Datasets:

APProjects
/

us-warn-act-layoff-notice-changes-daily

pretty_name: >-
  What changed in US WARN Act layoff notices (daily observed change log,
  amendments and disappearances)
license: cc-by-4.0
language:
  - en
task_categories:
  - tabular-classification
tags:
  - warn-act
  - layoffs
  - layoff-notices
  - amendments
  - revisions
  - change-log
  - point-in-time
  - data-provenance
  - time-series
  - public-records
  - government-data
  - daily-updated
  - labor
  - labor-market
  - united-states
size_categories:
  - n<1K
configs:
  - config_name: changes
    data_files: data/warn_notice_changes.csv
    default: true
  - config_name: suppressed_snapshot_pairs
    data_files: data/suppressed_snapshot_pairs.csv

What changed in US WARN Act layoff notices

376 observed changes across 13 daily builds (2026-08-30 to 2026-09-11), touching 336 distinct notices in the daily US WARN Act dataset.

This is the answer to a question no state labor department will answer for you: what did the layoff filing say before it said this? State WARN portals publish a page. When a notice is amended, re-filed or pulled, the page simply changes - there is no history, no diff, no archive. This dataset is that diff, taken every single day.

Read this before you use a single row

A difference between two of our daily builds has two possible causes, and we usually cannot tell which one it was:

the state agency amended, re-filed or removed the notice, or
our own parser or normalizer got better and now reads the same source differently.

We do not guess. The cause_class column labels only what is mechanically decidable, and every headline number below is computed from value_changed rows only.

change_type=row_absent means the row stopped appearing in our build. It does not mean the employer withdrew the notice. A state re-organising its portal, a paginated page timing out, or a filing ageing out of a source's own window all produce a row_absent row.

Workers whose headcount was substantively revised: 0. An earlier build of this pipeline was about to publish a four-figure "workers revised" total; inspection showed every one of those events was an empty field being populated for the first time, not a number being changed. That is why the number here is computed from value_changed rows alone.

Columns

column	meaning
`observed_date`	the build in which the change was first seen
`prev_observed_date`	the build it is being compared against
`id`	stable notice id, joins to the main dataset
`state`	two-letter state posting the notice
`company`	employer as filed at `observed_date`
`notice_date`	date on the filing
`change_type`	`amended` (a field differs) or `row_absent` (row gone from our build)
`cause_class`	`value_changed`, `field_populated`, `field_cleared`, `format_only`, `unknown`
`field`	which column changed (blank for `row_absent`)
`old_value` / `new_value`	the two values, as printed

What is in here right now

change_type	rows
amended	121
row_absent	255

cause_class	rows
value_changed	87
format_only	24
field_populated	10
field_cleared	0
unknown	255

Fields that were amended:

field	times
`location`	58
`company`	30
`notice_type`	25
`employees_affected`	4
`notice_date`	3
`effective_date`	1

The suppression ledger (5 day-pairs)

Our own coverage window has widened several times (adding states, and opening the free window from 2024+ back to 1988). On a day when the row count swings more than ~5%, treating every missing row as a disappearance would manufacture tens of thousands of fake withdrawals. On those day-pairs we suppress disappearances and publish the pair in suppressed_snapshot_pairs.csv instead, so the gap is visible rather than silent. Amendments are still published for those days.

Why this cannot be copied

Anyone can re-scrape 48 state portals and match our current table within a week. Nobody can reconstruct what those portals said last Tuesday. This file starts at 2026-08-30, grows by one build every day, and cannot be backfilled by us or by anyone else. Every day it is not captured is lost permanently.

Freshness

Rebuilt daily by an automated pipeline (gen_revisions.py) straight from the git history of the free dataset's data/warn_notices.csv. Latest build in this copy: 2026-09-11.

Daily US WARN Act notices - every notice, 48 states, free, no login.
Multi-state employers - who is laying off in more than one state.
Notice-period dataset - how many days' notice each filing gave.
Free CSV/JSON/RSS on GitHub - CC BY 4.0.
The site - state pages, employer pages, live search.
WARN Watch - same-day alert when an employer on your list files, $49/year.

License

CC BY 4.0. Underlying notices are US state public records. Attribution: "WARN Feed, https://approjects-warn-act-notices.static.hf.space".

Not affiliated with the US Department of Labor or any state agency.