Source catalog

The checked-in catalog defines 23 independently retrievable datasets. The table below summarizes acquisition and modeling metadata; run python -m chromantis.data catalog for the complete machine-readable specifications, signals, natural keys, and source URLs.

Dataset key	Provider	Fetcher	Cadence	Versioned	Native geography
cdc_nhsn_final	CDC	socrata	weekly	no	state, territory, hhs, nation
cdc_nhsn_initial_release	CDC	socrata	weekly	yes	state, territory, hhs, nation
cdc_nhsn_preliminary	CDC	socrata	weekly	no	state, territory, hhs, nation
cdc_nssp_daily	CDC	socrata	daily	no	state, nation
cdc_nssp_demographics	CDC	socrata	weekly	no	nation
cdc_nssp_trajectories	CDC	socrata	weekly	no	state, hsa
delphi_claims_inpatient	CMU Delphi	delphi_v5	daily	yes	county, hrr, msa, state, hhs, census region/division, nation
delphi_claims_outpatient	CMU Delphi	delphi_v5	daily	yes	county, hrr, msa, state, hhs, census region/division, nation
delphi_flusurv	CMU Delphi	delphi_v5	weekly	yes	state, nation, msa, misc
delphi_fluview_clinical	CMU Delphi	delphi_v5	weekly	yes	state, nation, hhs, census division
delphi_fluview_ilinet	CMU Delphi	delphi_v5	weekly	yes	state, nation, hhs, census division
delphi_nhsn	CMU Delphi	delphi_v5	weekly	yes	state, nation, hhs, census region/division
delphi_nssp	CMU Delphi	delphi_v5	weekly	yes	state, nation, hhs, county, hsa, hrr, msa, census region/division
delphi_nwss	CMU Delphi	delphi_v5	sample	yes	sewershed
delphi_nwss_aux	CMU Delphi	delphi_v5_aux	sample	yes	sewershed
derived_nwss_state_indices	Chromantis, derived from CMU Delphi NWSS	derived	weekly	yes	state, nation
hub_covid_current	Hubverse community	hubverse	weekly	yes	state, nation
hub_covid_legacy	Hubverse community	hubverse	weekly	yes	county, state, nation
hub_flusight_current	Hubverse community	hubverse	weekly	yes	state, nation
hub_flusight_legacy	Hubverse community	hubverse	weekly	yes	state, nation
hub_rsv_current	Hubverse community	hubverse	weekly	yes	state, nation
hub_rsvnet	Hubverse community	hubverse	weekly	yes	catchment, nation
pophive_kinsa_ili	PopHIVE (Kinsa Insights)	pophive_git	daily	yes	nation

The Versioned column includes both preserved initial releases and archives with multiple release dates. A yes does not necessarily mean every revision is available; see the NHSN distinction below.

For the consumer-facing organization and retained measures, see Shared selection: 23 acquisitions map to 11 logical groups.

Source families

CDC Socrata

Downloads use a source pre-count, deterministic ordering, complete offset pagination, and a post-download count check. Payloads are streamed rather than held in memory.

NHSN initial releases and revision history

cdc_nhsn_initial_release is the local catalog key for CDC's Weekly Hospital Respiratory Data (HRD) Metrics by Jurisdiction (Historical) (Socrata ID rhwp-grxi). It contains weekly hospital admissions, hospitalizations, bed capacity and occupancy, and reporting coverage for COVID-19, influenza, and RSV, beginning November 2024. CDC states that values are not updated or adjusted after initial publication for a reporting week.

CDC therefore preserves first-release NHSN values in this product, but it is not a complete archive of every subsequent revision.

Catalog key	Release semantics
cdc_nhsn_initial_release	Frozen first published values for each reporting week; revision_mode=initial_release.
cdc_nhsn_final	Current finalized weekly values, which can incorporate later corrections.
cdc_nhsn_preliminary	Preliminary weekly release, available before the finalized release; not a full revision archive.
delphi_nhsn	Report-time vintages for available NHSN signals; reference_time identifies the observation week and report_time identifies the vintage.

For example, if a week's admissions are first published as 100, then revised to 110 and later 115, the initial-release dataset retains 100. Comparing it with the current finalized dataset can reveal the change from 100 to 115, but does not recover the intermediate value of 110 or its publication date.

Use initial releases to study first-publication error against current values. For backtests requiring what was known at a particular forecast date, use the available delphi_nhsn report-time archive or historical snapshots captured at the relevant cutoff. Check vintage coverage first: an early observation date does not imply that all releases since that date are archived. The catalog's versioned=True for cdc_nhsn_initial_release means first-release preservation, not a complete sequence of vintages.

Delphi Epidata

The Delphi V5 sources listed above are enabled (NHSN, NSSP, NWSS with its auxiliary table, both claims feeds, FluView ILINet, FluView clinical and public health laboratories, and FluSurv-NET). Direct CDC and Hubverse sources are separate.

Following Delphi's Python migration guide, the downloader uses EpiDataContext.epidata_meta() to discover and validate live signals and geographic resolutions, then epidata_archive() or epidata_snapshot() to construct each query. Its public request_arguments() method provides the canonical HTTPS endpoint and wire parameters. These requests select all reference_time dates and geographies within each signal/geography partition; --report-time and --snapshot-date select the revision view.

The raw repository downloads the resulting CSV with format=csv instead of converting through .df(), so large archives can be streamed without loading them into memory. This preserves all publisher columns, timestamp precision, leading-zero identifiers, zeros, and missing values. Queries keep every published fill_method by default, as the official client does; no local imputation is performed. Keep fill_method in observation keys when selecting or aggregating series. An explicit --fill-method can restrict the query. Authentication uses the V5 token header.

Large pulls use bounded concurrency, complete gzip validation, whole-partition retries, and explicit resume directories. A saved request.json prevents reuse under different source, signal, geography, fill, or version selectors. A resumed unbounded archive or latest snapshot can include new upstream revisions in its newly downloaded partitions; use an explicit historical version selector when a fixed information cutoff is required. Staging directories without request.json cannot resume.

Source details:

• NHSN: weekly admissions, reporting coverage, and bed indicators.
• NSSP: weekly ED indicators. The configured pull excludes the MSA partition; it can be requested explicitly with --geo-type msa.
• NWSS: COVID-19, influenza A, and RSV concentrations at native sewershed support. reference_time is sample collection date. Keep nwss_source and sample_index in the natural key and retain pcr_target. Location and population metadata are in Delphi's separate aux_data table; this downloader does not yet acquire that table or construct state wastewater aggregates.

• Inpatient claims and outpatient claims: daily seven-day percentages for influenza, COVID-19, and other ARI. These use the claims_inpatient_adm_pct_* and claims_outpatient_ov_pct_* signal names. Suppressed windows are absent rows; these feeds need revision-aware evaluation because claims backfill substantially. Claims rows carry blank fill labels even though the source docs describe source; imposing that filter would discard these rows.

• FluView ILINet, FluView clinical labs and FluSurv-NET: the V5 replacements (September 2026) for the legacy V3 fluview, fluview_clinical and flusurv endpoints. FluView rows carry an extra age_group key column. The public health labs (fluview_resp_lab_ph) are not acquired (state rows are season totals). Release timing, archive gaps and the New York pool are in FluView and FluSurv-NET.

Full reference-date history is not necessarily full historical release coverage. Inspect the saved reference_time_range and report_time_range before selecting backtest periods.

PopHIVE Git history

pophive_git reads one PopHIVE Ingest standard/data.csv.gz table and turns its commit history into a report-time archive: a row's report time is the first commit that contains it. It lists the file's commits through the GitHub API and downloads each version from raw.githubusercontent.com rather than mirroring the 10 GB repository. Commits can be backfills, so the earliest report times are when a value entered PopHIVE, not when its publisher released it. Details, terms and the Kinsa commit history are in Kinsa (PopHIVE).

Hubverse

Repositories are maintained as read-only, blob-filtered Git mirrors. Exports are pinned to an exact commit; --hub-as-of resolves the last first-parent commit at or before a historical UTC cutoff.

Canonical target files without row release dates are also exported across their first-parent Git history. chromantis.data hub-history DATASET... adds that history to a new acquisition at the previously saved commit without refreshing the branch. The canonical selector and explorer expose it as Git commit history, with the availability assumptions documented in Vintages and geography.

Claims

Inpatient and outpatient claims provide daily trailing-seven-day percentages. The panel selects Saturday values for influenza and COVID-19. These are covariates, not admission counts. A finite conflicting same-date report establishes source availability but cannot supply an unambiguous model value; see conflict handling.

Kinsa

Kinsa is the national PopHIVE ILI series, recovered from pinned Git snapshots. Weekly values require all seven daily observations. Episode construction broadcasts the national value and reporting mask to every location; it does not create state-specific measurements. Commit time is evidence of repository availability, not an assumed provider release time. As-of coverage begins with the retained archive, even when the final series extends farther back.

FluView and FluSurv

delphi_fluview_ilinet, delphi_fluview_clinical and delphi_flusurv use Delphi V5 archives. ILINet provides ILI percentages, clinical laboratories provide influenza positivity, and FluSurv provides surveillance hospitalization rates. Geographic support and gaps remain source-specific; a region or surveillance catchment must not be relabeled as statewide support. Observation history can be much longer than usable release history. The scheduled-final protocol assumes a one-week lag for these three groups; this is an experimental assumption, not measured archive timing.

Wastewater

The panel uses derived_nwss_state_indices, rebuilt from versioned Delphi NWSS concentrations and auxiliary metadata by python -m chromantis.dataset.build nwss-indices. These WVAL-like indices are derived predictors, not the published CDC WVAL series.

At each report time, use only positive finite samples and auxiliary metadata available by that time. Group by sewershed, NWSS source, PCR target and laboratory method. For groups with at least 26 distinct weeks and positive log-scale standard deviation, compute exp((log(value) - log_value_p10) / log_value_sd) and the within-group log-value percentile. Average within a Saturday-ending week, take medians across groups at each site, then across sites per state and nationally. Each state or national week requires at least three sites. Same-release conflicts become missing; disappearing state-weeks receive explicit null statements. The formulas and release-boundary handling live in chromantis.dataset.nwss.

The earliest retained Delphi NWSS archive vintage is February 25, 2026. Older observation dates do not establish historical as-of availability. Run index construction after changing NWSS inputs and before rebuilding the panel.