AI trading bots can crunch numbers without sweating, but a 20-year test says they still cannot time markets.
According to the Wall Street Journal, researchers backtested several LLM-based trading strategies across 20 years and found that most failed to outperform the market. The machines were too cautious in bull markets, then too reckless when bears started chewing the furniture.
That is awkward for the AI booster crowd, which has been selling the idea that a chatbot with a spreadsheet habit can outsmart decades of market chaos.
UCLA professor of mathematics Mihai Cucuringu said the study showed the shine wears off once the test gets longer and messier.
“The key finding was that the apparent advantage of LLM-based strategies largely disappears when you evaluate them over longer periods and across a broader set of stocks,” he said .
Earlier studies had made AI trading look tasty, but Cucuringu said many used only a few stocks and shorter windows. His team dragged the bots through the 2008 financial crisis, the Covid crash and the bull runs between them.
The researchers included delisted stocks to avoid survivorship bias, the old backtesting wheeze where dead companies quietly vanish and the results look less rubbish.
Allison Investment Management partner David Allison said “A lot of models work well when the market environment is fairly stable. Financial markets are not stable, though.”
AI can spot patterns, but patterns from long bull markets often fall apart when shares suddenly go pear-shaped. Wars, pandemics, new technology and changing business models can turn yesterday’s signal into today’s expensive noise.
“I think investors should be skeptical of any trading strategy that uses AI to beat the market. It’s a big misconception that better models automatically translate into better trading performance,” he said.
The danger is that tomorrow’s bigger models may build fancier nonsense from the same noisy data. That does not make AI useless for investors. Allison said clients increasingly use it to shape basic investment questions, where the answers can be reasonably accurate.
If you have a financial planner who charges by the hour, you could save some time and money by using AI to narrow down your question,” Allison says.
AI can chew through annual reports, earnings call transcripts and historical financial statements without moaning about coffee. That is useful grunt work, provided investors remember it is not a crystal ball.
Human advisers can be fiduciaries, with legal duties to act in clients’ best interests. A bot has no such burden and will not look embarrassed when it stuffs up.
