Every Leaderboard row in Political Alpha carries a Compliance score, and almost everybody reads it as governance colour. It is not governance colour. It is a stable, repeatedly observed, per-filer behavioural measurement, computed on dozens or hundreds of independent events, sitting in a dataset where almost nothing else about a filer is measured that many times. That is an unusual object to have available and it is being wasted as a label.
The question worth asking is whether trade-to-filing delay carries information about the trades themselves. It is a proper feature-engineering question with a proper test, and the interesting part is that the two obvious hypotheses predict opposite signs, which is exactly the condition under which running the test is worth the effort.
What the module already measures, and at what granularity
The raw material is spread across three places and it joins at the filer.
The Leaderboard carries Compliance as the fourth per-row column, alongside Trades, Alpha 30d and Win Rate, across a population the page reports as 7083 politicians. The three rows visible on my capture read 100, 75 and 100, so it is a bounded score rather than a raw day count. The module documents a compliance leaderboard measuring punctuality of disclosure across politicians, which is the dedicated view behind that column.
The dashboard gives you the population-level parameters you need to size the exercise. Average delay reads 32.5d, subtitled "Trade to filing". Late filings in the current window read 330. So the central tendency is just over a month and there is a meaningful late tail to bucket against.
The third piece is the return side. Performance analytics are documented at 30, 60, 90 and 180 day forward horizons by politician, party and chamber. And the module documents a trade-versus-disclosure alpha decay measure described as the information advantage of late filings, which tells you the platform already treats the delay as an economically meaningful quantity rather than a hygiene statistic.
What I have not confirmed is a control that cross-tabulates compliance bucket against forward return directly. Assume you build that join yourself, because both sides are per-filer and the join key is obvious.
Two hypotheses, opposite signs
Hypothesis one is the information-preservation story. Delay is valuable to whoever traded, because the longer the gap between transaction and disclosure the more of the move happens before anyone else can see it. Under this hypothesis, delay is endogenous to the informativeness of the trade, and chronic late filers should exhibit stronger trade-date forward alpha than prompt filers.
Hypothesis two is the operational story. Delay is a property of the filing apparatus rather than of the trade. Filings are prepared by staff, from statements supplied by custodians and advisors, and a filer whose money sits in externally managed accounts has a longer and more error-prone path between a transaction occurring and a form being completed. Under this hypothesis, delay proxies for exactly the account structures where the named individual is furthest from the trading decision, and chronic late filers should exhibit weaker trade-date alpha than prompt filers, because more of their transaction flow is mechanical.
Both stories are plausible. Both have a clean directional prediction. They point opposite ways. A feature whose sign is genuinely contested before the test is a feature worth testing, and one whose sign everybody agrees on usually is not.
Note also that the two are not mutually exclusive and could be operating in different parts of the distribution, which is an argument for bucketing rather than fitting a linear term. Terciles or quartiles of the compliance score, with the population sized off the 7,083 filers and thinned by whatever minimum trade count you impose, will show you a non-monotonic relationship if one exists. A regression coefficient will average it away.
Two clocks, and only one of them is investable
This is the part that decides whether the study produces something you can deploy or something you can only publish internally.
Forward returns can be measured from the transaction date or from the disclosure date, and the two answer different questions. Transaction-date returns tell you whether the trade was informed. Disclosure-date returns tell you whether anything was left by the time the information became public, which is the only question that matters for a sleeve that trades on filings.
Run both. The interesting and quite likely outcome is that late filers show strong transaction-date alpha and no disclosure-date alpha at all, because the delay that made the trade look good is the same delay that consumed the opportunity. Under that result, hypothesis one is supported and the feature is still unimplementable as a long signal, which is a perfectly respectable finding provided you report it that way rather than quietly showing the transaction-date panel.
There is a usable version of the feature even in that case, and it is defensive rather than directional. If delay predicts decay, then per-filer expected delay is a forecast of how stale any given incoming filing is likely to be, and you can use it to route attention and to set an eligibility rule. Filings from filers in the prompt bucket get scored and sized. Filings from the chronic bucket get logged and excluded. That is a capacity-friendly use of the feature that does not require the sign of the alpha relationship to go your way.
The compliance score is cumulative, so using it raw is lookahead
Here is the defect that will invalidate the study if you skip it. A compliance score displayed today is computed over a filer's history, and that history includes filings that occurred after the trades you are scoring. Bucketing a 2023 trade using a 2026 compliance score means the bucket assignment was informed by three years of subsequent behaviour.
The magnitude is not trivial, because the score is bounded at 100 and a filer's rank within the distribution can move materially on a handful of late events. A filer who was prompt for a decade and then filed late repeatedly is in the chronic bucket today and was in the prompt bucket for most of the sample.
The fix is the ordinary one. Compute the feature as-of, using only filings dated before the trade being scored, with a minimum observation count before a filer becomes eligible. Ten prior filings is a reasonable floor and it will remove a large part of the population, which is a cost worth paying. Then keep the point-in-time series, because you will want to test whether a change in a filer's delay behaviour is itself informative, and that test is impossible from a single current snapshot.
The confounds you have to hold before the result means anything
Delay does not vary independently of everything else, and four confounds will masquerade as the effect you are looking for.
Account structure is the biggest. Trade count is a direct proxy for it, and the spread is enormous, with the top Leaderboard row carrying 49,087 disclosed trades against a population where most filers carry a tiny fraction of that. Bucket by trade count and check whether your delay effect survives inside buckets.
Size bracket is the second. Disclosure reports transaction size in ranges, the dashboard reports average trade as $40K explicitly labelled "Midpoint USD", and a very large share of activity sits in the lowest bracket. If delay correlates with bracket, and bracket correlates with informativeness, your delay coefficient is picking up size.
Chamber and tenure are the third and fourth, and they travel together. Committee position is a documented input to the per-trade signal score, committee position is a function of seniority, and seniority plausibly relates to staffing quality and therefore to filing promptness. Hold chamber fixed and stratify by tenure band.
Run the study with those four held and you have a result you can put in front of a risk committee. Run it without them and you have measured account administration, which is a real finding about the dataset and is not the finding you set out to test.