Coskewness and Cokurtosis

You build a portfolio because two assets have low correlation on ordinary days, so they diversify each other. Then a crisis hits and both assets crash on the same day, correlation you measured in calm markets simply vanished when you needed it. This is not a bug in your correlation estimate; it is a limit of what correlation can measure. Correlation is built from squared co-movements, which weight every day equally. It cannot tell you that assets tend to move together specifically on the bad days. Coskewness and cokurtosis are the tools built to catch exactly that.

The analogy

Picture two coworkers rated on how well they cover for each other. "Correlation" would ask: on a typical day, when one is busy, is the other also busy? A low score there sounds reassuring. But the real question a manager cares about is: on the day everything breaks at once, do they both vanish simultaneously? Two coworkers can look independent on ordinary Tuesdays and still both call in sick during the one week the office actually needs them. Coskewness and cokurtosis ask about that specific week, not the average Tuesday.

Building the idea from ordinary moments

Recall that an asset's own skewness measures whether its bad days are worse than its good days are good (negative skew = nasty left tail), and its kurtosis measures how fat its tails are relative to a normal bell curve. Both are single-asset ideas. Coskewness and cokurtosis extend them to pairs.

Coskewness of assets $i$ and $j$ with respect to $i$'s cube term is

$S_{i,ij} = \mathbb{E}\!\left[(R_i - \mu_i)^2(R_j - \mu_j)\right].$

In words: take asset $i$'s squared deviation (always positive, big on $i$'s extreme days regardless of direction) and multiply it by asset $j$'s deviation (positive if $j$ is up, negative if $j$ is down). Average over time. If $j$ tends to be down precisely on the days $i$ moves to an extreme, this product is negative — meaning the pair has bad coskewness, $j$ offers no shelter exactly when $i$'s risk is largest.

Cokurtosis goes one order further, replacing the square with a fourth power on one side:

$K_{i,ijj} = \mathbb{E}\!\left[(R_i - \mu_i)(R_j - \mu_j)^3\right].$

In words: it asks whether asset $i$'s deviations line up with asset $j$'s most extreme moves specifically (the cubed term amplifies $j$'s tail events far more than its typical days). High cokurtosis means the two assets are joined at the tail even if their ordinary-day correlation looks modest.

Set the correlation low here and look at the cloud of points. Ordinary correlation only cares about the overall tilt of that cloud. Coskewness and cokurtosis specifically care about the handful of points sitting in the far corners, and whether those corner points line up (both assets extreme together) even while the bulk of the cloud looks nicely scattered.

Worked example 1: coskewness by hand

Three days of returns for asset $i$ (equity) and asset $j$ (a "hedge"): Day 1: $i=+1\%, j=+1\%$. Day 2: $i=+1\%, j=-1\%$. Day 3 (a crash): $i=-8\%, j=-6\%$. Means: $\mu_i \approx -2\%$, $\mu_j \approx -2\%$.

Deviations: Day 1: $i: 3, j: 3$. Day 2: $i: 3, j: -3$. Day 3: $i: -6, j: -4$ (in percentage points).

$S_{i,ij}$ terms: Day 1: $3^2 \times 3 = 27$. Day 2: $3^2 \times (-3) = -27$. Day 3: $(-6)^2 \times (-4) = -144$. Average: $(27 - 27 -144)/3 = -48$.

The strongly negative value is driven entirely by Day 3: $i$'s squared deviation is huge (it is an extreme day for $i$) and $j$ is also negative, so the hedge failed exactly on the day it was needed. Days 1 and 2 nearly cancel because $j$'s sign flipped while $i$'s magnitude stayed modest — ordinary days contribute almost nothing, which is the entire point of the statistic.

Worked example 2: same correlation, different coskewness

Consider two candidate hedges for the same equity book, both with correlation roughly $-0.3$ to equities over a full sample. Hedge A loses money steadily in calm up-markets and gains steadily in calm down-markets. Hedge B is flat on ordinary down days but spikes hard on the single worst crash day, more than making up for its mediocre showing elsewhere. Correlation comes out identical for A and B, but coskewness with equity's squared crash-day deviation is far more favorable for B: its gains cluster exactly where $i$'s squared deviations are largest. A manager choosing purely on correlation would be indifferent; one checking coskewness would strongly prefer B as a genuine tail hedge.

What this means in practice

Multi-asset risk models that stop at a covariance matrix implicitly assume co-movement is symmetric and tail-blind, the same assumption behind 2008-era "diversified" books that collapsed together. Desks stress-testing a book look at conditional correlation (correlation computed only on the worst 5% of days) as a cheaper proxy for the same idea; see Tail Dependence Coefficients for the copula-based version.