- What you measured is not what you get
- Two wires connect any two holdings
- Part of the jump is in the measurement
- One number, many possible arrangements of days
- Some things look uncorrelated because nobody priced them
- The assumptions underneath a diversification calculation
- What diversification still does, including under stress
- What buying crisis-diversification costs
- Five ways this gets misread
- Where this sits in the app
- Common questions
What you measured is not what you get
A correlation is an average taken over a set of days that have already happened. Using it as the parameter that will hold during a future stressed period assumes the thing being measured is a fixed property of the two assets. It is not. It is a property of the assets, the period, the people who own them, and the way each holding gets a price.
That is the whole argument, and the rest of this article is the mechanism behind each of the four. Start with the one most readers have never considered, because it is the one that does the damage: the owners. Two companies in unrelated industries share no revenue, no customer and no regulator. They can still share a shareholder who bought both with borrowed money, and that shareholder is a channel between them which exists nowhere in either business.
The link matters only in one direction and only sometimes, which is exactly what makes it invisible in a calm-period measurement. On an ordinary day the leveraged holder is not doing anything, so the two return series carry no trace of him. On the day his loan is called he acts on both at once. A statistic built by averaging over the ordinary days will barely register a mechanism that only operates on the unusual ones.
None of this makes diversification a mistake or a myth. The arithmetic that removes company-specific risk is set out in full in diversification: what it removes and what it cannot, and that arithmetic is unaffected by anything below. What is affected is the estimate of how much the holdings share — and since the shared part is the entire floor a diversified portfolio converges to, an estimate made in the wrong conditions is an estimate of the wrong quantity.
Two wires connect any two holdings
It helps to picture two separate wires running between any pair of things you own, because they carry different signals and they are live at different times.
The first wire is the business. Two lenders share a rate cycle; two exporters share a currency and a customer economy; two companies with the same input share a commodity. This wire is what ordinary correlation is measuring, it is stable enough to reason about, and it is what most portfolio construction is built on. It is also almost entirely what a calm window contains.
The second wire is the balance sheet of whoever holds both. It is dark most of the time and carries an enormous current when it lights up, and its trigger has nothing to do with either company. Here is the arithmetic, with chosen figures that any reader can redo. A holder owns ₹20 lakh of positions, funded with ₹8 lakh borrowed against them and ₹12 lakh of his own money — a little under 1.7 times his capital. The market falls 25%. The positions are now worth ₹15 lakh; the loan is still ₹8 lakh, because a loan does not fall; his own money is ₹7 lakh. His leverage has risen to about 2.1 times without a single decision on his part.
To get back to where he started he must sell about ₹3.3 lakh of something. And here is the part that matters: nothing selects which something. His constraint is cash by a deadline, not value. What clears a deadline is whatever has a reliable bid at a knowable price, which is a completely different property from being overvalued, poorly run or exposed to the event that started the fall.
The named failure mode is the borrowed-money bridge. A fall in one holding produces selling in a second holding that has nothing to do with it, because a third party owned both against a loan. The two return series will show a co-movement on that day, and the co-movement is real — a reader looking only at prices cannot tell it apart from the two businesses being related, and it is not there in a month when nobody is being sold out.
This is why the useful question about a holding is not what its correlation has been. It is who else owns it, on what terms, and what would force them to sell. A holding owned mainly by people who cannot be forced is connected to the rest of a portfolio by one wire. A holding that is popular as collateral is connected by two, and the second one is invisible until it is not. The market-wide version of that day, when every panel on a screen turns the same colour at once, is described in what a risk-off day looks like.
Part of the jump is in the measurement
Before blaming the market for a correlation that rose, it is worth knowing that some of the rise is guaranteed by arithmetic and would appear even if nothing whatsoever had changed. This is the least known thing in the article and the easiest to check.
Take two series whose relationship is genuinely fixed — a single correlation that holds every day, forever, with no regime and no drama. Now measure it using only the turbulent days, which is what everyone does when they ask how two things behaved "during the fall". The coefficient you compute on that subsample is higher than the one that generated the data. Not because the relationship changed. Because you selected on the outcome, and correlation is a ratio in which the denominator carries how much each series moved.
The identity, for two normally distributed series with a true correlation of ρ measured only on days selected by one of them moving a lot, is ρ ÷ √(ρ² + (1 − ρ²)·(σ²/σc²)), where σc² is how much that series varied within the selected days. Put a chosen ρ of 0.30 through it. If the selected days carry three times the ordinary variation, the measured coefficient is 0.48. If they carry five times, it is 0.58. The parameter underneath never moved.
So a coefficient computed on a stressed window and compared with one computed on a full year is not a like-for-like comparison, and the gap between them is not all signal. Some of it is a sampling effect, and the amount depends on how violent the selected days were rather than on anything about the two assets. A reader who has understood what a correlation coefficient actually measures will recognise this as the same instrument being pointed at a sample it was not designed for.
Two honest limits on that, both of which matter. The identity assumes the two series are jointly normal, that the selection was made on one of them alone, and that the underlying parameter really is constant — three assumptions that are convenient rather than true, and which is exactly the point being made about every other model in this article. And it corrects your inference, not your account: an investor whose holdings all fell together lost the money regardless of how much of the measured jump was a sampling effect. The artifact tells you what you may conclude. It does not tell you what you did not lose.
One number, many possible arrangements of days
A correlation summarises a whole joint distribution in a single figure, and any summary discards the thing it summarised. Two pairs of assets can produce the identical coefficient while behaving in completely different ways on the days that decide an outcome.
The reason is that the coefficient is an average over arrangements. It adds up one contribution per day and divides. A pair that moved together mildly on almost every day, and a pair that ignored each other for months and then moved violently together for a week, can arrive at the same total by different routes. Nothing in the number distinguishes them, and the second pair is the one that determines what a portfolio does when it matters.
This has a specific consequence for the way risk is usually reported. Most portfolio risk figures — the spread of returns, the summary statistics on a factsheet, anything built on a covariance matrix — treat the shape of the joint distribution as settled, because they have to in order to reduce it to one number. The assumption is not hidden malice; it is the price of having a number at all. What it costs is that the arrangement of days is precisely what the assumption throws away. Which of those summary figures capture what, and which do not, is the subject of the risk measures article.
The direction of the move is the other thing an ordinary coefficient averages over. It treats a day when both rose and a day when both fell as the same evidence, because the arithmetic multiplies two deviations and does not care about their sign as long as they agree. An investor does care. Wanting to know whether two holdings fall together, specifically, is a different question from wanting to know whether they move together — and the ordinary coefficient answers only the second one.
Some things look uncorrelated because nobody priced them
There is a third route to a comfortable correlation number that has nothing to do with markets and everything to do with bookkeeping, and it is the one most likely to be sitting in a real portfolio right now.
A correlation can only be computed from a return series, and a return series can only be computed from prices. When a holding is priced by an appraisal, by a valuation committee, by an infrequent transaction, or by a model, its reported price is a smoothed version of whatever its value actually did. The measured correlation with a daily-traded holding falls accordingly — and nothing about the exposure changed. Only the reporting did.
The arithmetic is short. Suppose the true returns of a holding are independent from period to period, and its reported return is the equally weighted average of this period's true return and the last one — a deliberately crude model of a stale mark, with both choices made for legibility. Then the reported series has half the variance of the true series, the part of the co-movement that lands in the current period is halved, and the measured correlation comes out at about 0.71 times the real one. Take a true correlation of 0.6 — chosen for the illustration, not a figure taken from anywhere — and it reports as roughly 0.42, so the holding looks like a better diversifier than it is by the width of a bookkeeping convention.
The named failure mode is the unmarked holding. Property, unlisted holdings, anything valued quarterly and anything whose price comes from a committee will show a flattering correlation against a daily-marked portfolio, for exactly the same reason a photograph with a long exposure shows no motion. The test is not what the coefficient says. It is whether the price was observed or estimated — and if estimated, what would happen if it had to be sold by Friday.
Note that this failure and the previous two point in opposite directions, which is why they belong in one article. Selection on turbulent days pushes a measured correlation up; smoothed pricing pushes it down. Neither is a statement about the assets. Both are statements about how the number was made, and an investor comparing two holdings on their coefficients is comparing two measurement procedures at least as much as two exposures.
The assumptions underneath a diversification calculation
Every framework in this area rests on assumptions that are known to be imperfect, and the useful skill is naming them rather than avoiding them. Here they are, with the mechanism that breaks each one and the question that replaces it. There is no threshold in the last column and no instruction — deliberately.
| The assumption | Where it comes from | What breaks it | The question it replaces |
|---|---|---|---|
| The relationship is linear | Pearson correlation measures one shape and one shape only | A channel that operates on three days in sixty is diluted by the fifty-seven quiet ones | What is the mechanism, and on which kind of day does it carry current? |
| The parameter is stable | A window estimate is used as a forward number, because there is nothing else to use | Leverage, ownership and liquidity all change; the businesses need not | Which of the four inputs — assets, period, holders, pricing — is different now? |
| The link between two holdings is a business link | The models are built on cash flows, because cash flows are what an analyst can study | A forced seller sells what has a bid, imposing a common seller on unrelated holdings | Who else owns this, with how much borrowed money, and what would force them out? |
| The sample was not chosen using the answer | The comparison of a "crisis window" against a full period feels like the obvious test | Selecting days by their size raises the coefficient mechanically, by the identity above | Would this gap appear even if nothing had changed, and how much of it? |
| Every holding is priced by a trade | A return series needs prices, and every line in a statement carries one | Appraised, modelled and infrequently traded marks are smoothed, which lowers measured co-movement | Was this price observed or estimated, and what would a forced sale realise? |
| One period, no path | Variance arithmetic is single-period; it has no notion of order | Redemptions, margin calls and rebalancing rules all act at particular moments, in sequence | What could force a sale at the worst point, and is anything committed to selling then? |
Read the second row twice, because it is the one that carries the others. A correlation matrix is treated as an input to portfolio construction in the same way a lot size or an expense ratio is — a fact to look up. It is instead an estimate with conditions attached, and the conditions are not printed next to the number. The mix decision that consumes this input is the subject of asset allocation; what the input itself is worth is this article's subject, and the two are worth keeping separate in your head.
What diversification still does, including under stress
Here is the correction to the pessimistic version of this argument, and it is a real correction rather than a softening. Diversification does not stop working in a crisis. One specific expectation about it stops being met, and the rest continues exactly as before.
Split each holding's return, as the sibling article does, into a part specific to that company and a part shared with everything else. The specific parts are independent of each other, so averaging shrinks them — and it does so on the worst day as on any other. An accounting fraud at one company, a failed product at another, a fire at a third: these do not arrive together because a funding shock happened, and averaging still cancels them. That cancellation is the whole of what diversification ever promised, and it is delivered.
What a crisis does is move material across the line. Holdings whose returns had been independent of each other acquire a shared term — the leveraged owner, the redeeming fund, the risk manager reducing everything at once — and so a portion of what had been counted as company-specific turns out to have been shared all along, under conditions that had not yet occurred. The arithmetic did not fail. The classification was wrong, and it was wrong in a way that only one kind of day can reveal.
That is a much smaller claim than "diversification fails", and it is the one the mechanism supports. It also has a consequence worth stating plainly: since the floor a diversified portfolio converges to is set by the shared part and not by the count of holdings, adding more names cannot address any of this. Past the point where the count has done its work, more holdings are more of the same shared term. The reasoning behind that floor, and why the count stops mattering early, is in the diversification article.
What buying crisis-diversification costs
Diversification chosen for how it behaves under stress is bought rather than found, and it is paid for in three separate ways.
The first is that anything held for how it behaves on the worst days is also held on all the other days. That is not a criticism; it is a definition. A holding earns whatever it earns during the long stretches when the mechanism it was chosen for is dormant, and that stretch is the great majority of the time. Whether the trade is worth making is a question about a specific investor's situation and not one this article can answer — but the drag is the price, it is paid continuously, and it is paid whether or not the protection is ever needed.
The second is that the case for it rests on a mechanism rather than a measurement. No measurement exists for conditions that are not in the sample, so an argument for a stress-diversifying holding is always an argument about how something works and never about how it did. Reasoning from mechanism is the honest approach here and it is also weaker evidence than a measurement would be, and both halves of that sentence need saying. Anyone who tells you the historical record settles it is quietly assuming that a stressed period outside the sample resembles the ones inside it, which is the stability assumption this whole article is about.
The third cost is effort and it is the one people underestimate. Knowing who owns your holdings and on what terms, whether a price was observed or estimated, and which of your positions is committed to being sold at a bad moment — none of that is on a factsheet, and all of it goes stale. The alternative is not free either. A number in a matrix asks nothing of you, and gives you a parameter measured under conditions you are not in.
Five ways this gets misread
- Treating the matrix as a property of the assets. It is a property of the assets, the period, the holders and the pricing method. Change any of the last three and the number changes with nothing on either balance sheet having moved.
- Reading a stressed window against a full period as a like-for-like comparison. Selecting days by how large they were raises the coefficient by arithmetic alone, so part of any gap is the sampling and not the market. How much depends on how violent the selected days were.
- Adding holdings in response. The count addresses the independent part, which was never the problem being described here, and it is nearly exhausted early. What stress changes is the shared part, and more names do not touch it.
- Trusting a low number from an unmarked holding. A quarterly valuation produces a flattering coefficient against a daily-marked portfolio for reasons of bookkeeping rather than economics, and the flattery disappears the moment the holding has to be sold on a timetable.
- Assuming a rebalancing rule can be executed. Rules that buy what has fallen assume a bid on the other side and cash that is not needed elsewhere at that moment. The mechanics of that, and what makes a rule survive contact with a bad week, belong to the rebalancing article.
The trade-off in the whole instrument, stated plainly: a correlation is comparable across completely different assets precisely because it has averaged over every day, thrown away the arrangement of those days, and taken no interest in who owned anything. Comparability is what you bought. Those three things are what you paid, and there is no version of the coefficient that keeps them — a reader who wants to know about the worst days specifically is asking for a different measurement rather than a better reading of this one.
Where this sits in the app
The value of a correlation view is being able to check a link you can already state in words, over more than one window, without the panel choosing the story for you.
FNOTrader's Options Analytics app carries a cross-asset panel that computes Pearson correlation of Nifty's daily returns against a fixed list of ten drivers, over a window of 30, 60 or 90 observations, using the dates each pair has in common. Each window is shown separately rather than blended, which is the feature that matters for everything above: comparing the three is how a reader sees whether a coefficient is describing a relationship or describing a fortnight. There is no stressed-period view, no conditional correlation and no portfolio-overlap score — the panel reports pairwise coefficients over the window you pick, and the interpretation stays with you.
FNOTrader is not a SEBI-registered investment adviser or research analyst. Nothing here is a recommendation about any security, fund or allocation, and nothing here is a forecast of any market or any correlation. It is the arithmetic of a measurement, the mechanism that sits under it, and the assumptions both of them rest on.
Common questions
Do correlations really rise in a crisis?
Two mechanisms say yes, and both are mechanical rather than empirical: a common forced seller adds a shared term to two return series, and measuring on a selected set of turbulent days raises the coefficient by arithmetic alone. Both are worked through above. What actually happened to any particular pair of assets in any particular market fall is an empirical question that needs a stated dataset, universe and period — and this article does not answer it, because an unattributed historical claim is exactly the kind of thing that gets copied forward for a decade.
Why would two unrelated holdings fall together?
Because a third party owned both of them with borrowed money. When his positions fall, his loan does not, so his leverage rises without any decision on his part and he has to sell something to bring it back. His constraint is cash by a deadline, not value — so what gets sold is whatever has a reliable bid, which may be the holding that had nothing to do with the fall. The two businesses were never connected. Their owner was.
Does this mean diversification does not work?
No, and the distinction is worth being precise about. Averaging still cancels whatever is genuinely specific to one company, and it does that on the worst day as on any other — a fraud at one company and a failed product at another do not arrive together because of a funding shock. What a crisis reveals is that some of what was counted as company-specific was shared all along, under conditions that had not yet occurred. The arithmetic held; the classification was wrong.
If a correlation was measured over five years, is it not reliable?
A longer window gives a more precise estimate of the average relationship over those five years. It does not make the average the right number for a period that behaves differently, and it can make things worse by burying the unusual days in a much larger pile of ordinary ones. Precision about the wrong quantity is still the wrong quantity — the length of the window addresses sampling noise, not the assumption that the parameter is stable.
Why does measuring correlation during a fall overstate it?
Because you selected the days using the very thing you are measuring. For two normally distributed series with a fixed true correlation, computing it only on days when one of them moved a lot gives a higher answer than the parameter that generated the data — the identity is in the article, and with a chosen true correlation of 0.30, days carrying three times the ordinary variation return 0.48. Nothing changed. The sample was picked on the outcome.
My property or unlisted holding shows almost no correlation with equities. Is that real?
Check how it was priced before treating it as real. A correlation can only be computed from a return series, and a return series needs prices — so a holding valued by appraisal, by committee or by an infrequent transaction has a smoothed series, and smoothing lowers measured co-movement mechanically. With a crude two-period average the measured figure comes out at about 0.71 times the true one, so a true correlation of 0.6 — chosen for the illustration, not a figure from anywhere — reports as roughly 0.42. The exposure is unchanged; only the reporting is.
Would holding more stocks help?
Not with this. The count addresses the part of risk that is independent across holdings, and its contribution is nearly used up early, which the diversification article works through in full. What stress changes is the part every holding shares, and the floor a large portfolio converges to is set by that shared part rather than by how many names sit above it. Adding the twenty-first holding buys more of the same shared term.
What should I look at instead of the correlation number?
The questions that replace it are in the table above, and they are questions rather than a rule: what is the mechanism connecting these two, and on which kind of day does it operate; who else owns this and with how much borrowed money; was this price observed or estimated; and what could force a sale at the worst moment. None of those has a threshold to clear, which is the honest answer — the number felt like an answer partly because it was a number.
Is correlation the same as beta?
No, and confusing them is common. Correlation measures how consistently two series moved in the same direction and says nothing about size; beta is a slope and reports how much one moves per unit of the other. Two holdings can share a correlation of 0.8 while one of them barely moves and the other moves violently. Everything in this article about a coefficient's assumptions applies to a beta estimated over the same window, since it is estimated from the same days.
Continue reading
More in Advanced Investing · App: Options Analytics · Definitions: glossary · Free tools: calculators · All: every article