Forecast-Dojo: all 12 models improve with research tools and all 12 still trail Polymarket on 1,568 replayed events

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.28876 replays resolved Polymarket questions against 18.8M dated news articles, so agents can research and revise at successive historical dates without waiting for new events. Research tools lowered Brier score for every model, and forecasts improved most at steps with more newly dated evidence, but every model trailed the historical market on both Brier and accuracy. A belief notebook carried between dates cut research cost without reliably improving forecasts. The replay setup also provides training trajectories with real outcome feedback.