README.md

---
pretty_name: "knowable: US financial disclosures with a publication clock"
language:
- en
license: other
license_name: knowable-data-terms
license_link: LICENSE
annotations_creators:
- found
- machine-generated
source_datasets:
- original
size_categories:
- 10M<n<100M
tags:
- finance
- tabular
- timeseries
- sec
- edgar
- form-4
- 13f
- congress
- insider-trading
- point-in-time
- public-domain
- us-government
- open-government-data
configs:
- config_name: congress_trades
  default: true
  data_files:
  - split: train
    path: releases/2026-08-01/congress_trades/*.parquet
- config_name: house_filings
  data_files:
  - split: train
    path: releases/2026-08-01/house_filings/*.parquet
- config_name: senate_filings
  data_files:
  - split: train
    path: releases/2026-08-01/senate_filings/*.parquet
- config_name: fund_holdings
  data_files:
  - split: train
    path: releases/2026-08-01/fund_holdings/*.parquet
- config_name: insider_transactions
  data_files:
  - split: train
    path: releases/2026-08-01/insider_transactions/*.parquet
- config_name: coverage
  data_files:
  - split: train
    path: releases/2026-08-01/coverage/*.parquet
---

# knowable

US government financial disclosures as a point-in-time dataset: congressional
trades, SEC Form 4 insider transactions, and 13F fund holdings. Every row carries
**the instant it actually became public** and a **measured coverage denominator**.

**35,213,599 rows. 602 Parquet partitions. 958 MB zstd. Release `2026-08-01`.**

Immutable: corrections ship as a new release naming this one in `supersedes`,
never as a rewrite of these bytes.

## Access

```python
from datasets import load_dataset

trades = load_dataset("jwlutz/knowable", "congress_trades", split="train")
```

Query without downloading everything, which is what you want at 34M rows:

```python
import duckdb

duckdb.sql("""
    SELECT filer_name, ticker, side, amount_range_low, published_at
    FROM 'hf://datasets/jwlutz/knowable/releases/2026-08-01/congress_trades/*.parquet'
    WHERE ticker = 'NVDA' AND published_at <= '2026-01-01'
    ORDER BY published_at DESC
""")
```

Configs: `congress_trades`, `house_filings`, `senate_filings`, `fund_holdings`,
`insider_transactions`, `coverage`.

### Point-in-time queries

This is the reason to use this dataset rather than a scraped one. `published_at`
is when the filing became readable: not when the trade happened, not when we
fetched it. Filtering on it reconstructs what was knowable at a past instant.

```sql
-- exactly what a reader could have known on 2024-06-30, nothing later
SELECT * FROM congress_trades WHERE published_at <= TIMESTAMP '2024-06-30 23:59:59'
```

Do not use `transaction_date` for this. The median trade is disclosed 28 days
after it happens and 18.7% take longer than the statutory 45 days, so
`transaction_date <= T` silently includes filings nobody could read at `T`.

### Candella SDK

```python
from candella import knowable

trades = knowable.congress_trades(since="2026-01-01", ticker="NVDA")
```

The SDK wraps the HTTP contract below, so an SDK query and a direct API query
return identical rows.

**Status: the read API is not hosted yet.** The FastAPI app and a client both
exist and are tested against real release bytes, but no public endpoint serves
them today. Use the Parquet paths above until that lands. The contract, so you
can write against it now:

```
GET /v1/{resource}?cursor=&limit=&as_of=
    -> {"data": [...], "next_cursor": str|null, "as_of": timestamp}

resources: congress/trades, house/filings, senate/filings,
           insider/transactions, fund/holdings, coverage
```

`as_of` applies the point-in-time bound server-side. Paging is keyed on
`(published_at, id)`, so a cursor stays valid across releases.

## What it covers

| table | rows | one row is |
|---|---|---|
| `fund_holdings` | 34,030,485 | one 13F position line |
| `insider_transactions` | 1,067,193 | one Form 4 transaction line, all codes |
| `congress_trades` | 72,140 | one disclosed trade (House 55,796, Senate 16,344) |
| `house_filings` | 41,380 | one House filing with its publication clock |
| `senate_filings` | 2,394 | one Senate eFD document with its clock |
| `coverage` | 7 | one source-period with its measured denominator |

Congressional transactions span 2012 to present; House publication clocks reach
back to 2008-03-26. The SEC lanes cover 2024-01 through 2026-07, and every
statistic you compute from them inherits that boundary.

### Coverage, against each source's own denominator

This table ships **inside the release** as `coverage`, so incompleteness is
queryable rather than a claim in a README.

| source | captured | expected | coverage | known gap |
|---|---|---|---|---|
| `sec_form4` | 462,659 | 464,209 | 99.7% | 1,550 |
| `sec_13f_hr` | 86,123 | 86,124 | 100.0% | 1 |
| `house_clerk` | 41,380 | 41,454 | 99.8% | 74 |
| `senate_efd` (documents) | 2,394 | 2,394 | 100.0% | 217 |
| `senate_efd_ptr` (transcribed) | 1,791 | 2,394 | 74.8% | 603 |
| `house_clerk_ptr` (our parse) | 5,407 | 8,306 | 65.1% | 2,899 |
| `house_clerk_ptr_congresskit` (fill) | 372 | 2,899 | 12.8% | 2,527 |

**Read the two House PTR rows together: 5,779 of 8,306 documents, 69.6%.** That
is the weakest number here and it is not averaged away. The missing 2,527 are
2,437 paper filings that exist only as scanned images plus 90 e-filed documents
no public archive holds. Denominators come from the publisher's own index, never
from our own output.

Two scope caveats that change how a number reads:

- `senate_efd` expected counts **elected filers only**. The all-filer Senate PTR
  total is 11,830, so 1,791 transcribed is 74.8% of elected-filer PTRs but 15.1%
  of all Senate PTRs.
- The 217 gap on `senate_efd` is a **floor**. Third-party mirrors hold 217
  elected-filer documents that eFD's six-year retention dropped from its index,
  so the historical denominator is strictly larger than 2,394 and its true size
  is unmeasured.

## Speed and reliability

### Publication clock precision

| precision | share of `congress_trades` | meaning |
|---|---|---|
| `exact` | 88.4% (63,803) | a real second, from the publisher |
| `date_only` | 11.6% (8,337) | the source gave a date and no time |

`house_filings` is **100% `exact`**, back to 2008. The vocabulary is exactly
three values, `exact` / `first_seen` / `date_only`, and a row is never promoted
to a precision its source does not support. Midnight is not a time we observed,
so a `date_only` row says so rather than presenting `00:00:00` as an instant.

House clocks come from the clerk's own HTTP `Last-Modified` header. Senate
`exact` clocks come from bisecting eFD's `submitted_end_date` bound, which
accepts one-second granularity and so reads back the publisher's stored second.
**eFD states no timezone**: `America/New_York` is inferred from the filing-hour
distribution and corroborated against an independent 2020 scrape on 969 of 969
documents. That inference travels in `extraction_notes` on every row it affects.

### Disclosure latency

Transaction date to publication instant, over 72,135 congressional trades:

| p50 | p75 | p90 | p99 | within statutory 45 days |
|---|---|---|---|---|
| **28 days** | 40 days | 204 days | 819 days | **81.3%** |

Newest filing in this release published `2026-07-29`.

### Integrity

- Every partition carries a **sha256 in the manifest**, verified at publish time
  by reading hashes back from the Hub rather than trusting the upload succeeded.
- Releases are **immutable**; partitions are never rewritten.
- The build **refuses rather than warns**. A dangling foreign key, an
  unregistered parser version, a duplicate primary key, or a provider absent from
  the redistribution allowlist each abort the build before anything is written.
- 17 data expectations run over the real corpus on every build.

### Source errors are carried, not silently repaired

63 of 72,140 rows (0.118%) carry a `transaction_date` **after** their own
publication instant, which is physically impossible. They are filer typos: years
written as 3031, 2202, 2220. They ship as filed. Normalization reads the source's
vocabulary; it does not repair the source. Filter `transaction_date <=
published_at` if you need them gone.

Exclusions are handled the same way. 24,196 rows extracted by OCR from scanned
paper filings are **deliberately absent** from this release because their field
accuracy has not been human-adjudicated.

## Licence and citation

Code is Apache-2.0. **The data carries no licence, because it is not ours to
license.** These are US government records and facts are not copyrightable
(*Feist v. Rural Telephone*). We assert no copyright and grant none. See
`LICENSE` for the full statement and `ATTRIBUTION.md` for upstream credits.

**A statutory restriction binds you regardless of anything we say.**
5 U.S.C. 13107(c)(1) prohibits obtaining or using these reports for any unlawful
purpose, for commercial credit rating, for solicitation, or to determine or
establish credit rating. That is federal law, channel-independent, and not ours
to waive.

```bibtex
@misc{knowable2026,
  title  = {knowable: US government financial disclosures with a publication clock},
  author = {Lutz, Jack},
  year   = {2026},
  note   = {Release 2026-08-01},
  url    = {https://huggingface.co/datasets/jwlutz/knowable}
}
```