Congressional trading disclosures
The transaction date is on the filing, and it is not the date the information existed.
A Periodic Transaction Report is due up to 45 days after the trade, and the routine penalty
for missing that is a $200 fee. Keying a backtest on the transaction date is the same bug
../../../fin-core/skills/fundamental-and-macro-data/SKILL.md documents for a fiscal period
end, in a domain where it is easier to make because the filing prints both dates side by side.
Everything marked ✅ Measured is printed by scripts/congress.py (numpy + pandas, seed
20260910, about 15 s, no network). Its filings are synthetic and seeded - the lag
distribution is one this repo chose and then measured back, not a tally of real filings.
Everything marked ✅ source-verified was read at the URL given, on 2026-09-10.
1. ✅ The regime, at primary sources
| Fact | Value | Source, read 2026-09-10 |
|---|---|---|
| PTR deadline | "Not later than 30 days after receiving notification of any transaction", and "in no case later than 45 days after such transaction" | 5 U.S.C. 13105(l) (uscode.house.gov) |
| Reporting trigger | "the gross amount of a single purchase or sale transaction exceeds $1,000" - gross, so a losing trade still counts | House PTR form CY2025 (ethics.house.gov), 5 U.S.C. 13104(a)(5)(B) |
| Late-filing fee | "$200 penalty shall be assessed against anyone who files more than 30 days late" | 5 U.S.C. 13106(d); House PTR form; Senate PTR instructions |
| Civil penalty | base $50,000 for a knowing and willful falsification, inflation-adjusted to $75,540 for 2025 | 5 U.S.C. 13106(a)(1); 90 FR (2025-01-15) |
| Use restriction | unlawful for commercial purposes, credit rating or solicitation; penalty "not to exceed $10,000" | 5 U.S.C. 13107(c) |
| Online posting | STOCK Act sec. 8's "search, sort, and download" mandate was suspended by Pub. L. 113-7 (2013-04-15) for staff, but the carve-out expressly preserves "(C) Any Member of Congress. (D) Any candidate for Congress." | govinfo.gov PLAW-113publ7 |
🚨 The statute is 5 U.S.C. 13105(l), not 13104(l). 13104 is Contents of reports and has no subsection (l). Both numbers circulate; only one resolves.
🔑 The deadline is not the only clock. The 30-day leg starts at notification, and the $200 fee only attaches past 30 days late, i.e. day 75 after the trade. Nothing in the filing tells you which leg the filer was on.
The amount brackets - the amount itself is never disclosed
✅ House PTR form CY2025, columns A-J, plus a column K that has no numeric equivalent:
| bracket | width | bracket | width | |
|---|---|---|---|---|
| $1,001-$15,000 | 15.0x | $1,000,001-$5,000,000 | 5.0x | |
| $15,001-$50,000 | 3.3x | $5,000,001-$25,000,000 | 5.0x | |
| $50,001-$100,000 | 2.0x | $25,000,001-$50,000,000 | 2.0x | |
| $100,001-$250,000 | 2.5x | over $50,000,000 | open | |
| $250,001-$500,000 | 2.0x | K: spouse/dependent child over $1,000,000 | open | |
| $500,001-$1,000,000 | 2.0x |
🚨 Column K collapses G, H, I and J into one bucket for an asset owned by a spouse or dependent child in which the filer has no interest (5 U.S.C. 13104(e)(1)(F)). A $1.1M trade and a $60M trade are the same disclosure. ⚠️ The statutory categories at 13104(d)(1) start at "not more than $15,000"; the $1,001 floor is a form convention that follows from the $1,000 reporting trigger. ⚠️ The Senate's own bracket list was not verified - the Senate PTR instructions do not print it and the form sits behind the click-through in §2.
2. Where the filings live, and what is NOT knowable
✅ House - disclosures-clerk.house.gov/FinancialDisclosure, read 2026-09-10. Annual
archives at /public_disc/financial-pdfs/<YEAR>FD.zip, 2008-2026. The path segment is
literally financial-pdfs. A displayed notice quotes the 5 U.S.C. 13107(c) use prohibition;
there is no checkbox and no documented public API.
✅ Senate - efdsearch.senate.gov/search/home/, read 2026-09-10. Covers "Senators, former
Senators and Senate candidates filed from 2012 to present". Access is gated by a
click-through: a checkbox reading "I understand the prohibitions on obtaining and use of
financial disclosure reports." ✅ Both sites return HTTP 404 for robots.txt, so the
restriction is the statute and the agreement, not a crawl directive.
⚠️ What this repo did not verify, and will not assert:
- The internal structure of the House ZIP. Every description found of an XML index inside it came from a data vendor, not from the Clerk. Not opened here.
- What the Senate serves per report (HTML tables for e-filers, PDFs for paper filers is the common claim) and whether a bulk download exists. Both sit behind the agreement, and accepting an agreement on a user's behalf is not something this repo does.
🚨 The consequences for a pipeline, and they are the reason this skill ships no scraper:
- The unit of publication is a document, not a record. A PTR is a PDF or semi-structured HTML listing several transactions. Parsing it is an OCR/layout problem with its own error rate, and that error rate is a data-quality term nobody reports.
- Several aggregators resell this data. They add a ticker mapping, an asset-type guess
and a "disclosed" timestamp - all three are their inference, not the filing's content.
A vendor's
disclosure_datemay be its ingestion date, which is later again. - This library ships no scraper and no credential path for either site. Read the terms yourself, and if you use a vendor, ask which of its fields are transcribed and which are derived.
⚠️ Ticker resolution is a research step, not a lookup. Filings name assets in prose
("Apple Inc. - Common Stock"), and the same prose covers options, bonds and funds. Resolving
that string to a security is the same (identifier, date) problem as
../../../fin-core/skills/research-integrity-guards/SKILL.md §1, and it is where a
survivorship bias enters if you map against today's ticker table.
3. ✅ Measured - the lag this file chose, and what it produced
draw_lags() sets the distribution: a Beta squeezed onto [2, 45] days because a deadline is a
target rather than a mean, plus a 12% tail past the deadline because the fee is $200.
Measured off 1,500 seeded filings, calendar days from transaction to disclosure:
| mean | p05 | p25 | median | p75 | p95 | p99 | max |
|---|---|---|---|---|---|---|---|
| 43.1 | 11.8 | 22.3 | 31.2 | 38.6 | 141.2 | 306.6 | 471.2 |
✅ Measured: 11.7% past the 45-day deadline, 9.3% past 90 days, 0.5% past a year.
⚠️ These are properties of a generator, not a census of Congress. What is not a choice is the shape: the median filing is a month old when it becomes public, and the tail is long. Measure your own vendor's lag distribution before you quote any number from anywhere, including this table.
4. ✅ Measured - the A/B
One rule: long the disclosed name for 63 trading days, short the equal-weight universe, entered at the close of the key day so the first return earned is the next day's. Same trades, same names, same holding period; only the date you may act on differs.
| key | Sharpe | ann. return | ann. vol | retained |
|---|---|---|---|---|
| transaction-date (look-ahead) | 0.880 | 4.33% | 4.92% | 100% |
| disclosure-date (honest) | 0.362 | 1.82% | 5.01% | 41.2% |
Over 40 seeds, because one seed is one draw:
| mean | sd | p05 | p95 | > 0 | |
|---|---|---|---|---|---|
| transaction-date | 1.038 | 0.266 | 0.535 | 1.428 | 100% |
| disclosure-date | 0.449 | 0.275 | 0.055 | 0.839 | 100% |
| gap | 0.589 | 0.304 | 0.226 | 1.077 | 100% |
🔑 The gap is 0.589 of Sharpe and it is positive in every one of 40 seeds. The volatility of the two runs is the same to two decimal places - the look-ahead is not buying you a smoother ride, it is handing you a return you could not have earned.
5. 🚨 Measured - the gap is a ratio, not a number of days
Disclosure-date Sharpe as a fraction of the transaction-date Sharpe. Rows scale the whole lag draw (mean calendar days in brackets); columns are the half-life of the information in trading days. 12 seeds per cell.
| lag | half-life 5d | half-life 21d | half-life 63d |
|---|---|---|---|
| 0.1x (4 days) | 65.8% | 87.3% | 91.7% |
| 0.25x (11 days) | 48.5% | 79.4% | 86.7% |
| 0.5x (22 days) | 34.8% | 70.5% | 80.7% |
| 1x (44 days) | 19.9% | 52.3% | 65.3% |
| 2x (89 days) | 3.1% | 19.5% | 29.7% |
🔑 Read down a column, not across the table. At a 5-day half-life the statutory lag alone
removes 80% of the signal; at a 63-day half-life the same lag leaves two thirds of it. "45
days" says nothing until you say what it is 45 days of. ⚠️ The decay model is a choice
(exponential, stated in the script). The invariant is that the loss depends on
lag / half-life, so the first thing to measure on your own data is the half-life, by
running the disclosure-keyed rule at several holding periods.
6. ✅ Measured - what the brackets do to a size-weighted signal
Three measurements, in increasing order of how much they lean on the generator.
(a) Ordering - a property of the brackets alone. Two filings in the same bracket are indistinguishable no matter which imputation you pick. ✅ Measured on 18,000 pooled seeded filings: 22.6% of all pairs are exact ties, and the modal bracket ($15,001-$50,000) holds 32.5% of filings. Column K, the spouse/dependent-child catch-all, hides a p90/p10 ratio of 12.9x inside a single disclosed value.
(b) The dollar aggregate - "Congress bought $X of this name". Imputed total as a fraction of the true total, columns being the value you assign the open-ended top brackets as a multiple of their lower edge - a number nobody publishes:
| imputation | top x1 | top x2 | top x5 |
|---|---|---|---|
| bracket midpoint | 81.2% | 92.0% | 124.4% |
| geometric mean | 68.5% | 79.3% | 111.8% |
| lower bound | 43.0% | 43.0% | 43.0% |
🚨 ✅ Measured: 1.2% of filings sit in an open-ended bracket and carry 42.0% of the true dollars. The headline aggregate moves from 81% to 124% of truth on one unpublished assumption. A dollar aggregate is not a measurement; it is your prior about the tail.
(c) The Sharpe cost. Dollar-weight the disclosure-keyed portfolio five ways, 12 seeds.
Only true needs a number nobody published:
| weight | Sharpe | sd | cost vs true |
|---|---|---|---|
| true amount (unknowable) | 0.839 | 0.292 | - |
| bracket midpoint | 0.718 | 0.342 | -0.121 |
| geometric mean | 0.743 | 0.343 | -0.095 |
| lower bound | 0.799 | 0.334 | -0.040 |
| equal weight (ignore size) | 0.590 | 0.327 | -0.249 |
✅ Measured: size information is worth 0.249 of Sharpe here, and the bucketing throws
away about half of it. Re-valuing the open-ended brackets at 5x their lower edge moves the
midpoint run from 0.718 to 0.774 - the same free parameter as in (b). ⚠️ The level depends on
the alpha-to-size link the script sets (dollar_exponent=0.5, sublinear); the ordering does
not. 🔑 The lower bound beats the midpoint here, which is the point: the midpoint is an
estimator with a bias, not a neutral default, and each of these five is a trial to log in
../../../fin-core/skills/backtest-validation/scripts/trial_ledger.py.
7. Traps
- 🚨 Keying on
transaction_datebecause the column exists. §4: 0.589 of Sharpe, positive in 40 of 40 seeds. Use the date the filing became public, and prefer the vendor's ingestion timestamp over itsdisclosure_datewhen the two differ. - 🚨 Dropping the late filings. They are 11.7% of this panel and they are not missing at random - a filer who is late is different from one who is not. Dropping them is a selection on the disclosure process itself.
- 🚨 Backfilling an amendment onto its original date. An amended PTR is a new publication. Use the amendment's own date, the same rule as a restated 10-K.
- 🚨 Treating the bracket midpoint as the amount. §6. And the top bracket has no midpoint.
- 🚨 Aggregating "Congress" as one trader. 500+ filers with different mandates, staff and brokers. An equal-weighted aggregate over all of them is a survey, not a portfolio.
- 🚨 Selecting the members after seeing the returns. "Follow the five best members" chosen
in-sample is the same defect as an influencer panel - see
../social-and-influencer-feeds/SKILL.md§on selection, which measures it. - ⚠️ Sales are not the mirror of purchases. A sale can be tax, a blind-trust rebalance or a
divestiture required on taking a committee seat. The same asymmetry the insider literature
reports (
../insider-form-4/SKILL.md§on codes) applies here with less documentation. - 🚨 Filings are documents. Your parser's error rate belongs in the result. If you cannot state it, say the Sharpe is an upper bound.
8. Scripts and where this sits
scripts/congress.py - the transcribed statutory constants; draw_lags() and
lag_summary(); bracket_of() / impute_size() with the five imputations and the
open-ended-bracket parameter; make_panel(); portfolio() / run_ab() / ab_over_seeds();
lag_grid(); ordering_loss(), aggregate_bias(), weighting_table(). numpy + pandas,
seed 20260910, about 15 s, no network, no file writes, no scraper.
../insider-form-4/SKILL.md- the same argument with a two-business-day deadline instead of 45, and the transaction-code filter that decides whether there is a signal at all.../institutional-13f/SKILL.md- the same argument where the lag is measured from a quarter end, so the position can be four months old before you see it.../social-and-influencer-feeds/SKILL.md- the selection and deletion biases that a "top members" or "top accounts" panel shares with this one.../../../fin-core/skills/research-integrity-guards/SKILL.md§2 - the availability rule this skill is one instance of, and the(identifier, date)join that ticker resolution needs.../../../fin-core/skills/fundamental-and-macro-data/SKILL.md- the period / filed / acceptance distinction, and itspit_fundamentals.py, the same as-of read in code.../../../fin-macro/skills/real-time-macro-backtesting/SKILL.md- the identical A/B on macro vintages, and the finding that the timing leg usually dominates the value leg.../../../fin-core/skills/backtest-validation/SKILL.md- a Sharpe of 0.449 with sd 0.275 is not distinguishable from a good one; the imputation choices in §6 are trials.