The data behind every figure on this site

Every number on Hillscore is derived from the files below. They are published so a reader who doubts a headline can recompute it rather than take it on trust — including the sector benchmarks the comparisons are made against, and the reconciliation of the two sources, which are usually the parts nobody shows. The site's central claim, that a buy beat or trailed its own sector, rebuilds from these files exactly; a few longer-horizon figures need daily prices we do not publish, and we say which.

What is in them

They stack in that order. house_filings.csv is one of the two raw inputs, as read from the Clerk's PDFs and before any judgement was applied to it. combined_trades.csv is that source and the commercial feed reconciled, with every record the reconciliation set aside still present and carrying its reason. trades.csv is what came through it: every disclosed transaction the site can price, with the closing price on the transaction date attached, and a trade_uid on each row that joins back to the combined table. outcomes.csv is the derived layer: each disclosed buy with what the stock did afterwards and how that compares with its sector. sector_benchmarks.csv is what "compares with its sector" means in practice: the daily closes of the ETFs, joinable to outcomes.csv on benchmark and date. Everything else — the Composite score, the rankings, the signal comparison, the charts — is an aggregate of these.

Why the discarded records are published too. A pipeline that reconciles two sources is making choices, and a file that shows only the winners cannot be checked on the ones it dropped. Every record set aside is here with the reason: the same trade already held from the feed, a later filing restating an earlier one, a transaction dated in the future. A refiled report was being counted twice until a reader asked why a member's score had moved; it would have been one query against this file.

What the reconciliation does, and to how much

Two sources describing the same disclosures disagree in a handful of predictable ways. Each is handled by a rule rather than case by case, and every record it touches is in combined_trades.csv carrying the outcome, so the counts below can be checked against the file rather than taken from this page.

Nothing is corrected silently. A date that cannot be right is still not a date we know the answer to, and a pipeline that quietly rewrites 2202 to 2022 is guessing in a way a reader cannot see. So a record is either used, or set aside with the reason, or kept with the doubt written next to it. The last of those is the important one: a table that recorded only the decisions confident enough to delete something would look cleaner than the data actually is.

Which benchmark stands for which sector

One liquid, widely-tracked ETF per sector, mostly the SPDR Select Sector funds. 16 sectors, 14 ETFs: Insurance is measured against the financials fund and Telecom against the communications fund, because neither has a fund of its own worth using.

Reproducing the excess figures

Two rules decide whether your numbers match ours. Both are ordinary, and both will put you a few basis points out if you guess differently.

• Windows are calendar days, not calendar months. 1W is 7 days, 1M is 30 days, 3M is 91 days after the transaction date. Ninety-one days, not "three months later".
• Prices are the close on or before the target date. If a window lands on a weekend, a holiday or any other non-trading day, both the stock and the ETF roll back to the previous trading day's close. The same rule applies to the entry date.

Do that and every row's arithmetic closes: return_3M − benchmark_return_3M = excess_3M, and benchmark_return_3M rebuilds from the two closes in sector_benchmarks.csv. The benchmark file spans 2021-08-02 to 2026-10-30 — deliberately wider than the trades, so the earliest buy in outcomes.csv still has a sector to be measured against.

What these files cannot rebuild. They carry the closing price on each trade date, not a daily price series for every stock, so the one-week, one-month and three-month figures and everything derived from them can be recomputed — but the Composite score, the six- and twelve-month columns, the versus-ETF columns on the rankings and the estimated profit cannot. Those are computed from daily prices for every held position, which are not published here. We would rather say so than let you find out halfway through.

Read this before using it

• Amounts are ranges, not figures. Filings disclose bands, so amount_min and amount_max are the band, not what was spent. Anywhere the site needs a single number it uses the midpoint, and says so.
• Prices are closes, not execution prices. price_on_trade_date is the stock's ordinary closing price that day. Filings never disclose what was actually paid, and no figure here should be read as one.
• Returns are fractions. -0.030445 is −3.0445%. excess_* is the same window's move for the stock's sector benchmark ETF subtracted — the benchmark column names which ETF and benchmark_return_* gives its move, so the subtraction is visible in the row rather than taken on trust.
• Blanks are missing, not zero. An empty return_3M means the window has not elapsed yet or no price was available, and an empty sold_date means the position is still open. Treating either as zero will change your answer.
• This is not all of Congress. Every House member with ten or more disclosed stock purchases in the window, plus the senators our feed covers — not all 535. Who is tracked and why.

A sample of trades.csv

The 20 most recent rows, so you can see the shape before downloading 3.6 MB.

Licence

Published under a Creative Commons Attribution 4.0 International licence (CC BY 4.0). You may copy, redistribute, adapt and build on this data for any purpose, including commercially, provided you credit Hillscore with a link and say if you changed anything.

The licence covers this compilation — the matching, pricing, sector assignment and scoring — not the underlying facts. The filings themselves are United States government public records and are not anyone's to license, including ours.

Rebuilt daily alongside the rest of the site; both files carry whatever the latest run held. If a figure here disagrees with the filing it came from, the filing is right — tell us and it gets fixed.

Data last updated 31 October 2026.