Foresight V3 was trained from historical open-web news. The SDK generated prediction questions from information available at each article timestamp, resolved them from later sources, and Foresight Learning reinforced the reasoning paths that produced better probabilities. The result reached #1 on Prophet Arena, where every model receives identical context.
What we did
- Generated forecasting questions from historical open-web news using only information available at the source timestamp.
- Resolved each question from later evidence, so outcomes came from what actually happened instead of human labels.
- Fine-tuned with Foresight Learning, where multiple probability rollouts are scored against the resolved outcome.
- Evaluated on Prophet Arena and live Polymarket questions against frontier models.
Read more
Primary artifacts for this case study.
Blog post blog.lightningrod.ai
How We Built the #1 AI Forecaster
Leading live prediction-market benchmarks against much larger general models
Blog post blog.lightningrod.ai
Foresight-32B Beats Frontier LLMs on Live Polymarket Predictions
Sharper calibration and profitable live trading without frontier-scale compute
Blog post blog.lightningrod.ai
Using the Future to Train Prediction Models
Turning time-stamped history into supervision your leadership team can trust
Paper arxiv.org
Future-as-Label: Scalable Supervision from Real-World Outcomes
Methodology for replacing manual labels with resolved events
Paper arxiv.org
Outcome-based Reinforcement Learning to Predict the Future
How outcome-based training improves calibration and accuracy
Notebook github.com
Polymarket backtesting
Evaluate LLM forecasting performance against historically resolved Polymarket prediction markets