Correlation Betrays You in Crises

Diversification is counted in drivers

Holding more names reduces single-name risk and says nothing about whether the risks are independent. That distinction is the whole lesson. Twelve semiconductor companies share a demand cycle, a rate sensitivity and a set of customers, so a shock to any of those moves all twelve; twelve names across utilities, healthcare, energy and materials share far less. The number of tickers is a count of positions, and the number of **drivers** is a count of the ways the portfolio can lose. The effective-bet formula turns that into a number: with N equally weighted names and an average pairwise correlation ρ, the portfolio behaves like N ÷ (1 + (N − 1)·ρ) independent bets. At N = 12 and ρ = 0.75 that is 1.3. Push the correlation down to 0.25 and it becomes 12 ÷ (1 + 2.75) = 3.2; at a correlation of 0.1 it is 5.5. The formula is worth internalising because the answer is so much smaller than the count: it says that a portfolio of a dozen correlated names is a portfolio of one or two ideas, and it explains why adding a thirteenth name of the same kind changes almost nothing. There is a second half to the argument, which is what happens in a crisis. Correlation is regime-dependent: it is lowest in quiet markets and rises towards one under stress, because a general flight from risky assets is a sale of everything at once. Assets that look diversifying on calm-market data therefore deliver least diversification exactly when the portfolio needs it most, and the crisis is the moment the whole calculation is being relied on. The honest response is to diversify across drivers that are structurally different — different demand cycles, different rate sensitivities, different geographies — and to hold something genuinely uncorrelated or defensive for the tail rather than assuming the existing names will behave differently than they did in 2008. The same twelve names, three correlations — ρ = 0.75: 1.3 effective bets · ρ = 0.25: 3.2 effective bets · ρ = 0.10: 5.5 effective bets ← · What the count of names says: 12 — and it is the wrong number The formula assumes equal weights and a common correlation, which is a simplification. It is used here because the direction and rough size of the correction are what matter: the count of holdings is an upper bound on diversification, never the answer.

Caps are what make the arithmetic hold

If a cluster can cost a fifth of the portfolio on one view, the portfolio needs a limit on how much of that view it can hold. Two caps do most of the work. A **single-name cap** — commonly 10 to 20% of capital — keeps one company from deciding the outcome; at $500,000 a 10% cap is $50,000, which is a size a household can lose without a change of plan. A **cluster cap** — often 25 to 35% per sector or theme — keeps a single driver from doing the same thing through several tickers, and it is the cap that the twelve-name portfolio above fails. The distinction between the two matters because the single-name cap is the one people apply and the cluster cap is the one that binds. A portfolio can hold twelve names and pass every individual limit while three of them carry half the book, which is the same concentration as owning one position and calling it diversified. Auditing a portfolio therefore means sorting by driver and adding, which is a different operation from checking each position — and it is the reason the caps have to be checked on the total rather than on the largest holding. A third element, less discussed and worth stating: concentration is a legitimate choice as long as it is managed as one. An investor who genuinely wants to hold a large position in a single business can do so, but then the rest of the book should be sized as the hedge and the total exposure to that idea should be pre-agreed — which is a plan rather than a position limit. What does not work is a concentrated bet that is not recognised as one, because the sizing, the stops and the drawdown expectations are all computed on the assumption that the portfolio’s risk is spread. The cluster in the workbench — Semiconductor cluster: 52% of $500,000 = $260,000 · Loss if it falls 40%: $104,000 — 20.8% of the portfolio ← · Sold to cap the cluster at 35%: $85,000 · One name at a 10% cap: $50,000 Thematically labelled positions are the usual way a cluster hides: an AI fund, a chipmaker, a data-centre REIT and a power utility can be four tickers in one trade without sharing a sector name.

Correlation describes the middle, and it is noisy

Correlation is a single summary of a relationship across the whole range of outcomes, and the range you care about is not the whole range. Two assets can be almost uncorrelated in ordinary months and move together almost perfectly in the worst one, because independence in the middle of a distribution says nothing about the joints in the tails. That asymmetry has a name — **tail dependence** — and it is the reason a diversified book can look well behaved for years and then behave like one position in a single week. The mechanism is usually not statistical at all: in a crisis, positions get sold to raise cash, and what gets sold is what can be sold, which means everything falls together regardless of what it was built to track. The second problem is estimation error, and it is worse than most people assume. A correlation is computed from a sample, and the sample estimate of a correlation is one of the least stable statistics there is: with sixty overlapping observations, the confidence interval around an estimated correlation is wide enough that 0.4 and 0.8 cannot be reliably told apart. Overlapping windows make it worse, because daily returns are not independent, and using a short window makes the estimate respond to noise while using a long one makes it respond slowly to genuine change. The practical response is not to find a better correlation but to stop treating the point estimate as a fact: **assume correlations rise in the states that hurt you**, and size the book on the assumption rather than on the measured average. There is one more trap worth naming, because it is what makes the counted-drivers exercise in this lesson a discipline rather than a formality. Correlation is computed from **prices**, and two positions can be highly correlated in their price history while depending on entirely different drivers that happen to have moved together for a period — or uncorrelated in price while sharing a single driver, because one of them is hedged. Grouping positions by what they are (asset class, sector, ticker) is the fastest method and the least reliable one. Grouping them by what would have to happen for each to lose money is slower, requires thinking, and is the exercise that actually protects the book. • Correlation summarises the middle of the distribution and is weakest exactly in the tail. • A sample correlation is unstable: short windows are noisy, long windows are stale. • Assume correlations rise in the states that hurt you, and size on that assumption. • Group positions by what would have to happen for each to lose, not by asset class label. A correlation matrix produced by a tool is a description of the past in a particular regime. Using it as an input assumes the regime persists, and the assumption is least reliable exactly when the matrix is most comforting.

Estimating the number, and finding who actually carries the risk

The effective-bet formula takes the correlation as an input, and the input is an estimate — which means the number that decides how diversified a book is comes with an error bar that is usually ignored. Three properties of the estimate matter. First, **the window**: a correlation computed over sixty daily observations can differ from the same pair’s correlation over five years by a wide margin, and the shorter window is far more responsive to what has happened lately, which is exactly the regime-dependence the previous read described. Second, **a handful of days dominate**: correlation is a covariance divided by two standard deviations, and covariance is an average of products, so the largest joint moves in the sample carry disproportionate weight. A pair with one shared crash day in the window will report a meaningful positive correlation even if the other fifty-nine days were independent, which is how two genuinely diversifying assets can look correlated on paper. Third, the estimate is **biased in the direction of the sample**: a period with no crisis produces a confident low correlation, and the low correlation is a statement about the period as much as about the assets. Two distinctions prevent the most common misreading. The first is between **correlation and beta**: correlation measures how tightly two series move together on a standardised basis, while beta measures how much one moves for a given move in the other. An asset can be weakly correlated with the market and highly sensitive to it — a small-cap stock with idiosyncratic news has a low correlation and a beta above one — and the risk that matters for the portfolio is closer to the beta. A book of low-correlation high-beta names is not diversified in any useful sense, and the effective-bet formula will not warn you because it is asked only about correlation. The second is that the right input for a portfolio is not the pairwise correlations but their interaction: a new position reduces risk by an amount that depends on its correlation with **what is already in the book**, not with each name individually, so adding a name that is uncorrelated with most holdings but highly correlated with the largest one adds less than the average pairwise number suggests. The tool that gets at the interaction directly is the **risk contribution** decomposition. Instead of asking how many names there are, it asks how much of the portfolio’s total volatility each position accounts for, where the answer depends on the position’s size, its own volatility, and its average correlation with the rest of the book: the marginal contribution is the position weight multiplied by its covariance with the portfolio, and the sum of the contributions equals the portfolio’s variance. Worked on the earlier example — twelve names at a 0.75 average pairwise correlation, one of them at 40% of the book — the largest position can easily account for more than half of the portfolio’s risk while the smallest accounts for under a percent, so the question “how many positions do I hold” is answered by the count and the question “how many risks do I hold” is answered by the decomposition. They rarely give the same answer, and the decomposition is the one that describes what a bad month will do. Two questions, one book — How many positions?: Twelve — the count, and the number most people report · How many effective bets?: N ÷ (1 + (N − 1)ρ) ≈ 1.3 — the driver count ← · Who carries the risk?: Weight × covariance with the portfolio; the 40% position is often over half the risk ← · Correlation versus beta: Tightness of co-movement versus sensitivity — a low-correlation name can still be high-beta The estimation warnings and the decomposition point the same way: the only robust conclusion is comparative. A correlation estimated on a calm sixty-day window is a lower bound on the crisis value, and a risk decomposition is a description of the last sample rather than a forecast. Both are useful for choosing what to trim, and neither is precise enough to justify a position that only works if the number is right.

Sources

• Effective number of bets and portfolio concentrationMeucci, "Managing Diversification" (2009)
• Correlation convergence in crisesLongin & Solnik on extreme correlation; 2008 and March 2020 evidence
• Concentration as a risk decision rather than a conviction signalDamodaran, "Investment Valuation" (portfolio chapters)
• Why 2022 failed the stock-and-bond diversification assumptionStandard 2022 asset-class return data; CFA Institute commentary