use AI on stocks, use it to read, sort and say no. Not to press buy every morning.

Sources

  • Chague, De-Losso and Giovannetti (2020), Day Trading for a Living?, SSRN.
  • Barber, Lee, Liu and Odean, studies of day traders in Taiwan 1992 to 2006.
  • Other figures are my own tests on data from Yahoo Finance, Newsweb and Nordnet.

This is not investment advice.