ZipLime/congress-trading

US Congress Trading Disclosures

Securities transactions disclosed by members of the U.S. House and Senate under the STOCK Act, normalized into a single schema and rebuilt on a schedule.

The pipeline that produces this dataset lives in recipe/ inside this same repository, at the same revision as the data. Nothing here was assembled by hand — see PIPELINE.md for the full method, including the parts that are still incomplete.

Configs

Config	Rows	What it is
trades	~134k	One row per disclosed transaction. Splits: house, senate.
holdings	growing	Asset snapshots from annual disclosures — Schedule A.
liabilities	growing	Debts from annual disclosures — Schedule D.
features	~108k	Pre-aggregated per (ticker, disclosure day) numeric features.
pit	~178k	The same trades in point-in-time shape, as a Delta table. See below.
filings	~46k	Index of every disclosure filing, with extraction status.
legislators, legislator_terms, committees, committee_members	—	Reference data from unitedstates/congress-legislators.

from datasets import load_dataset

trades = load_dataset("<namespace>/congress-trading", "trades", split="house")

Or straight from Parquet, which keeps the exact dtypes:

import polars as pl
trades = pl.read_parquet("hf://datasets/<namespace>/congress-trading/data/trades/*.parquet")

The one thing to know before backtesting

Use filing_date, not transaction_date and not notification_date, as your point-in-time key. A trade becomes public only when the report reaches the clerk's office. Keying on transaction_date gives your strategy information nobody had at the time — the median PTR is filed 28 days after execution, the 90th percentile 54 days.

notification_date is not the publication date and must not be used as one. On the House form it is the day the filer learned of a trade in a managed account, which precedes the filing: measured against filing_date, keying on it opens positions a median of 13 days early, and 15.6% of rows would trade on the transaction day itself. It is kept in the schema because it means something in its own right, but it is no longer the key.

Three caveats the data itself carries:

• Filter superseded_by IS NULL unless you specifically want amended reports. Senate amendments resubmit the whole report, so originals and amendments otherwise both appear.
• date_quality flags rows the extractor could not vouch for rather than silently repairing them: notify_before_tx (notification earlier than the transaction, impossible), out_of_range, unparsed, and before_tenure — a transaction dated more than two years before the member took office, which is nearly always a misread year digit.
• ticker holds only values that look like exchange tickers (^[A-Z]{1,5}(\.[A-Z])?$). Anything else the form named — CUSIPs, bond stubs like JST-E, fund names, foreign listings — lives in asset_identifier. 4,291 rows across 1,954 distinct values sit there; they were previously in ticker, where they aggregated into features as instruments that do not exist.

Point-in-time view

data/pit/congress_trading.delta carries the same trades in the point-in-time schema described by manifest.json:

system   entity_id, event_date, knowledge_date, knowledge_estimated
values   representative, chamber, transaction_type, asset_ticker,
         amount_low, amount_high, disclosure_lag_days, knowledge_source

knowledge_date is the day the trade became public: filing_date where the document carries one, else notification_date, else the STOCK Act statutory limit of transaction date plus 45 days, with knowledge_estimated set on that last case. knowledge_source names which of the three produced the date, so you can filter on provenance instead of guessing. On the current snapshot the filing date covers 242,414 rows, the notification date the remaining 538, and the statutory fallback fires for none.

The table is append-only: a rebuild adds rows it has not seen and never rewrites or deletes existing ones. Changing what was considered known, after the fact, would break the reproducibility of any backtest built on it.

disclosure_lag_days is the distance between the two dates and is the quickest way to spot bad extractions: PTRs land at a median of 28 days, annual Schedule B rows at 407 (they are filed the following year), and anything past 2,000 days is almost certainly a misread year digit.

Two views of the same data exist because two contracts do. data/trades works with the ziplime release that reads custom bundles through DataBundleSource; data/pit matches the newer point-in-time contract. Neither is a subset of the other: trades keeps all 36 columns including provenance, pit keeps the knowledge-date semantics that trades cannot express.

Two disclosure forms, one trades table

Congress discloses trades twice, and both are in trades, distinguished by source_form:

• ptr — a Periodic Transaction Report, filed within 30–45 days of the trade.
• annual_b — Schedule B of the annual disclosure, filed the following May–August, listing every transaction of the year.

Most annual rows restate a trade already filed as a PTR; those carry superseded_by pointing at the PTR row, which was disclosed earlier and is therefore the correct point-in-time entry. The rows that remain are trades with no PTR at all — operations you cannot learn about any other way. Filtering superseded_by IS NULL keeps exactly one row per real trade, at its earliest public date.

Do not pool the two forms in one signal without accounting for the difference. They are not two sources of the same event; they differ in how fast the news travels and in how much of the row is filled in:

	ptr	annual_b
rows (current, superseded_by IS NULL)	130,700	112,278
median disclosure lag	28 days	407 days
90th-percentile lag	54 days	647 days
ticker populated	85.1%	56.4%
min_amount_usd populated	99.5%	98.8%
transaction_type populated	100%	100%

An annual_b row is public more than a year after the trade, and it is missing a ticker four times as often. A momentum or event study that mixes them measures two different things at once and will read as far weaker than the PTR signal alone. Either filter to source_form = 'ptr', or split the analysis by form. The features config keys on knowledge_date, so it places each row at its own public date correctly — but a horizon shorter than a year still sees almost only PTRs, and one longer than a year sees a mixture.

Why holdings matter

trades alone cannot tell you position size. A $15,000 sale by a member with a $50,000 portfolio and by one with a $5,000,000 portfolio are different events. holdings is the Schedule A snapshot — every asset the member reported, with its value band, income type, and income band, as of the annual filing date.

Both value and income are bands, not exact figures, and value_min_usd is null where the form itself prints None — divested assets and closed accounts are listed without a value.

Schema — trades

The first 29 columns match ziplime's congressional-data contract exactly, so this dataset is a drop-in source for it.

Column	Type	Notes
id	str	Stable primary key, {chamber}_{doc_id}_{n}.
chamber	str	house | senate
bioguide_id	str	Resolved against congress-legislators; null for 0.3% of rows.
member_name, member_first_name, member_last_name	str
party, state, state_district	str	As of the transaction date, not the member's current affiliation.
owner	str	self | spouse | joint | child
ticker	str	Only values matching ^[A-Z]{1,5}(\.[A-Z])?$; null where the filing names no ticker. See ticker_source and asset_identifier.
asset_name, asset_category	str	stock | option | bond | municipal | futures | crypto | non_public | other
transaction_type	str	purchase | sale_full | sale_partial | exchange | exercise | redemption | distribution | other
transaction_date, notification_date	date	notification_date is when the filer learned of the trade, not when it became public.
min_amount_usd, max_amount_usd	i64	Official disclosure bands. max is null for the open-ended top band.
is_option, option_type, option_quantity, strike_price, expiration_date, days_to_expiration	mixed	Populated for option trades only.
description, doc_id, report_url, filing_year, inserted_at	mixed

Provenance and quality columns added on top of the ziplime contract:

Column	Values	Meaning
filing_date	date	Day the report reached the clerk's office — the point-in-time key.
filing_id	str	Join key to the filings config.
asset_identifier	str | null	The raw identifier when it is not a ticker: CUSIP, bond stub, fund name, foreign listing.
extractor	pdf-inspector+rules, legacy-openrouter, legacy-pandas-read-html	How the row was extracted.
confidence	f64	1.0 for the deterministic parser.
superseded_by	str | null	id of the amendment that replaces this row.
amount_quality	ok | snapped | invalid	snapped = coerced onto an official band.
date_quality	ok | out_of_range | unparsed | notify_before_tx | before_tenure	before_tenure = dated >2 years before the member took office.
ticker_source	disclosed | name_lookup | null	name_lookup = recovered from the asset name via in-dataset evidence, never guessed.
source_form	ptr | annual_b	Which form disclosed the trade.

Schema — features

Aggregated by (ticker, disclosure day). All numeric columns are Float64.

Daily columns: n_disclosures, n_purchases, n_sales, n_members, buy_notional_usd, sell_notional_usd, net_notional_usd, max_single_notional_usd, n_option_trades, median_disclosure_lag_days.

Trailing-window columns: n_disclosures_30d, net_notional_usd_30d, n_members_30d, and the same three at _90d.

Read the window columns at any frequency coarser than daily. ziplime downsamples custom bundles with .last() per column, so a daily count read monthly reports only the final day of the month. A trailing 30-day window read monthly is still a valid 30-day window.

Notional is the midpoint of the disclosed band, or the lower bound for the open-ended top band — a single "Over $50,000,000" row would otherwise dominate any sum.

Using it with ziplime

from ziplime.data.data_sources.huggingface_congress_data_source import (
    HuggingFaceCongressDataSource,
)

source = HuggingFaceCongressDataSource.from_env(dataset="features")

See examples/ingest_data_huggingface_congress.py in ziplime for a full ingest. Note that ingestion resolves tickers against ziplime's asset database: of the ~9,800 distinct tickers here, about 3,300 resolve, covering ~83% of rows. The rest are bonds identified by CUSIP, delisted names, and foreign listings.

Coverage and known gaps

	Covered	Gap
House PTRs	2013 – present, 8,353 filings indexed	~12% of recent filings are scanned paper; share was 60% in 2014
Senate PTRs	2013 – present, 1,194 electronic filings	491 paper filings not yet extracted (29%), and nearly all of 2012–2013
House annual disclosures	Schedules A, B and D extracted	Schedules C and E–J (earned income, positions, gifts, travel) not extracted

Other things worth knowing:

• The House Clerk rewrites historical catalogs. Between April and August 2026, 4,305 filings vanished from the 2012 catalog — all pre-STOCK-Act annual reports. The filings config keeps them, with last_seen_at frozen at the run that last saw them.
• 4,094 House rows and 1,690 Senate rows have no ticker. The previous pipeline filled this field by asking a model to "infer from context", which produced values like US_TREASURY, MUNI and N/A. Those are not tickers, so this build leaves the field null instead.
• Disclosure bands are ranges. There is no exact position size or P&L in this data, and there never will be.

Schema — holdings

id, chamber, bioguide_id, member name fields, party, state, state_district, owner, ticker, asset_name, asset_category, filing_year, as_of_date, value_min_usd, value_max_usd, income_type, income_min_usd, income_max_usd, tx_over_1000, description, doc_id, report_url, inserted_at, filing_id, extractor, confidence, value_quality, ticker_source.

as_of_date is the filing date — the day the snapshot became public. filing_year is the year the report covers, which is the year before.

Schema — liabilities

id, chamber, bioguide_id, member_name, owner, creditor, liability_type, date_incurred, filing_year, as_of_date, min_amount_usd, max_amount_usd, description, doc_id, report_url, inserted_at, filing_id, extractor.

Liability bands are not the same as asset bands: they start at $10,001 and have no $1–$1,000 or $1,001–$15,000 step.

How the extraction was verified

recipe/audit_extraction.py checks the bundle against its own sources: every value must appear literally in the PDF it was extracted from. On a 200-document sample, transaction and notification dates match 2,351/2,351; every non-null amount bound and ticker matches; holdings value bounds match 6,785/6,785; and asset names match 6,889/6,897 (99.88%). Bundle-wide invariants are clean: no amounts off the official bands, no duplicate ids, no inverted ranges.

Where the deterministic parser and the previous LLM extraction overlap (5,699 documents), the parser yields 53,895 rows against the LLM's 53,415, agreeing exactly on row count in 97% of documents.

The residual ~0.1% of holdings whose name does not match is caused by two consecutive assets merging when the form breaks a row in an unusual place. Those rows carry correct amounts and dates; only the name is a concatenation.

Licence and terms

House financial disclosures are U.S. Government works in the public domain. Reference data from congress-legislators is CC0.

Senate disclosures are obtained through efdsearch.senate.gov, whose access agreement prohibits use of the reports for "commercial purpose" other than by news media, among other restrictions under the Ethics in Government Act. Consider whether your use falls within those terms.

This dataset is provided for research. It is not investment advice, and the extraction is imperfect in the ways documented above.