Research
A 20-year backtest across more than 100 stocks by researchers from Edinburgh, UCLA and Sungkyunkwan found buy-and-hold beat every one of the 20-plus AI trading strategies tested on a risk-adjusted basis, ahead of the team's presentation at ACM SIGKDD this week
On steadier, low-volatility stocks, plain buy-and-hold returned about 7.9% a year and topped the field once risk was accounted for. One AI strategy posted a raw 14% return but with swings large enough to lose on a risk-adjusted basis, and a 55-year-old ARIMA statistical model outperformed most of the AI tools tested.
The headline finding, that a do-nothing strategy beat more than twenty AI trading approaches, is less interesting than why the researchers think earlier studies missed it. Most prior work on AI trading agents ran backtests over a handful of months on a small, often hand-picked set of stocks, sometimes as few as three names over three months, conditions that flatter whichever strategy happens to catch a favourable stretch of price action. This study instead ran over more than twenty years of data across more than a hundred stocks, deliberately including the 2008 financial crisis, the COVID crash, multiple bull markets and delisted stocks to avoid survivorship bias, the kind of design that is expensive to run and unglamorous to publish but is what actually tests whether an edge is real or an artefact of a lucky sample window.
The team, drawn from the University of Edinburgh, UCLA and South Korea's Sungkyunkwan University with UCLA and Oxford mathematics professor Mihai Cucuringu among the authors, is explicit that the paper has not yet been peer reviewed, which is worth keeping in view given how strong the headline claim is. What raises confidence in the result regardless is the behavioural pattern they document underneath the raw numbers, the AI strategies as a group traded cautiously when markets were rising and aggressively when markets were falling, which is close to the opposite of sound risk management and a plausible mechanical explanation for why paper returns looked fine while risk-adjusted returns did not.
That behavioural finding lands at an uncomfortable moment for anyone who watched July's AI-linked selloff play out across the funds already covered on this site, Millennium, Jupiter's Minghong-linked quant fund, the SOX index's worst month since 2008. If AI trading strategies really do lean into losing positions during drawdowns rather than cutting them, that is a mechanism, not just a correlation, for why AI-heavy books took the disproportionate damage they did last month, and it is a testable hypothesis the next drawdown will either confirm or undercut.
The paper's real test starts this week, not with peer review but with the presentation itself, at the ACM SIGKDD conference running August 9 to 13 on Jeju Island, in front of an audience with every incentive to poke holes in a twenty-year backtest that contradicts a fair amount of published AI-trading literature. For a niche where marketed retail AI trading tools routinely cite short-window backtests as evidence of edge, a rigorous long-horizon study reaching the opposite conclusion is exactly the kind of result that needs adversarial scrutiny before anyone treats it as settled, and exactly the kind of result retail platforms selling AI-picked portfolios have no particular incentive to publicise if it survives that scrutiny.
Read the original: Yahoo Finance - A 20-year backtest across more than 100 stocks by researchers from Edinburgh, UCLA and Sungkyunkwan found buy-and-hold beat every one of the 20-plus AI trading strategies tested on a risk-adjusted basis, ahead of the team's presentation at ACM SIGKDD this week. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.