An independent review of public financial disclosures. Not affiliated with or endorsed by the U.S. House, the U.S. Senate, or any government agency.
Seal of the U.S. House of RepresentativesSeal of the U.S. SenateCongress Trading ReviewStock trades of the U.S. House and Senate
Methodology

How the tests work

Download the full technical methodology (PDF, 0.9 MB)

Each test compares a member's disclosed trades with a specific signature of trading on nonpublic information. The question is always the same: does this record look more like an ordinary managed account, or like someone trading on what they know?

Four readings, not a score

Every test produces one of four readings. We do not combine them into a single probability of misconduct: the tests overlap, a single number would hide which evidence drives it, and a probability would need a prior on wrongdoing that public data cannot supply.

Fits pattern
The evidence resembles a known insider-trading signature at the stated confidence level.
Open question
The evidence is inconclusive, or it raises a compliance or conflict-of-interest concern.
Does not fit
The evidence rules out a meaningful insider-trading signature for this test.
Not assessed
Too few trades to decide either way, or the required inputs are not yet available.
Screening signal
An open question whose unadjusted p-value is below 0.05. This is weaker than a fit: across all members, about 28 such signals would appear by chance alone. It is a kind of open question, not a fifth reading, and it is never counted as a fit.

Statistical framework

Each test contrasts two explanations. Under the null, trades are chosen by a process independent of nonpublic information, so signed abnormal returns average zero and timing is unrelated to the member's information calendar. Under the alternative, some trades are timed or chosen using that information.

Equivalence as well as significance

Failing to find an effect is weak evidence when a test has little power. So each test is paired with a two one-sided tests (TOST) equivalence check against a margin δ fixed before seeing the data. With a 90% interval for the oriented parameter θ:

Reading={fitsθ^−>0 after multiplicity adjustmentdoes not fitθ^+<δopen questionotherwise\text{Reading} = \begin{cases} \textit{fits} & \hat\theta^- > 0 \text{ after multiplicity adjustment} \\ \textit{does not fit} & \hat\theta^+ < \delta \\ \textit{open question} & \text{otherwise} \end{cases}

Multiplicity

With many tests per member, and hundreds of members, some results will reject by chance. Family-wise error within a member is controlled with Holm's step-down procedure; Benjamini–Hochberg q-values are reported as a screening complement across members.

reject H(k) for k<k∗,k∗=min⁡{k:p(k)>α/(m−k+1)}\text{reject } H_{(k)} \text{ for } k < k^\ast,\quad k^\ast = \min\{k : p_{(k)} > \alpha/(m-k+1)\}

Date-clustered bootstrap

Trades on the same day share the same market shock, so intervals resample trading dates rather than individual trades, with amount imputation nested inside each replication, and use bias-corrected and accelerated (BCa) quantiles.

Power

For a mean signed abnormal return from N trades with H-day idiosyncratic volatility σ and within-date correlation ρ, the minimum detectable effect is shown below. Members with few trades get wide intervals, which usually leads to open question or not assessed rather than a false clean bill.

MDE=(z1−α/2+z1−β) σHN1+(nˉcl−1)ρ\mathrm{MDE} = (z_{1-\alpha/2} + z_{1-\beta})\,\frac{\sigma_H}{\sqrt N}\sqrt{1 + (\bar n_{\text{cl}} - 1)\rho}

Range-reported amounts

Disclosures give a value band, such as $1,001 to $15,000, never the exact amount. Every statistic carries that uncertainty in one of three ways: sharp bounds evaluated at the band endpoints (partial identification), a point estimate at the geometric midpoint for descriptive figures, and log-uniform draws inside the bootstrap for intervals.

v~i=LiUi,vi(b)=exp⁡(ln⁡Li+Ui(b)(ln⁡Ui−ln⁡Li))\tilde v_i = \sqrt{L_i U_i}, \qquad v_i^{(b)} = \exp\big(\ln L_i + U_i^{(b)}(\ln U_i - \ln L_i)\big)

The tests

Size and breadth

T1Trade sizeAre the member's trades unusually large, especially just before news?T2Breadth and concentrationIs trading concentrated in a few companies rather than spread across a diversified portfolio?

Timing

T3Cadence and event proximityDo trades arrive in routine rebalancing batches, or sporadically just before company news?T4Event timing (abnormal returns)Do purchases rise and sales fall more than the market in the weeks after the trade?T5Direction before market-wide shocksBefore the market's biggest moves, did the member's trades point the right way?T12Anticipation of large stock movesWhen a traded stock made a large unexpected move soon after the trade, was the trade on the right side?

Control and behavior

T6Control and independenceDoes the member direct the trades, and does trading change when the member's access to information changes?T10Behavior after scrutinyDid trading change after press scrutiny or an ethics inquiry?T14Family-account tradesDo trades in a spouse's or dependent's name do better than the member's own?T15Short-swing round tripsDoes the member buy and sell the same stock within six months, and do those round trips beat the market?T16Options and leverageDoes the member use options, and do option trades point the right way?

Compliance

T7Disclosure timelinessWere trades disclosed within the STOCK Act's 45-day limit?

Policy and committee overlap

T8Committee overlapDoes the member trade industries their committees oversee, and do those trades do better?T9Policy-topic overlapDid trades in stocks affected by the member's legislative work come before votes and move the right way?

Peer comparison

T11Performance relative to CongressAfter accounting for noise, does the member's trading beat that of other members?T13Co-trading with other membersDoes the member trade the same stock in the same direction as other members at the same time more than chance predicts?

Robustness

Every quantitative test is re-run with a four-factor model in place of the market model, at horizons of 5, 21 and 63 trading days, with equal and dollar weighting (the latter at both band bounds), excluding the largest batch date, and excluding spouse and dependent accounts. A reading is labelled robust only if it holds across all variants; each member's scorecard lists the variants.

Limitations

  • The tests measure resemblance to known patterns in public data. They cannot show what a member knew, and a fitting result is not a finding of wrongdoing.
  • Only disclosed trades are tested. Late, missing or misreported filings, and assets held through funds or trusts, are outside what the data can see.
  • Amounts are bands, so dollar-weighted statistics are reported as ranges or intervals.
  • Committee jurisdiction and policy-topic maps involve judgment; the maps are published with the data.

The full technical document, with every estimator, interval and algorithm, is available as a PDF download.