The authors introduce an agentic forecasting environment built from over 2,100 resolved Polymarket questions. They train a Qwen3.5-35B-A3B model using single-epoch GRPO with a Brier-score reward, where the agent acquires its own context through web search, page reading, and financial time series, all filtered to information available before each question's cutoff.
Training improved calibration by 30-40% and reduced average search attempts from 3.8 to 2.25 per rollout, indicating learned evidence discipline. The trained policy outperformed every frontier model tested, including Claude Opus 4.5 (soft-Brier 0.254 vs. 0.256 on 265 held-out questions), at roughly 5% of the inference cost, with the largest advantage on the hardest questions.
This work demonstrates that folding evidence-gathering into the reinforcement learning loop, rather than freezing research context before training, significantly improves forecasting performance. The environment, dataset, and per-rollout records are released as a reusable harness for temporal forecasting agents.