Why Correlation Goes to 1 in a Market Crash

In a severe market decline, assets that spent years behaving independently begin moving together, and a portfolio built to spread risk across six ideas discovers it held one. This is not bad luck and it is not a failure of measurement. It is a change in what is pricing the assets, and it happens reliably enough that it belongs in the sizing decision rather than in the post mortem.

The short answer: in calm conditions prices are driven mostly by things specific to each company, which differ. Under stress they are driven by one shared thing, and one driver produces one behaviour.

Normal conditions price idiosyncrasy

On an ordinary day, a semiconductor firm and a regional bank are being priced on different questions. One is about chip demand, inventory and a product cycle. The other is about loan books, deposit costs and local credit conditions. Those inputs are close to independent, so the daily returns are too, and the measured correlation across a calm year is genuinely low.

That low number is a real description of that period. The mistake is treating it as a property of the pair rather than a property of the conditions the pair was measured in.

Stress replaces many drivers with one

When a serious drawdown arrives, the question being asked of every risk asset becomes the same question: does anyone want to hold this at all right now?

Several mechanisms push in that direction at once.

Forced selling ignores fundamentals. A fund meeting redemptions sells what it can, not what it would prefer to. A trader meeting a margin call sells whatever is most liquid. Neither sale carries any view about the specific company, and both push unrelated assets down together.

Liquidity thins at the same moment for everything. Spreads widen across the board, and the price impact of a given order size rises everywhere at once. That shared mechanism alone creates shared movement.

Risk appetite is a single variable. The decision to reduce exposure is taken at the portfolio level, not the position level. When many participants take that decision in the same week, the effect lands on every holding regardless of its merits.

None of those mechanisms care which stocks you chose. That is precisely why the correlation you measured in the calm period stops describing the one you now face.

The measurement that would have told you

A single correlation figure across a year averages the quiet days, which dominate by count, with the violent ones, which dominate by consequence. The quiet days win the average, and the resulting number flatters the portfolio.

A more honest version is conditional. Filter both return series down to only the days the benchmark fell more than some threshold, then correlate only those. The reading is usually substantially higher than the all-period figure, because it is measuring the regime you actually care about.

Doing this for each pair you hold takes about as long as the ordinary calculation. The method is identical; only the filter changes, and the mechanics are in our guide to calculating correlation between two stocks.

It is worth doing once, mainly for the shock. Pairs that read 0.3 across the full year commonly read far higher on the days that mattered, and seeing that gap on your own holdings is more persuasive than any argument about it.

Sizing for the number you will get

If the correlation that applies under stress is higher than the one you measured, there are three honest responses, and they compound.

Cap exposure per driver, not per position. Decide the maximum loss you will accept from one economic story, then split it across every position that depends on that story. The portfolio heat guide works through the arithmetic across a full book.

Keep the driver count small enough to name. If you cannot list the distinct economic reasons your positions might fall, you do not know how many bets you hold. Six tickers resolving to two drivers is a two position portfolio with extra commission.

Assume convergence rather than hoping against it. Size as though the measured correlations will rise materially when tested. A book that survives that assumption survives the calm case comfortably. A book that only works at calm-period correlations has no margin in the one scenario the diversification was for.

Watching breadth helps as an early signal here, because correlation convergence tends to show up in internals before it shows up in a correlation matrix you recalculate monthly. Our guide to market internals covers what to watch.

Recognising it while it is happening

Convergence is not instantaneous, and it tends to announce itself before a monthly correlation matrix notices. Three things are usually visible first.

Dispersion collapses. On ordinary days, the gap between the best and worst performers in an index is wide, because company news is doing most of the work. As a shared driver takes over, that gap narrows: more names finish the day clustered around the same return. Narrowing dispersion across several sessions is often the earliest sign that individual stories have stopped mattering.

Breadth stops confirming the index. When the index holds up while the number of advancing names deteriorates underneath it, the index level is being carried by a shrinking group. That is a correlation story as much as a breadth story, and our guide to market internals covers how to read it.

Positions move together on news that should only affect one. The most practical signal of all, and it requires no data at all. If a sector-specific headline moves holdings in three unrelated sectors, the market has stopped pricing those holdings on their own merits.

None of these is a timing tool, and none tells you how far a decline will run. What they tell you is that the assumption your position sizing rests on has weakened, which is a different and more useful piece of information. The response is to reduce total exposure to the cluster rather than to predict the next move.

What does not converge

The convergence is not universal, and overstating it leads to its own error, which is concluding that diversification is pointless and concentrating instead.

What converges is assets sharing a driver. Equities broadly share one, which is why equity-only diversification degrades so sharply. Genuinely different drivers degrade less, and cash is the only holding whose behaviour does not change at all when everything else does. Sitting in cash during these periods is itself a position, and a defensible one.

The practical conclusion is narrower than “diversification fails”. It is that diversification measured inside one asset class, in calm conditions, using a single average figure, overstates itself — and it overstates itself most in exactly the week you were relying on it.