Every cluster screen I have seen ranks by member count, because member count is the field that exists and it sorts cleanly. It is also the wrong field, and the reason is the same reason a survey of one household is not a survey of five people just because five people live there.
Five members of the same state delegation, same party, filing the same ticker in the same fortnight are not five independent observations. They talk to each other. They sit in the same meetings, they hear the same regional briefing, several of them use the same handful of wealth managers, and in some cases they attended the same event where somebody mentioned a company. The information reaching all five arrived through one channel. Counting it five times is double counting with extra steps, and a ranking built on it puts the least informative cards at the top.
What a member count actually counts
Look at how the card is built. The summary line gives a member count, a direction and a USD volume. Underneath sits a table with Party, Politician, Type and Amount columns, and the rows are individual filings.
On the card I have in front of me the summary reads three politicians, and the visible rows carry the same party letter on all three, with one member's name appearing on two of them. So the headline count and the number of distinct people are not necessarily the same number, and the number of distinct people and the number of distinct information sources are not the same number either. Two collapses, one after the other, both in the direction of overstating breadth.
The other field worth reading sceptically is the summary direction. This card reads Sale on the header while the rows include a Sell and two Buys. A group whose members are trading in opposite directions is not a group with a view, and a ranking that treats it as one is ranking a border rather than a signal.
The breadth the tab gives you, and the breadth you have to derive
Party breadth is available directly. The Bipartisan Only view filters to cards whose members are not all from one party, and the Party column lets you see the composition at row level. That is a real axis and it is the easiest one to use.
Chamber is not exposed as a filter here. The legislators counter in the header reads 7,083 across Senate, House, Lords and MPs, so the underlying dataset knows which body each filer sits in, but on this tab you get it by recognising the names or by joining the member roster yourself. If you are running this systematically, join it, because chamber is the axis that carries the most information per unit of effort and you do not want it depending on whether an analyst recognises a name.
Delegation and committee are the same story. Both are derivable from a roster you maintain anyway, neither is a control on the card. Build the join once, attach state, chamber, party and committee assignments to every row, and then every card gets a breadth vector instead of a headcount.
A breadth score you can compute this week
The scoring rule I use is a deliberately crude penalty applied to distinct members, on the principle that a rough correction in the right direction beats a precise number computed on the wrong quantity.
Start from distinct members after deduplicating repeated rows by the same filer. Then apply a multiplier for each independence axis the cluster spans. Members from more than one chamber, multiply up. Members from more than one party, multiply up. Members from more than two state delegations, multiply up. Members sitting on committees with different jurisdictions, multiply up. A cluster that spans none of these, five members of one delegation from one party, gets the harshest treatment, and I would score it at close to the value of a single member because that is roughly what it is.
Work the two examples through. A six-member cluster, one party, one state, one chamber, scores as something between one and two effective sources. A three-member cluster with a senator and two representatives from different states and different parties scores as close to three, because there is no plausible single channel that reaches all three of them. The three-member card outranks the six-member card, which is the entire point of the exercise and is exactly the ordering a headcount sort will never produce.
The statistical framing behind the multiplier is familiar to anyone who has clustered standard errors. Correlated observations inflate apparent sample size and deflate apparent standard errors, and the correction is to count clusters rather than observations. A political trade cluster is that problem in its most literal form, and the fix is the same.
Why chamber is the strongest axis available
Party breadth is good and it is noisier than people assume, because members of opposite parties on the same committee share an information environment even while they disagree about everything else. Two members of one committee who happen to be from different parties are not two independent sources, and a bipartisan label on that card is misleading.
Chamber does more work because the two bodies run on different calendars, hold different hearings, have different committee jurisdictions and different memberships in every practical sense. A senator and a representative arriving at the same ticker inside a two week window are two people whose days do not overlap much. Whatever brought them both to the name had to be visible in two separate environments, and something visible in two environments is either widely available public information, which is checkable, or a broader development than a single briefing.
Add one important exclusion. The dataset spans jurisdictions, with Lords and MPs alongside the US chambers, and a cluster mixing US and UK filers is not a breadth win. Those are different disclosure regimes with different deadlines and different reporting conventions, and grouping them into one card creates apparent breadth out of a data join. Split by jurisdiction before you score, and treat cross-jurisdiction cards as two clusters rather than one broad one.
What the ranking does to the sleeve, and how to check it
Reranking by breadth cuts the actionable population hard. Cross-chamber, cross-party clusters on names outside the mega-cap complex are not common, and a screen tuned this way will produce a handful a month rather than a page a day. That is the intended effect, and it also means the sleeve trades less, holds fewer names and has more concentrated single-name risk, which needs saying out loud in the construction document rather than discovered at the first drawdown.
Now the honest part. Everything above is an argument from source independence, not a backtested result, and I have not shown you that breadth-ranked clusters outperform headcount-ranked ones. That is a testable proposition and you should test it before you commit capital to the ranking. The Performance tab gives you the benchmark to test against, and it is a demanding one in the useful direction. Across 5,000 analysed trades the population showed average alpha of minus 0.25 percent at 30 days, minus 0.34 percent at 90 days, and a win rate of 48.2 percent. That is the null. A breadth-ranked subset has to beat a population that is currently a slightly losing coin flip, and if your ranked subset cannot clear it out of sample, the ranking is decoration.
Run it as a straight horse race. Same universe, same holding period, same costs, two selection rules, breadth-weighted against headcount, and report both against the population number rather than against zero. Whichever wins, you will have a defensible sentence about why the sleeve selects the cards it selects, which is the question an allocator actually asks.