Model runs¶
One line per experiment phase, newest first. Each report's "Model choices" table says what its models trained on, what they were scored on and with which seasons (Model choices A–F). Submissions are in the submission log; how to run an experiment is in the workflow.
Influenza submission, October 2026¶
- Submission choice (8 October): why B7 replaced the on-time System2 file, from the B7 versus System2 comparison on real reports.
- B7 production (7 October): three recipes retrained on all four seasons, ten seeds each, and their 7 October forecast; this is the merged submission.
- B7 folds (7 October): individual recipes retrained with the prescribed 2025–26 reporting errors, three held-out seasons (versus System2).
- B6 (7 October): search, extension and the System2 selection; results: original campaign, extension, System2, saved ensembles, by horizon, reporting regime on 7 October.
- Season peak (7 October, stopped): whole-season paths versus peak loss and autoregressive rollout.
- B4.polish and B5 (6–7 October): new seeds, training windows, B5 exploration and confirmation (plan).
- B3 pilot and B4 flu study (5–6 October): consolidated analysis, refineTop2, B3 pilot (plan), B4 versus Google.
Reporting delays, 1–5 October 2026 (training routes since deleted)¶
- End-of-run summary and combined technical results; artificial vintaging, separate or joint nowcasting, matched comparison with Google, 256 versus 2,048 samples.
- Expanded B2 revision comparison, C1 with local nowcast errors, B2 reporting-error augmentation, input-encoding controls, input replay.
- Nowcasting: context-conditioned nowcaster, seasonal online nowcasting, local reporting triangles.