Sizing Correlated Positions Together

You own five regional bank stocks, each at 3% of the book, comfortably under a 5% single-name cap. On paper this looks diversified: five separate ideas, five separate stories, five separate risk budgets. In a rate shock or a regional-banking liquidity scare, all five move together, and what you actually hold is one 15% position in "regional bank stress," sized as if it were five unrelated 3% positions because nobody added them up.

The cap that catches what per-name caps miss

Per-name concentration caps are built to catch a single stock blowing up. They say nothing about a shared driver across several stocks that individually look small. The fix desks use is a cluster or theme cap: group positions by their dominant common factor, sector, geography, a shared supply-chain input, exposure to a single commodity or rate, and apply a separate, tighter limit to the sum of the group, independent of how each member scores against its own single-name cap.

The grouping is a judgment call, and it is worth erring toward grouping too aggressively rather than too loosely. Two stocks do not need to be in the same GICS sector to move together; a memory-chip maker and a construction firm can both be levered bets on the same interest-rate path even though nothing about their business description says so.

Worked example

Single-name cap is 5% of a $150m book ($7.5m). A trader holds five regional bank longs:

Every single position clears the 5% cap comfortably. If the desk's theme cap for "regional bank / rate-sensitive credit" is 8% of book ($12m), this cluster is 50% over its real limit even though it never triggered a single-name alert. The correction is to trim the cluster as a whole, likely dropping the two lowest-conviction names entirely rather than shaving all five, since a thin trim across five correlated names barely reduces the shared exposure that actually matters.

Building the check into the routine

This is not a one-time exercise at trade entry. Correlations between names you did not think were related can rise sharply during stress, the "everything correlates to one" effect, so the cluster check belongs in the same regular book review as the stale-position review, not just at the moment a new position is added.