The pool is a gross risk budget. It does not know your positions are related.

What the sum assumes

Adding four risks together is a statement about how they will fail.

The exposure layer sums active risk across open positions and compares the total against the authorised pool. That arithmetic is exactly right when the positions are independent, because then the pool is a worst-case bound that is very unlikely to be reached — some trades hit their stops, others do not, and the realised loss on a bad day lands well short of the sum. It is also exactly right in the opposite extreme, where all four are the same trade in different clothing, because then the worst case is not merely a bound but a plausible outcome. What the arithmetic cannot do is distinguish between those two situations, and they call for different amounts of exposure.

Where correlation comes from

In a currency book, related positions are the default rather than the exception.

This matters more here than it would in a diversified portfolio, because currency pairs share components by construction. Two positions on different pairs that both involve the same currency on the same side are substantially one bet on that currency. A view that a particular economy is strengthening can express itself across several pairs at once, and it frequently does — the setups appear together precisely because the same underlying move is generating them. So the situation the pool arithmetic handles least well is not an unusual one requiring an exotic scenario. It is the ordinary consequence of a coherent market read producing several qualifying setups in the same session.

Why it was built this way

A correlation-aware budget would need an input nobody can supply reliably.

The obvious repair is to weight the sum by the relationships between positions, and the reason the system does not is worth being honest about. Doing it properly requires a correlation estimate, which is a statistic computed over some historical window and which is famously unstable in exactly the conditions that make it matter — relationships that held for months converge during stress. A budget that relaxed exposure limits on the strength of a low correlation reading would therefore be most permissive shortly before the periods when the assumption fails. Between a constraint that is conservative and legible and one that is theoretically better and unstable, the system takes the first, which is the same trade it makes with fixed thresholds in the regime engine.

Which way the error runs

The limit is one-sided, and the side it errs on is the safe one.

It is worth being precise about the direction, because a stated limitation invites the assumption that the number is unreliable in both directions. It is not. Summing gross risk always produces a figure at least as large as the true aggregate downside, and equal to it only when the positions are perfectly related. So the pool constraint is never too loose relative to the correlation question — it is potentially too tight, binding an operator whose four positions really are independent. The failure mode of the limitation is forgone deployment rather than hidden leverage. That is the correct side to err on, and it means the layer remains sound as a ceiling even though it is imprecise as a measurement.

What the operator does about it

The correction lives in trade selection, where discretion already lives.

Since the arithmetic will not make the adjustment, the adjustment belongs where the system already places human judgement: in deciding which setups to take. Four qualifying setups that all express one view are a reason to take fewer of them, and that decision sits comfortably inside the discretion the system preserves, because it is a selection question rather than a sizing question. The sanctioned instrument if a smaller footprint is wanted for the cycle is the manual tier cap, which lowers the whole envelope and is logged. What is not available is quietly deploying the full pool across four related positions on the grounds that the exposure layer reported no problem — the layer did not report no problem, it reported a sum, and the sum was never the whole answer.

• Gross summing is exact for related positions and conservative for independent ones.
• Currency pairs share components, so relatedness is the default case, not the exotic one.
• The correction belongs in selection; the tier cap is there if a smaller cycle is wanted.

The key idea

A constraint is more useful when its blind spot is documented than when it is not.

Every risk limit rests on assumptions, and the ones that cause trouble are the assumptions nobody wrote down — because an undocumented assumption gets treated as a guarantee by whoever inherits the system, including the person who built it, two years later. Saying out loud that the pool is a gross budget which does not model relatedness costs nothing in credibility and buys an operator the ability to know when the number is doing less work than it appears to. The alternative is a limit that is silently approximate, which is the same thing without the warning.