Skip to content

Prime Radiant

Calibrated forecasting. Two threads share this repo and a calibration thesis:

  • FluSight epi forecaster — CDC FluSight quantile forecasts (WIS-scored) for weekly confirmed-flu hospital admissions. Pure statistical pipeline, no LLM calls, ≈$0 compute. Registered as JGracey-prime_radiant.
  • Metaculus LLM bot — LLM retrieval → ensemble → aggregation on binary questions, Brier-scored. Parked at its retrieval phase.

Start here

  • Live dashboard — choropleth of predicted 3-week change, per-state fan charts, reliability curves, league tables, all served from a frozen, byte-regenerable bundle.
  • Model card — architecture, data, honest evaluation (including the seasons it loses and the under-coverage debt).
  • The Prime Radiant Record (PDF) — the complete build write-up: eight phases, the gate chain, and the adversarial ledger.
  • Repository — README carries the three-season league tables; CONTRIBUTING carries the gates; AI-USE declares how AI built this; CHANGELOG is release-automation-managed.

The one-paragraph method

NHSN weekly admissions (hub target data, git-vintaged) → per-100k 4th-root transform with per-location scale/center fitted per origin → pooled LightGBM quantile regression (23 levels, horizon-as-feature, deterministic, exact-pinned) → monotone sort, inversion, integer rounding at the hub boundary → per-quantile-median ensemble with a validated baseline replica → WIS/coverage scoring on natural and log scales, pinned-truth, common-task relative skill. Leakage invariants are property-tested, not commented.