An AI agent called Raven-Agent can now do what most forecasting models can't: actually trade on its own predictions.
Researchers introduced Raven-Agent as what they describe as the first autonomous trading agent designed specifically for prediction markets. The core problem it addresses is a gap that recent benchmarks have made hard to ignore — models that produce well-calibrated probability estimates still tend to lose money when those estimates get converted into trades. Raven-Agent adds a "belief-to-trade" layer that bridges that gap. Tested against a controlled replay of archived decisions, it was the only policy among all evaluated approaches to achieve both a positive return and a positive risk-adjusted return. The code is publicly available on GitHub.
The result matters because prediction markets are increasingly used as a serious evaluation ground for AI forecasting — if a model is genuinely good at predicting the future, it should be able to profit from that knowledge. Most can't, which suggests that raw probability calibration and decision-making under uncertainty are different skills. Raven-Agent is an early attempt to build the connective tissue between the two.
One controlled replay over archived data is a long way from live-market performance, and the researchers are careful to note the gap themselves — but the benchmark is more rigorous than a leaderboard score, and the open-source release invites scrutiny that abstract claims rarely get.