baselinenowcast comparator

Every new task=finalize, task=nowcast, and pipeline nowcast-stage evaluation includes an independent baselinenowcast point comparator. The goal is to improve on this reference, not only on unchanged reports. This does not implement the package's uncertainty estimation or sampling; beating the point comparator does not establish superiority to its full probabilistic forecasts.

The equations follow upstream estimate_delay.R and apply_delay.R. The Python implementation independently computes incremental reports from archived cumulative vintages, completes the triangle column by column using ratios of summed increments to earlier exposure, and computes a delay CDF from the completed column totals. It has no spatial pooling, identity prior, robust median fitting, or learned correction-strength selection.

Assumptions and adaptations

• Fit separately by signal, location and issuance using the latest 52 event weeks whose scheduled initial report date has passed. Delay 12 approximates maturity; these settings are fixed, not selected on evaluation scores. The window differs from Chromantis's 52 eligible pairs per delay.
• Only scheduled vintages available at issuance enter the fit, including causally available reports during held-out seasons. Reference truth never enters delay estimation. The current event week's visible reports may enter the triangle.
• Missing archived vintages are not zeros. Exclude an event row if any already-due vintage is missing or negative. Require at least one fully observed row and a finite positive delay CDF. Signed revision increments remain signed; a CDF above one permits downward corrections. These are low-level signed-revision equations, not a claim of matching all upstream preprocessing defaults.
• Counts use upstream's exposure denominator floor of one and successive increment expectation (current cumulative + 1 - previous CDF) / previous CDF * next PMF. Add each expected increment before processing the next delay. Continuous signals use a positive exposure denominator without that floor and report / CDF(age), preserving invariance to rate units. This continuous extension is an adaptation, not upstream count semantics. Clip estimates to nonnegative and source-specific physical bounds. Preserve unrounded count expectations.
• At age 12 or older retain the visible report. Missing current reports or unsupported triangles retain the persistence anchor. Finalization uses its existing last-observed-in-12-weeks / training-median fallback. Target nowcasts use the latest visible value at or before the target week within 12 weeks, or zero if none exists. No future event week or final truth fills these gaps. Status and retained triangle-row counts accompany predictions; fallback cells remain in scores rather than silently changing evaluation support.

Outputs and interpretation

Finalization saves baselinenowcast, baselinenowcast_status, and baselinenowcast_support alongside each prediction in finalizations.csv.gz. The method appears in all finalization metric/ranking tables and age-performance plots; baselinenowcast_scores.json also records fallback coverage.

Target nowcasts save baselinenowcast.npz with point predictions, truth, masks, dates, location order, status and support. nowcast_scores.json and nowcast-ranking.csv include its normalized MAE and baselinenowcast_point_crps_ratio: model fair CRPS divided by baseline point MAE (less than one favors the model). MAE is the CRPS of a deterministic distribution; both use the model's training-only Q95 scales and identical scientific cell weights. A zero baseline denominator produces an undefined ratio. Pipeline files live under nowcast/; forecasting continues to use its existing evaluation.

Existing attempts lacking baseline artifacts are incomplete under the updated manager. Add the comparator without refitting using python -m chromantis.experiment.planner baseline -e NAME, then run status and rank. The command checks pinned inputs, aligns the saved scored cells, and records separate baseline provenance without changing the original fit manifest. For new fits, re-plan to refresh pinned source before run.

Decision log

• 2026-10-01: Added the independent comparator to both per-signal finalization and neural nowcasts, including the pipeline stage, at the user's request. Implemented the point estimator first and explicitly distinguished continuous-signal adaptations and deterministic CRPS from the upstream uncertainty model.

• 2026-10-01 audit: fixed a count-expectation shortcut that did not reproduce upstream sequential application. Version baselinenowcast_point_v2 matches the pinned R low-level functions on 27 real NHSN triangles (flu/COVID/RSV, NC/CA/US, three issuances), with maximum CDF error below 1e-15 and point error below 3e-12 admissions. Before correction the largest discrepancy was 0.13 admissions. Regenerated the stored baseline.

• The configured comparator is not the upstream high-level default: upstream redistributes negative increments by default, and its default history allocation usually uses 18 weekly rows at maximum delay 12 with sufficient history, versus our 52-row signed-revision fit. Neither configuration has been selected using pre-evaluation validation. Continuous signals remain an explicit adaptation. Do not interpret this comparison as beating the default package.