ZipLime/us-government-contracts

US Federal Contract Awards

Federal contract actions paid to publicly traded companies — and, for each one, the day the public could first see it, which for the Department of Defense is three months after the contract was signed.

13 372 892 contract transactions · 10 533 005 awards · 1 703 listed companies · FY2005 to today · $4.45tn obligated

69% of those rows are the Department of Defense and the Army Corps of Engineers, and not one of them is dated less than ninety-two days after the contract was signed.

The pipeline lives in recipe/ at the same revision as the data. See PIPELINE.md for the method and its limits.

The problem this dataset exists to solve

The Army signs a contract with Lockheed Martin on 3 March. FPDS has the record the same afternoon — the Pentagon reports its contract actions immediately.

And then FPDS does not publish it for ninety days.

"Publication of Department of Defense (DOD) and U.S. Army Corps of Engineers (USACE) data in FPDS is delayed 90 days." — USAspending.gov

Here is what that does to a row:

W58RGZ26C0001  Lockheed Martin Corporation     Department of the Army
  action_date              2026-03-03
  initial_report_date      2026-03-03 14:02 UTC    ← in FPDS the same day
  knowledge_date           2026-06-03 22:00 UTC    ← public three months later
  knowledge_date_method    dod_publication_delay
  publication_lag_days     92

A dataset keyed on action_date makes that award visible on 3 March. A dataset that reads initial_report_date as a publication time makes it visible on 3 March too — and that is the subtler trap, because the field looks exactly like the answer. Measured over a week of DOD contract actions, initial_report_date sits a median of zero days after the action date.

Either way a backtest buys the defense sector a quarter before the market could have known, and calls the result alpha.

So every row here is dated on knowledge_date, and action_date is carried alongside for reference and used for nothing.

How each row is dated

knowledge_date_method names the evidence, and it is on every row.

Method	What it rests on	Estimated
collector_first_seen	the daily sweep watched this record appear	no
fpds_initial_report	the FPDS report instant the source states, plus the two days USAspending takes to publish	no
dod_publication_delay	ninety days of statutory silence, then the same two days	yes
reporting_deadline_rule	three federal business days to report, plus two to publish	yes

The rules are USAspending's own, quoted in metadata/sources.yaml:

"Procurement policy allows contracting officers three business days to submit a contract action report to FPDS; after they do so, those data are made available to USAspending.gov on the following morning and automatically published to the website the day after that."

with a thirty-day allowance instead of three business days where the award was made under FAR 6.302-2 or FAR subpart 18.2 — unusual and compelling urgency, and emergency acquisitions — which the competition code identifies row by row.

The ninety-day window follows the agency that reports the action, not the one whose money it is. Measured against the live API over the thirty days to 14 September 2026: 345 DOD-awarded contract actions were visible, where an ordinary month carries hundreds of thousands — and 497 GSA-awarded actions funded by DOD were visible over the same window, on the ordinary civilian schedule. So is_dod reads the awarding agency, and funding_agency_name is published beside it for anyone who wants to slice by whose appropriation it was.

USAspending states the day a record reaches the site and never the minute, so the modelled instant is stamped at 18:00 Eastern, after the New York close. The first session in which a row can be traded is the next one. The model is allowed to be late; it is never allowed to be early.

Records that were public before the action they describe

About one row in sixty-five reaches FPDS before its own action date, and a few hundred of those are dated years forward — an option or a delivery order whose action date is set well ahead of the paperwork. publication_lag_days goes negative for them, and they are kept as they are: the record genuinely was readable on that date, and rewriting it would be inventing a publication instant to make a column tidy.

They are a tail — 1.5% of rows have any negative lag, and 0.02% cross a mid-year boundary because of it — and they can never occur on the delayed half of the dataset, where publication is floored at ninety days after the action. A reader who would rather not have them writes publication_lag_days >= 0.

knowledge_estimated is narrower than it looks, and it is worth being precise about. It says whether the FPDS report instant was read or ruled — not that the publication instant is exact. Except where the collector watched the record arrive, the two days between FPDS and the website, and the hour of the day, are always modelled. There is no column here called exact_publication_timestamp, because USAspending does not publish one.

What the collector adds

Rules are what is left when nobody watched. Every weekday morning a scheduled job asks USAspending for every contract record whose FPDS entry changed in the last four days — filtered on last_modified_date, not on the action date, because a DOD action signed three months ago becomes visible today, and that is the fact worth writing down.

It is one request for the whole federal government, and what is written down is narrower: the recipients this dataset publishes. A ledger of every contract action in the country would grow by a gigabyte a year to record first sightings of rows nothing here joins to. A recipient that becomes resolvable later starts being observed from that day and keeps the modelled rule for its earlier rows.

metadata/first_seen.parquet keeps one row per transaction: the identifier, the content hash, when it was first seen and when it was last seen. first_seen_at is written once and never revised — revising it would be rewriting when the market could have known something. From then on that row's knowledge_date is a measurement rather than a rule, and knowledge_date_method says so.

When a published record is later corrected, metadata/revisions.parquet gains a row carrying the superseded hash. A correction never erases what was public before it.

Identity: CIK, not ticker, and not the recipient either

A contract recipient is a legal entity. A strategy trades a listed parent. These are not the same thing, and the gap between them is where a dataset like this is most easily made to lie.

LOCKHEED MARTIN CORPORATION and LOCKHEED MARTIN CORP are two UEIs of one company. SIKORSKY AIRCRAFT CORPORATION was United Technologies' until 6 November 2015 and Lockheed Martin's after it. RAYTHEON COMPANY was its own SEC filer until April 2020 and part of RTX after.

Every link in entity_map therefore carries three things:

• provenance — match_type and a sentence of evidence;
• a window — valid_from and valid_to, applied against the action date;
• a confidence, and a rule that the pipeline never promotes a name similarity to canonical.

Recipient UEIs resolved to a listed company	2 908
Listed companies behind them	1 703
Links carrying a validity window	3 915 of 4 149
From an exact SEC filer name	272
From the curated subsidiary table	3 877
From a name similarity	0

Measured against the whole recipient index — 92 842 awardee entities, the P parent level excluded so nothing is counted twice — 39.5% of federal contract obligations resolve to a listed parent. The rest is privately held, and the largest of it is named in the unmatched report rather than glossed over.

The listed side comes from ZipLime/security-master, which knows what a filer was called and which symbol it carried on a date, rather than what it is called now.

What is deliberately left unmatched

Joint ventures and national-laboratory operating companies — Triad National Security, UT-Battelle, Lawrence Livermore National Security, Consolidated Nuclear Security, Savannah River Nuclear Solutions — are consortia whose economics do not flow to a single listed parent. Assigning them to one member would be inventing a fact. They are among the largest recipients in the country, they stay unmatched, and they sit at the top of metadata/top-unmatched-recipients.parquet, which is both the honest counterweight to the coverage claim and the work queue for metadata/public_parents.yaml.

Universities, non-profits and privately held contractors — TriWest Healthcare Alliance, Atlantic Diving Supply, Bechtel, Deloitte, Peraton — are unmatched for the plainer reason that there is nothing to match them to.

Three amounts that are not interchangeable

federal_action_obligation   money this action actually placed on the contract
award_current_value         the contract's value as of this action
award_potential_value       the ceiling, which nobody has promised to spend

An indefinite delivery vehicle can carry a ten-billion-dollar ceiling and never be ordered against, and a press release quoting that number is quoting a ceiling. is_idv, is_task_order and has_parent_award are published so the structure is visible, and the feature layer is built on obligation flow alone — the ceiling is in the event table for anyone who wants it and is deliberately not made into a feature.

Negative obligations are kept with their sign. A deobligation is money coming back off a contract — a cancellation, a scope reduction, an unused option — and a pipeline that took absolute values would turn cancellations into wins. The quality gate fails a build that has none.

One award, many transactions

A contract with forty modifications is one award and forty rows. Treating every positive obligation as a new contract turns federal procurement into a signal that is mostly option exercises and funding increments.

is_new_award is decided per award across the whole history — the earliest action, with the earliest knowledge instant breaking ties — not within a fiscal year, because an award signed in one year and modified in another would otherwise produce two "new" contracts.

The awards table is a roll-up and current state. It is rewritten whenever a modification arrives, and a historical strategy that reads it is reading the present. transactions and pit are what history is for.

Point-in-time usage

Everything a strategy may know at an instant:

import polars as pl

AS_OF = pl.datetime(2026, 3, 31, 20, 0, 0, time_zone="UTC")   # 16:00 Eastern

known = (
    pl.scan_parquet("data/company_contract_events/**/*.parquet")
      .filter(pl.col("knowledge_date") <= AS_OF)
)

Contract acceleration, ranked:

panel = pl.scan_parquet("data/features/*.parquet")

accelerating = (
    panel.filter(
        (pl.col("as_of_date") == pl.date(2026, 3, 31))
        & (pl.col("contract_obligation_90d_change_yoy") > 0.5)
        & (pl.col("contract_obligation_365d") > 1e8)
    )
    .sort("contract_obligation_90d_change_yoy", descending=True)
)

Defense exposure:

panel.filter(pl.col("dod_obligation_share_365d") > 0.5)

Concentration and growth, and what they are measured against:

# dod_obligation_share_365d, top_agency_share_365d and top_psc_share_365d
# divide money placed by money placed, so they are in [0, 1]. A share of the
# *net* year would exceed 1 whenever deobligations offset it.
panel.filter(pl.col("dod_obligation_share_365d") > 0.5)

# Growth is null where the prior year came to less than $1,000 -- a base that
# small is an administrative action, not a business. Above it, growth is
# unbounded and honestly so.
panel.filter(pl.col("contract_obligation_365d_change_yoy") > 0.5)

Only instants that were measured, discarding the modelled ones:

known.filter(~pl.col("knowledge_estimated"))

What not to do:

# Wrong: action_date makes a DOD award visible three months before FPDS
# published it.
pl.col("action_date") <= AS_OF.dt.date()

# Wrong: the FPDS report instant is when the Pentagon filed the record, not
# when anyone could read it.
pl.col("initial_report_date") <= AS_OF

Tables

Table	Rows
company_contract_events	11 057 926	One contract transaction linked to a listed company — the table most readers want
transactions	13 372 892	One federal contract transaction, partitioned by fiscal year
awards	10 533 005	One prime award, rolled up — current state, not history
pit	13 372 892	The two dates, how the second was decided, and the content hash
recipients	181 945	Every contract recipient the source knows, matched or not
entity_map	4 149	UEI joined to a listed parent, with evidence and window
features	189 061	One company, one month end, computed only from what was knowable
psc	2 918	Product and service codes
naics	1 741	Industry codes
agencies	111	Top-tier federal agencies

Where the knowledge instants come from

knowledge_date_method	Transactions	Median lag	Obligations
dod_publication_delay	9 260 879	92 days	$3.40tn	ninety days of statutory silence, then two more to reach the site
fpds_initial_report	2 558 125	2 days	$861.8bn	the instant FPDS states, plus the two days to publish
reporting_deadline_rule	1 524 776	7 days	$175.2bn	no usable report instant; the allowances applied to the action date
collector_first_seen	29 112	5 days	$16.9bn	the daily sweep watched the record appear

The ten largest recipients

Ticker	Company	Obligations	Transactions
LMT	LOCKHEED MARTIN CORP	$877.6bn	841 915
BA	BOEING CO	$430.9bn	250 447
GD	GENERAL DYNAMICS CORP	$381.3bn	258 240
NOC	NORTHROP GRUMMAN CORP	$246.6bn	217 751
RTN	RAYTHEON CO/	$178.0bn	120 904
MCK	MCKESSON CORP	$150.8bn	564 053
HON	HONEYWELL INTERNATIONAL INC	$117.8bn	137 957
UNH	UNITEDHEALTH GROUP INC	$115.8bn	3 673
HUM	HUMANA INC	$104.5bn	5 293
HII	HUNTINGTON INGALLS INDUSTRIES, INC.	$98.9bn	32 557

The transaction counts in that table vary by two orders of magnitude for companies of similar size, and that is the source, not a defect: military health contracts are a handful of very large awards, and pharmaceutical distribution is a very large number of small ones. A signal built on transaction counts rather than obligations will find only the second kind.

Coverage

Fiscal year	Transactions	Awards	Recipients	Obligations	Stated FPDS instant
FY2005	294 546	231 796	1 150	$100.4bn	37%
FY2006	402 830	342 511	1 210	$114.2bn	20%
FY2007	473 264	402 176	1 265	$135.1bn	22%
FY2008	524 357	446 615	1 309	$148.5bn	22%
FY2009	423 116	338 858	1 340	$160.7bn	28%
FY2010	389 119	296 702	1 394	$170.9bn	30%
FY2011	396 347	296 978	1 416	$172.2bn	32%
FY2012	367 381	273 681	1 450	$169.0bn	38%
FY2013	309 373	219 258	1 456	$163.9bn	45%
FY2014	326 255	239 015	1 498	$147.6bn	50%
FY2015	723 921	636 029	1 543	$153.0bn	70%
FY2016	718 100	626 054	1 555	$179.2bn	100%
FY2017	654 973	557 018	1 651	$189.6bn	100%
FY2018	499 587	405 822	1 656	$210.1bn	100%
FY2019	591 994	504 232	1 701	$229.2bn	100%
FY2020	858 350	761 326	1 795	$268.5bn	100%
FY2021	919 229	824 471	1 818	$251.5bn	100%
FY2022	764 335	665 554	1 843	$273.6bn	100%
FY2023	877 753	780 924	1 899	$301.5bn	100%
FY2024	744 478	640 832	1 968	$285.3bn	100%
FY2025	1 172 857	1 066 957	2 019	$333.7bn	100%
FY2026	937 346	854 994	2 182	$291.9bn	100%
FY2027	3 381	3 338	281	$455m	100%

The newest fiscal year is short because it is not over. metadata/coverage.parquet carries the same table with field-completeness shares, and metadata/publication-lag-stats.parquet breaks the lag down by year, agency and award type — the Department of Defense rows should sit at ninety-two days, and anything else there is a bug.

The table above is generated from the data on every build, so it says exactly which years are in and no more. The backfill runs newest-first and stops at FY2008: the recipient-filtered endpoint refuses any date before 1 October 2007, and reaching FY2001 means pulling whole-universe files instead — a different and much heavier route.

What this dataset is not

• Not corporate revenue. An obligation is money placed on a contract, not revenue recognised. The performance period can run for years.
• Not a backlog. award_potential_value is a ceiling nobody has promised to spend.
• Not complete procurement. Classified programmes are not in FPDS, and micro-purchases below the reporting threshold were not reported at all in the early years.
• Not same-day award news. That is the whole point: the earliest civilian rows become knowable about five days after the action, and DOD rows about ninety-two.
• Not evidence that the listed parent receives the money. A subsidiary's award is the subsidiary's; how much reaches the parent's income statement is a question this dataset does not answer.
• Not subawards. Prime contracts only. Subcontract reporting has different semantics and a different recipient graph, and mixing the two into one fact table would make both untrustworthy.
• Not grants, loans or other financial assistance. Contracts only.

Provenance and rights

Source: USAspending.gov, U.S. Department of the Treasury. Original data source: the Federal Procurement Data System (FPDS) and the federal agencies that report contract actions to it.

Federal procurement records are produced by the United States government in the course of awarding contracts and are not subject to copyright protection in the United States (17 U.S.C. 105). USAspending publishes them in bulk expressly so that they can be reused. Every row keeps its transaction and award identifiers and a source_url, so any value can be traced to the official record. See metadata/source_licenses.yaml.

Neither the Treasury nor any reporting agency endorses, certifies or verifies this dataset. No commercial procurement-data vendor is a source for it, and no vendor's licensing text is inherited.

Updates

Job	Cron (UTC)
us-government-contracts-update	0 11 * * 1-5	record what became public, fold it in, publish
us-government-contracts-refresh	0 3 * * 6	re-read the recent fiscal years in full

The daily update is one request, not a hundred and forty-six: the sweep already asks USAspending for every contract record whose FPDS entry changed in the last four days, so the published years are folded forward from that download rather than re-fetched. It is also what turns knowledge_date from a rule into a measurement, one weekday at a time.

The Saturday refresh re-reads the recent fiscal years in full, because a modification feed only reports what the source says changed.

The historical backfill runs as one-off jobs, newest first: FY2020–FY2026, then FY2010–FY2019, then FY2008–FY2009. FY2001–FY2007 is a separate problem — the recipient-filtered endpoint refuses any date before 1 October 2007, and those years have to come from whole-universe files instead.

Building it yourself

pip install -e ".[publish]"
export USASPENDING_USER_AGENT="Your Project you@example.com"

python -m recipe.cli build --from-year 2016 --to-year 2026   # the backfill
python -m recipe.cli update --days 4      # the daily path: sweep, fold in, derive
python -m recipe.cli derive               # rebuild the derived tables only
python -m recipe.cli verify               # the quality gate