ZipLime/insider-trading

ZipLime US Insider Trading Disclosures (PIT)

Use knowledge_date, not transaction_date, when backtesting. transaction_date says when a trade occurred; knowledge_date says when the filing became observable through EDGAR. Using the former as the signal date introduces look-ahead bias.

visible = trades.filter(pl.col("knowledge_date") <= simulation_time)

This dataset normalizes corporate-insider transactions reported in U.S. SEC EDGAR Form 4 and Form 4/A ownership filings. It includes non-derivative (Table I) and derivative (Table II) rows, source provenance, reporting-person roles, amendment lineage, deterministic economic classifications, and point-in-time (PIT) features.

The official source is U.S. SEC EDGAR. The deterministic recipe lives in recipe/ in the same repository and revision as the published data. See PIPELINE.md for the collection, validation, amendment, and failure contracts.

Configs

Config	Rows	Grain	Purpose
transactions	9,873,330	One disclosed Table I/II transaction row per filing revision	Normalized source facts, provenance, classifications, quality flags, and PIT dates.
filings	4,034,227	One accession number	Processing ledger for Form 4 and 4/A.
owners	4,351,105	One (accession_number, reporting_owner) relation	Owner names, CIKs, roles, titles, and joint-filing status without multiplying transaction notional.
features	6,617,994	One (issuer/ticker, knowledge_day) row	PIT-safe daily and trailing 7/30/90-day aggregates, including cluster-buying inputs.
pit	9,873,330	One knowledge event/revision	The point-in-time view a backtester mounts, published as Parquet and as a Delta table. See below.

Covering 2006-01-03 to 2026-06-30 without a gap: 17,996 issuers, 201,812 reporting persons, 23,304 distinct tickers, and 99,404 amendments.

Which one to read. Simulate against pit — it carries the amendment lineage and one row per knowledge event, which is what an as-of query needs. Take ready-made signals from features. Reach for transactions when you want the source facts and their provenance, and for filings or owners to answer who filed what. The owners config partitions by ingest date, not knowledge date, because a reporting-owner relation is a property of a filing rather than an event of its own.

from datasets import load_dataset

# `pit` is the one to simulate against; see "Which one to read" above.
events = load_dataset("ZipLime/insider-trading", "pit", split="train")

Polars preserves the Parquet dtypes directly:

import polars as pl

events = pl.read_parquet(
    "hf://datasets/ZipLime/insider-trading/data/pit/knowledge_year=*/*.parquet"
)

The pit directory also holds the Delta table a ziplime bundle mounts, so read its year partitions rather than globbing the directory whole.

Pin a Hugging Face revision= (or a repository commit) for a reproducible research result.

Point-in-time semantics

The core contract is:

event_date      = transaction_date
knowledge_date  = accepted_at (UTC)
visible at T    = knowledge_date <= T

accepted_at is the EDGAR acceptance timestamp. It is an official, reproducible proxy for public availability, not an exact sec.gov first-availability timestamp. SEC says filings are often available on sec.gov within 1–3 minutes of the EDGAR timestamp, while ownership-form submissions begun after the 22:00 ET cutoff may be disseminated the next business day. Intraday or latency-sensitive studies should apply a conservative policy suitable for their use; the core dataset does not invent an availability timestamp that SEC does not publish.

When the acceptance timestamp cannot be recovered, the dataset uses a documented conservative fallback and records both:

• knowledge_date_source: edgar_acceptance_datetime, filing_date_eod_fallback, or first_seen_fallback;
• knowledge_estimated: true whenever the timestamp is not the EDGAR acceptance datetime.

filing_date_eod_fallback means 23:59:59.999999 in America/New_York converted to UTC. If even the filing date is absent, the UTC first_seen_at timestamp is used. Both choices are intentionally visible and conservative.

Point-in-time view

data/pit/insider_trading.delta carries the same events as the pit Parquet config, in the Delta table a ziplime bundle mounts. manifest.json describes it: bundle_storage_data.table_uri points at the table, and schema lists its columns.

system   entity_id, event_date, knowledge_date, knowledge_estimated
values   issuer_cik, ticker, insider_name, insider_cik, primary_role,
         transaction_code, economic_type, signal_type, shares, price,
         transaction_value, shares_owned_after, holding_change_pct,
         disclosure_lag_days, is_10b5_1, is_derivative,
         logical_transaction_id, revision, amendment_status

entity_id is the issuer CIK, which is the stable identity across ticker changes. Delta has no list type that round-trips cleanly across readers, so reporting_owner_ciks and reporting_owner_names travel as |-delimited strings in this table; the Parquet config keeps them as lists.

The table is rewritten in full by each rebuild, so it carries one version rather than the history of every run. Pin a repository revision, not a Delta version, for a reproducible backtest.

Do not replace a historical CIK/ticker with today's ticker. issuer_cik is the stable issuer identity; ticker is the symbol disclosed or mapped using only information available at that point in time. ticker_source makes that choice auditable.

Amendments (Form 4/A)

Amendments are new knowledge, not permission to rewrite history. The raw and PIT layers are append-only: an original Form 4 remains visible at its original knowledge time, while a later Form 4/A becomes visible only at its own knowledge_date. Form 4/A XML exposes dateOfOriginalSubmission, but does not identify the original accession number and often contains only changed rows; the dataset therefore never assumes that an amendment is a full replacement.

Transaction lineage is represented with logical_transaction_id, revision, supersedes_transaction_id, superseded_by, and amendment_status. Possible statuses include original, amended, added_by_amendment, removed_by_amendment, unchanged, and unresolved. Reconciliation is conservative: an ambiguous amendment stays unresolved instead of being linked by guesswork. The authoritative raw lineage points backward from the later row with supersedes_transaction_id; superseded_by remains null in append-only raw history so future knowledge is not written into an older row. Omission from a partial Form 4/A is never treated as a removal.

For an as-of query, first restrict all knowledge events to the simulation timestamp and only then resolve revisions. Filtering against amendment links from a future snapshot can itself introduce look-ahead bias.

What Form 4 contains

The dataset retains every transaction code, not only purchases and sales.

SEC code/designation	Typical meaning	Default economic class
P	Open-market or private purchase	open_market_purchase
S	Open-market or private sale	open_market_sale
A	Grant or award	grant_or_award
M	Exercise or conversion of a derivative	option_exercise
F	Tax withholding/payment with securities	tax_withholding
G	Gift	gift
D	Disposition back to issuer	issuer_disposition
K indicator	Equity swap or similar transaction	equity_swap
J, V indicator, other/unknown	Other/voluntarily reported transaction	other

The raw transaction_code is never discarded. In structured ownership XML, the visual K designation is normally carried by equitySwapInvolved alongside the main transaction code, and voluntary-reporting status is expressed through form/timeliness fields rather than a literal V code. economic_type and signal_type are deterministic, versioned derived fields. In particular, grants, option exercises, tax withholding, and gifts are not silently treated as discretionary insider purchases.

The XML aff10b5One value is a document-level Form 4/5 field available from 2023-04-01. Row-level is_10b5_1 = true is assigned only when a transaction has specific footnote evidence or a flagged filing contains exactly one transaction. Rows in a mixed flagged filing remain null when attribution is unclear. false means the filing-level checkbox reports false; it is not, by itself, proof that a transaction was discretionary, including for some older plans. null means the historical form did not expose the field or row-level attribution is unknown. A plan adoption date is not a structured transaction field and is populated only when it can be reliably recovered from explanation/footnote text. Table II rows are retained, but derivative rows are excluded from the default insider-buy signal.

Schemas

The canonical, executable schemas are in recipe/schema.py. The main field groups are summarized below.

transactions

• Identity/provenance: id, transaction_id, logical_transaction_id, accession_number, filing_transaction_index, table_number, source_url, ownership_xml_url, footnote_ids, parser_version, inserted_at.
• Issuer: issuer_cik, issuer_name, ticker_raw, ticker, ticker_source, security_title.
• Reporting person: insider_cik, insider_name, SEC role flags, officer_title, primary_role, and reporting-owner arrays for joint filings.
• Source facts: transaction_date, deemed_execution_date, transaction_code, acquired_disposed, shares, price, shares_owned_after, and ownership fields.
• Derivatives: security title, exercise/conversion price, underlying security and shares, exercise/expiration dates, price, and post-transaction holdings.
• 10b5-1: source filing_aff10b5_one, conservative row-level is_10b5_1, and footnote-supported plan_adoption_date when available.
• Derived/quality: transaction_value, economic_type, signal_type, holding_change_pct, disclosure_lag_days, date_quality, price_quality, and quality_flags.
• PIT/amendments: event_date, knowledge_date, knowledge_estimated, knowledge_date_source, revision, and lineage/status fields.

transaction_value is calculated only when both reported shares and reported price exist. Missing prices are never filled with market prices. Weighted-average prices remain reported averages; footnotes are not expanded into fictional exact executions.

Filers occasionally mistype a price or share count, producing a notional larger than any real transaction. Those rows keep their reported values verbatim and are marked instead: price_quality = "implausible", plus implausible_price and/or implausible_transaction_value in quality_flags. The thresholds flag only what is structurally impossible — a price above $2,000,000 per share or a notional above $1e12 — because the dataset holds no market data and must not second-guess a price it cannot check. Filter on quality_flags before summing notional.

filings

Contains accession/form identity, filing and acceptance dates, issuer fields, amendment pointers, document-level filing_aff10b5_one, source URLs, first/last seen timestamps, parser/XML versions, parse_status, parse_error, and sec_presence_status. A filing that later disappears from an SEC index is retained because it was historically observable.

owners

Contains one relation per reporting owner: accession, owner CIK/name, SEC role flags, title/other text, is_joint_filing, and insertion timestamp. This table supports many-to-many analysis without double-counting transaction value.

features

Daily fields include transaction/buyer/seller counts, purchase/sale/net notional, CEO/CFO/director purchase counts, 10b5-1 and derivative counts, median disclosure lag, and maximum holding change. Trailing fields cover 7/30/90-day purchases, notional, unique buyers, CEO/CFO activity, and a 14-day cluster-buying window. All windows are calculated by knowledge time.

knowledge_day is the America/New_York calendar day of the underlying knowledge events. A daily feature row is exposed at feature_available_at (also used as its knowledge_date): midnight America/New_York at the start of the following calendar day, converted to UTC. input_max_knowledge_date makes the no-future-input invariant auditable.

is_cluster_buy uses the documented baseline of at least three distinct insiders making open-market purchases in a trailing 14-calendar-day window; the component counts are published so researchers can choose another rule.

pit

The compact PIT view begins with pit_event_id, operation, entity_id, event_date, knowledge_date, and knowledge_estimated, followed by issuer/ticker, insider/role, transaction, notional/position, 10b5-1/derivative, source identity, and amendment fields. operation is UPSERT, RETRACT, or UNRESOLVED; unresolved amendment rows are excluded from the default corrected as-of view unless explicitly requested.

Coverage and known gaps

	Covered	Gap
Form 4 / 4-A	2006-01-03 to 2026-06-30, 4,034,227 filings, no missing month	2003-06-30 to 2005-12-31 is not covered: SEC's quarterly extracts begin at 2006 Q1
Acceptance timestamps	99.9996% of transaction rows carry a real EDGAR acceptance time	41 rows fall back to the end of the filing day
Parsing	4,034,227 filings, zero parse failures	—

Electronic filing of Forms 3/4/5 became mandatory on 2003-06-30 under SEC Release 33-8230. The start date here is set by the source rather than by the rule: this build is loaded from SEC's quarterly Form 3/4/5 extracts, which begin at 2006 Q1.

Values are stated to two decimals. The extracts round share counts and prices, where the filing itself may state more precision. Measured on one quarter, that touches about 22% of rows on price and 4% on share count. The values themselves agree with the filings; only the trailing digits are lost.

About 7% of rows carry fewer footnote references than the filing does. The extracts expose footnote citations per field and do not reproduce every link the XML carries. The visible consequence is price_quality: rows a full parse would label reported_average may appear as footnote_referenced instead.

Other things worth knowing:

• Scope is Form 4 and Form 4/A only. Forms 3 and 5 travel in the same extracts and are deliberately not included: they report holdings and deferred transactions, which this schema does not model.
• 3,070,384 transaction rows (31.1%) have no reported price, and 54,765 (0.55%) have no ticker. Both are properties of the filings: a grant or gift often carries no price, and an issuer without a disclosed trading symbol has no ticker to record. Neither is filled in by inference.
• 2,708,865 rows (27.4%) are derivative (Table II). They are retained but excluded from the default insider-buy signal. Their price and post-transaction holding live in derivative_price and derivative_owned_after; price and shares_owned_after belong to Table I and stay empty on those rows.
• The owners config carries no knowledge timestamp of its own, so its knowledge_year= directory reflects when the rows were ingested, not when the relations became public. Join to filings or transactions on accession_number for a point-in-time view of who reported.
• A filing can omit ticker, price, owner CIK, transaction date, or other facts. Missing source facts stay null; they are not guessed by an LLM.
• holding_change_pct is only meaningful where the before/after-position logic is unambiguous, which is 26.0% of rows. Read holding_change_quality with it: 41.2% of rows are multiple_same_day_transactions (the position moved several times that day and no single row owns the change), 27.4% are derivative, 3.8% ambiguous, 1.5% zero_previous_position.
• Suspicious dates and price semantics are retained with quality flags rather than silently corrected. 4,965 rows coded P are not marked as an acquisition and 6,050 coded S are not marked as a disposition; these are what the filers submitted.
• 130 rows carry a transaction date before 1990 — a filer typing a two-digit year into a four-digit field, so 2012 becomes 0012. They are flagged date_quality = "out_of_range" and kept as filed. Filter on date_quality before computing anything from disclosure_lag_days.
• superseded_by is always null, deliberately. Lineage points backwards only: the later row names what it supersedes through supersedes_transaction_id. Writing the reverse link would put knowledge from a future amendment into a row that was published before that amendment existed, which is the look-ahead this dataset exists to prevent. Resolve lineage forward from pit instead, after filtering by knowledge_date.
• The data is an extraction of public disclosures, not a complete record of beneficial ownership or actual market executions.

How the data was verified

Every published figure comes from the filing it cites; nothing is inferred by a model. Three checks carry the most weight, all documented in full in PIPELINE.md:

• The two routes agree. This build is loaded from SEC's quarterly Form 3/4/5 extracts. An earlier build parsed the same filings out of the daily dissemination archives, one submission at a time. On a shared quarter the two agree on all 65,957 submissions, on the exact Table I and Table II row counts with no per-filing discrepancy, on every price to the cent, and on the value distribution of all 61 published fields. Both routes feed one normalisation, so classifications cannot drift between them.
• Acceptance timestamps. The bulk metadata EDGAR publishes stamps many historical acceptance times in Eastern local time with a Z suffix, with no field saying which encoding a value uses. Reading those literally moves a filing four to five hours early — straight into look-ahead. Sampled across years, that misstatement persists into 2019 at rates from 14% to 83% and is gone by 2020. The correction in recipe/acceptance.py was measured against 68,000 filings of SEC-HEADER ground truth and reproduced it exactly: 45,209 resolved, 0 wrong. It settled 88% of this corpus; the remaining 491,819 filings had their acceptance instant read directly from their submission header, one ranged request each. Corrected values are labelled edgar_bulk_acceptance_corrected and read ones edgar_acceptance_datetime: an inferred timestamp stays distinguishable from a read one.
• Publication gates. recipe/quality.py runs before every publication and fails the build on duplicate accessions or transaction IDs, unsupported form types, negative shares, invalid A/D codes, non-monotonic PIT lineage, feature rows built from future input, and an accepted_at missingness above 10%. This revision passes all of them (data/quality_report.json).

Where knowledge_date comes from, across the 9,873,330 transaction rows:

Source	Rows	Share
edgar_bulk_acceptance_corrected	8,703,378	88.150%
edgar_acceptance_datetime (read from the header)	1,169,911	11.849%
filing_date_eod_fallback (estimated)	41	0.000%

Using it with ziplime

manifest.json declares everything a bundle needs: data_type: PIT_DATA, the storage class, and bundle_storage_data.table_uri pointing at data/pit/insider_trading.delta. Ingest reads the Delta table; entity_id is the issuer CIK, so resolving to ziplime assets means mapping CIK (or the point-in-time ticker) to the asset database.

Reading it directly, the as-of query has two steps that must happen in this order:

import polars as pl

pit = pl.read_delta("data/pit/insider_trading.delta")

# 1. Restrict to what was knowable at the simulation timestamp.
visible = pit.filter(pl.col("knowledge_date") <= simulation_time)

# 2. Resolve revisions: the newest state of each logical transaction.
latest = visible.sort(
    ["logical_transaction_id", "revision", "knowledge_date", "pit_event_id"]
).unique(subset=["logical_transaction_id"], keep="last", maintain_order=True)

# 3. Only now drop what is no longer in force.
as_of = latest.filter(pl.col("operation") == "UPSERT")

Each step depends on the one before it:

• Filtering by knowledge_date first is what keeps an amendment filed after simulation_time from deciding which row survives.
• Sorting by revision before knowledge_date matters because an estimated knowledge date is an upper bound, pushed to 23:59 of the filing day. A later revision can legitimately carry an earlier timestamp, and ordering by time would then return the superseded row.
• Dropping non-UPSERT rows last matters because a retraction is itself a revision. Filtering it out earlier would resurrect the row it retracted.

recipe/build_pit.py implements exactly this as as_of(pit, at), including the include_unresolved option for ambiguous amendment rows, which are excluded by default.

Pin a repository revision in every reproducible backtest.

Reproducibility

The history is loaded from SEC's quarterly Form 3/4/5 extracts, one archive per quarter, and shaped into the same intermediate form the XML parser produces so that a single normalisation computes every derived field. Knowledge time is resolved separately, from the bulk company index with its era correction and from submission headers for whatever that leaves undecided — because knowledge_date is baked into each row when it is written, a timestamp arriving afterwards would mean rewriting history rather than appending to it.

Identity is stable: accession, table and row ordinal. Writes are append-only and publication gates cover schemas, duplicates, quality, idempotency and PIT invariants. Commands are documented in PIPELINE.md.

Updates are designed to run on Hugging Face Jobs from jobs/run.py, a self-contained UV script stored in this repository, so the code that produced a revision ships with that revision. Each run restores the prior published state first, so a daily lookback cannot replace the historical bundle with only recent rows. Lint, tests and the dataset validator all run before upload; a failed gate leaves the previous revision as the latest valid release.

No schedule is registered yet. This revision was built and uploaded from a local rebuild. The daily job is ready to be turned on.

Every run publishes its own summary under data/run_log/, including the recipe hash, row counts, parser failures, and the resulting dataset revision.

License, source terms, and disclaimer

The repository's ZipLime-authored recipe, documentation, schema, and database compilation are offered under the Apache License 2.0, to the extent ZipLime has rights to license them. The underlying filings are created by reporting persons and issuers and are made available through SEC EDGAR; this project does not claim that those filings are U.S. Government works or that all source-document rights are owned by ZipLime. Source URLs and accession numbers are retained for verification. See NOTICE for scope and source details.

Use of SEC systems must follow its developer resources and Fair Access guidance. This dataset is provided for research and informational purposes, without warranty. It is not investment, legal, or compliance advice, and it is not endorsed by the SEC.