Most venue exposure limits are inherited rather than derived. A venue gets onboarded, someone puts a number on it that felt defensible at the time, and the number survives every subsequent review because nobody has a better one. Attribution by venue is the obvious candidate for a better one, and it is genuinely useful, but not in the way the request usually arrives. Realised P&L per venue is a weak input to the limit itself. It is a strong input to where inside a band you sit.
A note on sourcing before the argument. The capture I am working from shows the Performance tab of the Portfolio module as it loads, which gives the view controls and the summary tiles. It does not show me the venue breakdown grid, so I will not specify its columns. The framework below assumes only that you can obtain a count of trades and a realised P&L figure grouped by venue, which is the minimum any attribution surface provides.
Why realised P&L is the wrong primary input
A counterparty limit answers a solvency question. It is the amount you are willing to have unrecoverable if the venue halts withdrawals, gets seized, reprices your collateral unilaterally, or simply stops answering. That question is about the venue's balance sheet, its custody arrangements, its regulatory perimeter and its history under stress. None of those are observable in your own P&L, and your P&L can be excellent right up to the morning the venue fails. Using realised profit to justify a larger balance held at a counterparty inverts the causality that matters.
There is a second problem, which is selection. The venues where you have made the most money are usually the venues where you have taken the most risk, held the largest balances, and used the most leverage. Sizing the limit on realised P&L therefore hands the largest credit exposure to the counterparty you are structurally most exposed to already. That is the concentration you were meant to be governing.
What venue attribution legitimately does is tell you whether the operational cost of maintaining a relationship is being repaid. A venue that consumes onboarding, reconciliation, monitoring and a slice of your credit budget, and returns nothing you could not have got elsewhere, should be closed or reduced. That is an economics decision, and attribution is the right evidence for it.
The floor, the ceiling, and the band attribution can move
Construct the limit in three steps and let attribution act only on the third.
The floor is operational. It is the working balance required to run the mandated activity at the venue across a full settlement cycle, including margin buffer at the stress level your risk policy specifies, plus whatever the venue's withdrawal latency forces you to strand. Below the floor the relationship does not function and should be terminated rather than starved.
The ceiling is credit. It is the largest loss from a single counterparty failure the fund can absorb without breaching a client-facing constraint, expressed as a share of net assets. This number comes from your risk policy and your diligence file, not from performance data. If the ceiling sits below the floor, the venue is uninvestable at your size and no amount of good attribution changes that.
The band between the two is where venue attribution earns its place. Inside it you are choosing how much optional balance to leave at a counterparty that could hold more, and the sensible basis is realised economics per unit of exposure. That is the number the third step needs, and it is not the raw P&L figure.
Normalising so venues are comparable
Raw P&L per venue answers "where did money appear", which is dominated by where you deployed the most capital for the longest time. Three normalisations turn it into something a committee can compare across counterparties.
Divide by average balance held at the venue over the period, so the answer is a return on the credit exposure rather than a dollar total. Then divide by time, so a venue used for two months is not compared with one used for twelve. Then subtract the venue-specific costs that never appear in trade P&L, which is where most of the real dispersion lives. Fee tier, funding rates paid on carried positions, financing on borrowed collateral, withdrawal charges, and the spread cost implied by depth. For a book turning over meaningfully, ten basis points of fee differential is usually a larger effect than any skill differential the attribution grid is showing you.
Sample adequacy comes next, and it is the constraint most attribution exercises fail. The tiles in the capture read 87 total trades for the whole book over 1Y. Split across counterparties, the per-venue counts are in the tens. The standard error on a win rate at n equal to 20, near a true rate of a quarter, is about 10 percentage points, so a two standard error band spans roughly 40 points. Any ranking of venues built on those rows is noise wearing a table. State the per-venue counts next to the per-venue figures in the memo, and let the reviewer see the sample rather than discover it later.
Two families of statistics on one screen
The summary tiles in the capture illustrate something worth handling explicitly, because it will come up the first time an outside reader sees your report. The same view shows Sharpe 1.23 and Sortino 9.84 alongside a profit factor of 0.08, expectancy of negative 1.20% and a win rate of 25.29%. Those readings point in different directions, and both can be correct without anything being wrong.
The general reasons two performance figures on one page legitimately diverge are worth writing into your own documentation as a standing footnote. They are computed over different measurement windows. They are computed on different return series, and specifically a per-trade series and an account balance series are different objects with different lengths, different weights and different treatment of periods where nothing is open. One may be annualised and the other not. One may be sensitive to the tail and the other to the median. And the sample lengths differ, since a ratio built on daily balance observations has far more points than one built on 87 closed trades.
I am not going to assert which of those explains this particular pair, because that requires knowing the exact construction and I do not. Nor is there any reason to assume a calculation is wrong. The way to resolve it for your own reporting is empirical. Fix one variable at a time, change the period from 1Y to ALL and watch which tiles move, switch the account selector from cumulative to a single named account and watch again, and you will quickly see which figures track the balance series and which track the trade list. Then cite the one that answers your question and label it as such.
What the limit memo has to record
The paper that sets the limit needs to be reconstructible by someone who was not in the room. Record the view state that produced every figure, meaning the reality toggle, the source filter, the account selection and the period, since those are the four dimensions the page exposes. Record the as-of date, because venue data backfills and today's figure for last quarter is not necessarily last quarter's figure. Record the per-venue trade counts alongside the per-venue economics. Record the costs you subtracted and the ones you could not obtain.
Then state the limit in the form that survives a bad outcome. Not "we allocate more to this venue because it performed", which reads badly in a post-mortem. Rather that the credit ceiling is X percent of net assets from policy, the operational floor is Y from the trading requirement, and the chosen balance sits at Z inside that band because the normalised realised economics over a stated period and a stated trade count supported it. If the venue fails, the first sentence is the one that defends you, and it was never a performance sentence.