AI Agents That Compete, Not Agree, Forecast Markets 27% Better
A squad of rival AI forecasters beat one big model — your money could benefit.
A team of nine researchers has published a new method called CompEvo for forecasting numbers — things like stock moves, energy demand, or shipping volumes — by combining news headlines with historical data. The problem they tackled is a quiet one in AI: when you put several AI assistants (called "agents," meaning AI that can act on its own) on the same task, they tend to copy each other. One finds a piece of evidence, the rest pile on, and you end up with nine versions of the same opinion. Their other complaint was that the rules for updating each agent's strategy were basically guesswork.
The fix borrows from biology and game theory. Instead of cooperating politely, the agents compete. Each one reads news differently and makes its own call. The agents whose forecasts turn out accurate gain more influence next round; the others adjust or fade away. Think of a newsroom full of analysts, each following different sources — the ones whose predictions pan out earn more weight over time, while the specialists who spot unusual signals keep their niche instead of being steamrolled by consensus.
The results are notable. Across four real-world datasets, CompEvo cut forecast error by 27.3% (a measure called RMSE) and 26.2% (MAPE) compared with strong existing methods. Translated: predictions were roughly a quarter more accurate. The researchers also checked that the agents stayed diverse and specialized rather than collapsing into one shared voice — the exact failure they set out to prevent. This matters for market risk monitoring and resource scheduling, where being a bit early and a bit right is worth real money.
The honest catch: this is an academic paper, not an app you can download. It was tested on historical data, and real markets can behave in ways the past never showed. Forecasts can also sound confident while being wrong. Still, the direction is clear — banks, utilities and logistics firms all want sharper predictions, and whoever deploys this first gets an edge.
- CompEvo uses several AI assistants that compete rather than agree, so they don't all repeat the same opinion.
- On four real datasets, forecasts were about 27% more accurate than strong existing methods.
- It's a research paper, not a product — no app, and real-world markets are messier than historical tests.
Why It Matters
Sharper forecasts could mean better risk alerts, energy planning and investing — but it's lab work, not a live product.