Prime Radiant

Prime Radiant

[](https://github.com/JeremyGracey-AI/prime-radiant/actions/workflows/ci.yml) [](./AI-USE.md) [](./LICENSE) [](https://huggingface.co/spaces/jeremygracey-ai/prime-radiant) [](https://jeremygracey-ai.github.io/prime-radiant/) [](https://pypi.org/project/prime-radiant/) [](https://app.codecov.io/github/JeremyGracey-AI/prime-radiant)

Calibrated forecasting, measured honestly. Two threads share this repo:

• FluSight epi forecaster (prime_radiant.epi) — CDC FluSight-format quantile forecasts of weekly US influenza hospital admissions: LightGBM quantile regression in the shape of the FluSight-2023/24-winning flusion model, ensembled with a validated replica of CDC's own baseline. Pure statistical pipeline; no LLM calls.
• Metaculus bot (prime_radiant.metaculus) — LLM forecasting on binary questions (parked at its retrieval phase; client + tests remain green).

Backtest results (three seasons, vintage-honest)

Rolling-origin backtests over 85 weekly origins across the 2023-24, 2024-25 and 2025-26 seasons. Every forecast is trained only on the data snapshot a live forecaster would have held on the Wednesday submission evening (the hub's git history is the vintage store); scored against final truth as of 2026-07-09, horizons 0-3, all 53 jurisdictions. `wis_rel` = mean WIS relative to the official FluSight-baseline on the common task set across all six models (lower is better; the official "scaled relative skill" collapses to this ratio on i