# Forecast-Dojo study improves AI forecasting accuracy through replay training

_Published Tuesday, October 27, 2026 at 8:06 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers behind the Forecast-Dojo preprint report that training a Qwen forecasting model on another model's recorded interactions improved its evaluation accuracy from 34.8% to 42.8%. Their environment reconstructs resolved prediction-market events as dated steps, allowing agents to gather evidence and revise predictions in reproducible tests.

The environment contains 1,568 Polymarket events, divided temporally into training and evaluation sets, and 18.8 million dated news articles. Across 12 models, research tools consistently improved forecasting performance. Carrying a belief notebook between dates reduced research costs without consistently improving forecast quality, and all models still trailed historical market forecasts.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2609.28876)

---
Canonical: https://techandbusiness.org/newswire/y1O645lvoZBdCtDG_gpV3B
Published: 2026-10-27T12:06:33.446Z
Story chronology: 2026-10-27T04:00:00.000Z
Retrieved: 2026-10-27T14:40:29.337Z
Publisher: Tech & Business (techandbusiness.org)