How to Rebalance When Correlations Change

Correlation is usually measured once, at entry, and then left alone. That is where the risk accumulates. Positions opened in June as four separate ideas can be one bet by October without a single new trade being placed, because the relationships between them changed rather than the holdings.

Rebalancing for correlation means checking the book against the relationships that exist now, then reducing exposure to whichever driver has quietly taken over. It is a sizing exercise, not a view on any individual holding.

How a spread book becomes concentrated on its own

Three mechanisms do most of it, and none involves buying anything.

A new driver takes over. A rate decision, a commodity move or a single macro theme starts pricing assets that previously responded to their own fundamentals. The holdings did not change; what moves them did.

Winners grow. A position that has doubled is now a larger share of the book than when it was opened. If two of your winners sit in the same cluster, that cluster’s weight has risen twice without any decision being taken.

Theses converge. A company changes what it does. An acquisition moves it into a sector you already hold. Two holdings that genuinely were independent become competitors in the same market.

All three are invisible on a per-position risk report, because every individual line still shows the risk it was opened with.

Spotting it

The measurement is the same one used at entry, run again. Recalculate pairwise correlations across open positions on sixty days of daily returns, and compare them with the readings you had when each position was opened. The method is in how to calculate correlation between two stocks.

What you are looking for is not a precise number but a change of category: pairs that have moved from weak to strong. A pair crossing from below 0.3 to above 0.7 should be treated as one position for sizing from that point.

Two things are worth more than the matrix, though.

Run the sentence test again. Write the reason each position should rise, as of today rather than as of entry. Theses drift, and the sentences often converge before the returns do.

Watch whether several pairs rise together. One pair tightening is ordinary. Most pairs tightening at once usually means a shared driver has taken over the whole book, which is the pattern described in why correlation goes to 1 in a crash and which tends to precede the week you most need the diversification you thought you had.

Deciding what to cut

Once a cluster is identified, the question is which position to reduce. Three rules, in order.

Cut to restore driver count, not position count. The aim is more independent stories, not fewer tickers. Reducing two positions inside one cluster achieves that; reducing one from each of two clusters does not.

Within a cluster, cut the weakest thesis. Not the newest, not the one showing a loss. If you rewrote each sentence and one of them is now vaguer than the rest, that is the one whose reason to exist has weakened.

Trim winners that have outgrown their share. Uncomfortable, and necessary. A holding that has grown into a large share of a cluster contributes more to a bad week than its thesis supports. Trimming it says nothing about whether it continues to work; it is a statement about how much of the account is riding on one story.

The mistake that undoes the exercise

The most common error is rebalancing by adding.

The book looks concentrated, spare capital is available, and a new position feels like the constructive response. If that position shares a driver with the cluster being diluted, the concentration has increased while feeling like it decreased, and the account now carries more total exposure as well.

Adding only works when the new holding brings a driver that is not already represented, and when total exposure after the addition is still inside the limit. Where neither is true, the answer is reduction. Sitting in cash is itself a position, as sitting in cash is a position sets out, and it is often the correct one at the end of this exercise.

What rebalancing costs

Every adjustment is paid for, and the costs argue for acting on category changes rather than on small drifts.

Spread and commission are charged on each trim, and a book rebalanced aggressively can spend a meaningful share of its edge on the adjustments alone. This is the ordinary arithmetic covered in trading costs, and it scales with how often you act rather than with how much you move.

Tax consequences, where they apply, land on realised gains. Trimming a winner to reduce cluster weight can create a liability that the risk reduction does not obviously justify, which is a reason to prefer reducing the cluster’s weakest holding where that achieves the same thing.

Exit timing risk. Reducing a position means acting on a sizing judgement rather than on the thesis, and sometimes the trimmed holding then runs. That is the cost of the rule working as designed, and it is worth accepting rather than relitigating each time.

Together these argue for a threshold rather than continuous adjustment: act when a pair changes category or a cluster breaches its cap, and leave everything else alone.

A weekly routine

Fifteen minutes, once a week, on everything open.

1. Rewrite the one-sentence reason for each position as of today. Note any that now restate each other.
2. Recalculate pairwise correlations over sixty days. Flag any pair above roughly 0.7.
3. Group into clusters and total the risk in each.
4. Where a cluster exceeds the cap you set for one driver, reduce inside that cluster.
5. Record what you changed and why, so the next review has something to compare against.

The output is usually nothing. Most weeks the book has not drifted enough to act on, and recording that is part of the value: it means the weeks when something has changed stand out rather than arriving as a surprise after several positions have already fallen together.