A bank does not lose money on one risk at a time. Loans default together in a recession, insurance lines take the same storm, and markets fall together in a crisis. A single correlation number cannot say how often the worst outcomes coincide. A copula can: it separates how each risk behaves alone from how they move together.
Two borrowers each have a 5% chance of defaulting this year. That is all the marginalMarginal distribution: the distribution of one risk on its own, ignoring the other. distributions say. The joint distributionJoint distribution: the probabilities of two (or more) risks taking values together, F(x, y) = P(X ≤ x, Y ≤ y). adds how they default together: here both default with probability 2%.
B defaults
B survives
Total (A's margin)
A defaults
2%
3%
5%
A survives
3%
92%
95%
Total (B's margin)
5%
95%
100%
Read the table
Guess, then reveal
If the two were independent, how likely would a joint default be?
A has defaulted. How likely is B to default now?
Could the same two margins come with a different joint table?
The rules
Joint, marginal, conditional. The three views of a pair, and independence.
The probability integral transform. Any continuous risk becomes a uniform through its own cdfProbability integral transform: for a continuous X with cdf F, U = F(X) is uniform on (0, 1); and F⁻¹(U) has cdf F..
Sklar's theorem and two copulas. A copulaCopula: a joint distribution of variables that are each uniform on (0, 1). It describes dependence with the margins removed. is the joint distribution of the ranks; Sklar's theoremSklar's theorem: any joint cdf F(x, y) equals C(F_X(x), F_Y(y)) for some copula C, unique when the margins are continuous. glues it to any margins. The GaussianGaussian copula: the dependence structure of a bivariate normal with correlation ρ. and tt copula: the dependence structure of a bivariate Student t with correlation ρ and ν degrees of freedom. copulas are the workhorses.
Tail dependence. Do the extremes come togetherTail dependence λ: the limit of P(V ≤ q | U ≤ q) as q → 0 (lower) or the matching limit in the upper corner. Zero for the Gaussian copula with ρ < 1, positive for the t copula.?
F is the joint cdf and f the joint density of (X, Y); FX, FY are the marginal cdfs; C is a copula, a joint cdf on [0, 1]² with uniform margins, unique when the margins are continuous; Φρ and tρ,ν are the bivariate standard normal and Student t cdfs with correlation ρ (ν degrees of freedom); Φ⁻¹ and tν⁻¹ are the univariate quantile functions; U = FX(X), V = FY(Y). Sources: Sklar (1959); Nelsen (2006), chapter 2; McNeil, Frey and Embrechts (2015); Demarta and McNeil (2005).
Try it
The same 1,000 normal pairs, turned into a Gaussian copula and into a t copula with the same ρ. Count the pairs in the shaded corners, where both risks are in their worst (or best) q. Then lower ν, or raise ρ, and watch the t corners fill up.
Try itSame correlation, different tails
t copula degrees of freedom ν
Tail corner q
seed 1
Gaussian copula, ρ = 0.5t copula, ρ = 0.5, ν = 5
1,000 draws of (U, V), each uniform on its own. Shaded: both below q (red, joint losses in the lower tail) and both above 1 − q (green). The t draws use the same normal pairs as the Gaussian ones.
Expected per corner if independent, n q²2.5
Gaussian: lower corner7
Gaussian: upper corner14
t: lower corner15
t: upper corner21
Gaussian: P(V < q | U < q)14%
t: P(V < q | U < q)30%
Limit as q → 0, Gaussian0
Limit as q → 0, t (tail dependence)0.207
Worked examples
1. Two borrowers
Step 1 of 2
1
Margins from the joint
Add across a row or down a column: P(A defaults) = 2% + 3% = 5%. The margins are the table's edges.
2. Ranks there and back
Step 1 of 4
1
A loss to a rank
Line A's yearly loss is exponential with mean €2m. A loss of €4.6m has rank U = 1 − e−4.6/2 = 0.8997: about its 90th percentile.
3. Same correlation, different joint tails
Step 1 of 3
1
The experiment
10,000 seeded draws from a Gaussian copula and from a t copula with ν = 4, both with ρ = 0.5. Count the pairs where both are below their 1% quantile.
Common mistakes
Each line below is a tempting slip. Spot what is wrong, then check.
Reaction
Check what stuck
Question 1 · objective 1
P(A) = 0.1 and P(A and B) = 0.04. What is P(B | A)?
Question 2 · objective 1
Two events are independent, with P(A) = 0.1 and P(B) = 0.2. What is P(A and B)?
Question 3 · objective 2
X is exponential with mean 1, so F(x) = 1 − e⁻ˣ. What is U = F(X) when X = ln 4?
Question 4 · objective 4
With 2,000 independent pairs of uniforms, how many would you expect to fall in the corner where both are below 0.05?
Question 5 · objective 4
A Gaussian and a t copula have the same correlation of 0.5. What differs?
Glossary
Joint distribution
Joint distribution: the probabilities of two (or more) risks taking values together, F(x, y) = P(X ≤ x, Y ≤ y).
Marginal distribution
Marginal distribution: the distribution of one risk on its own, ignoring the other.
Conditional distribution
Conditional distribution: the distribution of one risk given a value or event of the other.
Independent
Independent: the joint distribution is the product of the margins, so one risk says nothing about the other.
Probability integral transform
Probability integral transform: for a continuous X with cdf F, U = F(X) is uniform on (0, 1); and F⁻¹(U) has cdf F.
Copula
Copula: a joint distribution of variables that are each uniform on (0, 1). It describes dependence with the margins removed.
Sklar's theorem
Sklar's theorem: any joint cdf F(x, y) equals C(F_X(x), F_Y(y)) for some copula C, unique when the margins are continuous.
Gaussian copula
Gaussian copula: the dependence structure of a bivariate normal with correlation ρ.
t copula
t copula: the dependence structure of a bivariate Student t with correlation ρ and ν degrees of freedom.
Tail dependence
Tail dependence λ: the limit of P(V ≤ q | U ≤ q) as q → 0 (lower) or the matching limit in the upper corner. Zero for the Gaussian copula with ρ < 1, positive for the t copula.
Nelsen, An Introduction to Copulas: Archimedean copulas (Clayton, Gumbel, Frank) and measures of dependence such as Kendall's tau. Advanced.
Origins
Abe Sklar, “Fonctions de répartition à n dimensions et leurs marges”, Publications de l'Institut de Statistique de l'Université de Paris 8, 229–231, 1959.
Roger B. Nelsen, An Introduction to Copulas, 2nd edition, Springer Series in Statistics, Springer, 2006 (chapter 2, definitions and basic properties; chapter 5, dependence).
Alexander J. McNeil, Rüdiger Frey and Paul Embrechts, Quantitative Risk Management: Concepts, Techniques and Tools, revised edition, Princeton University Press, 2015.
Stefano Demarta and Alexander J. McNeil, “The t copula and related copulas”, International Statistical Review 73(1), 111–129, 2005.