Portfolio Correlation Checklist Before You Add Risk

Most correlation damage is done at the moment of entry. A position looks independent, the per trade risk is within the rule, and it quietly doubles an exposure that was already there. The risk report confirms everything is fine, because each line is fine.

These seven checks run before opening a position. Together they take a few minutes and catch the overlaps that arithmetic alone misses.

1. Name the reason in one sentence

Before anything numerical: write down why this position should make money. One sentence, plain language, no chart references.

Then read the sentences you wrote for everything already open. If the new one restates an existing one in different words, you are adding size to an existing bet rather than adding a bet. That is a legitimate choice, but it should be a decision rather than an accident.

This check costs under a minute and catches more real overlap than any correlation matrix, because it sees relationships that have not yet shown up in returns.

2. Count drivers, not tickers

List the distinct economic forces your open positions depend on: a rate decision, an input cost, a single customer, a regulatory outcome, a commodity price.

If six positions resolve to two drivers, you hold two positions. The ticker count is irrelevant to the risk and actively misleading, because a longer list of holdings reads as more diversification on every screen you look at. The full argument is in correlation and portfolio heat.

3. Run the pairwise number against existing holdings

Calculate the correlation between the candidate and each open position, using daily returns over at least sixty trading days. Anything above roughly 0.7 is the same cluster for sizing purposes.

This is the arithmetic check, and it is a floor rather than a ceiling: it confirms overlaps you already suspected and occasionally surfaces one you did not. The procedure is in how to calculate correlation between two stocks.

4. Check it again on down days only

Filter both return series to days the benchmark fell meaningfully, then correlate only those.

The conditional figure is usually higher than the headline, sometimes dramatically. It is also the figure that describes the conditions under which several positions fail at once, which is the only scenario where correlation does real damage. Why that gap exists is covered in why correlation goes to 1 in a crash.

5. Price the cluster, not the position

Add the candidate’s risk to the risk already carried by the cluster it joins, then ask whether that combined number is one you would knowingly accept on a single idea.

If three positions each risking one per cent sit in one cluster, the honest reading is a three per cent bet, and the fourth entry makes it four. Deciding in advance what the per cluster cap is, and splitting risk within the cluster to respect it, is the whole mechanism. The position size calculator handles the split once the cap is set.

6. Ask what a bad week does to the total

Take the clusters, apply a plausible adverse move to each, and sum the result.

The output is a number for a bad week rather than a bad trade, and it is the figure that determines whether the account survives to trade the edge. If that total is larger than you are comfortable reading on a statement, the position to cut is in the largest cluster, not the newest entry.

What this does to long run results is covered in expectancy and position sizing.

7. Confirm the new position is not just the index

A final sanity check: if the candidate correlates strongly with the broad market and so does most of the book, the portfolio is an index position with extra steps and higher costs.

That is worth knowing explicitly. It is a reasonable thing to hold deliberately and an expensive thing to hold by accident, and the difference is whether the correlation was checked.

The checklist on a single entry

A worked pass, to show how fast it runs in practice.

The account holds four positions: a semiconductor firm, a cloud software company, an oilfield services business and a regional bank. Each risks one per cent. A fifth candidate appears, a second semiconductor name with a clean setup.

Check 1. The sentence for the candidate is “chip demand recovers faster than expected”. The sentence already written for the existing semiconductor position is nearly identical. The overlap is visible before any data is opened.

Check 2. Driver count across the existing four: chip demand, enterprise software spending, the oil price, and regional credit conditions. Four positions, four drivers, which is a genuinely spread book. The candidate adds no new driver.

Check 3. Pairwise correlation between the candidate and the existing semiconductor position comes back high, as the first check suggested it would. Against the other three it is unremarkable.

Check 4. On down days only, the candidate’s correlation with the software position rises too. Two drivers that looked separate in calm conditions are partially shared under stress, which is information the headline figure was hiding.

Check 5. The chip cluster currently carries one per cent. Adding the candidate at a full position would make it two. The decision becomes explicit: either accept a two per cent bet on chip demand, or split the existing one per cent across both names at half each.

Check 6. Summing the clusters under an adverse move gives a bad-week figure. If the split version keeps that total inside the limit and the full-size version does not, the checklist has made the decision.

Check 7. The book is not simply the index: four distinct drivers, only partly overlapping. Nothing to correct.

Elapsed time, with the spreadsheet already built: a few minutes. The output is not a prediction. It is the difference between holding four ideas and believing you hold five.

Running it weekly

The entry checks prevent the problem. A weekly pass over everything open catches the other half, because correlations drift: positions opened as independent ideas can converge as conditions change or as one story starts driving both.

Fifteen minutes a week, re-running checks three, four and six across open positions, is enough. What you are looking for is not precision. It is the pair you had been counting as two ideas that has quietly become one.