Model card — JGracey-prime_radiant
Summary
Quantile forecasts of weekly confirmed-influenza hospital admissions for the US and its states/territories, in the CDC FluSight format: 23 quantile levels, horizons 0–3 weeks, reference dates on Saturdays. The submitted model is an ensemble: the per-quantile median of a pooled LightGBM quantile regressor and a validated replica of FluSight-baseline.
Intended use
Research and public-health situational awareness via the FluSight hub and the dashboard. Not a clinical decision tool; forecasts carry wide, honestly-quantified uncertainty and should be read through their intervals, never their medians alone.
Architecture
- LightGBM component — one booster per quantile level (23), trained jointly
across all locations with horizon as a feature, predicting the change in
4th-root per-100k admission rates. Per-location scale/center statistics are
fitted per forecast origin from as-of data only. Post-processing sorts
quantiles in transformed space, inverts, clips at zero, and rounds integers
at the submission boundary.
lightgbm==4.7.0exact-pinned: its determinism is binary-scoped. - Baseline component — a replica of FluSight-baseline, cross-validated against the official implementation to a season relative WIS of 0.999989 on fingerprint-matched vintages.
- Ensemble — per-quantile median of the two.
Data
NHSN weekly confirmed-flu hospital admissions, read exclusively through the
FluSight hub's target-data/ git history as a vintage store: training,
scoring, and anchoring never see data committed after the forecast origin.
Population denominators come from the hub's season-correct auxiliary-data
snapshots. No other data sources, no LLM calls, ≈$0 compute.
Evaluation
Rolling-origin backtests over three seasons (2023-24, 2024-25, 2025-26),
scored with WIS on natural and log(x+1) scales against a truth vintage pinned
to a stated as-of date; relative skill is computed on the common task
intersection vs FluSight-baseline. Full league tables: reports/backtest_*.csv.
| Season | ensemble rel. WIS | lgbm rel. WIS | Best official comparator |
|---|---|---|---|
| 2023-24 | loses to UMass-flusion | loses | UMass-flusion |
| 2024-25 | loses to UMass-flusion | loses | UMass-flusion |
| 2025-26 | competitive | 0.609 (wins) | UMass-flusion 0.625 |
Known limitations (stated, not hidden)
- Interval under-coverage: the LightGBM model's 50% intervals empirically cover ~34–40% of observations, worsening with horizon. Season-bagging is the sanctioned fix, deliberately parked.
- Single data source and single target; no rate-change/peak/ED-visit targets.
- 2022-23 is not backtestable (the hub's vintage history begins Oct 2023).
- The dashboard serves a frozen backtest bundle, not a live feed.
Provenance
Built agent-assisted with Claude Code under human-gated go-live controls;
every phase adversarially verified. See AI-USE.md and CONTRIBUTING.md.
Accountable: Jeremy Gracey.