4.12.1 Copulas from Zero

Basic Prerequisites: Probability & Measure Theory (joint vs marginal distributions).

---

1. Intuition & Practical Objective

This page builds the why of copulas with no prior copula knowledge needed. The objective is one idea: the risk of a portfolio is not the sum of the risks of its positions; it is the sum plus a term that depends entirely on how the positions move together - and that term is a separate modelling choice you have to make, not something the individual risks tell you.

Start with the dumbest question: two positions, each with the same loss distribution. Is the portfolio risk determined? The beginner says yes: “just add the VaRs.” The beginner is wrong. Consider two assets, each with a standard-normal return, in an equal-weighted book. The book’s loss is $-(A+B)/2$. Change only the dependence between $A$ and $B$:

1. Independent - $A$ and $B$ are unrelated. The book’s volatility falls to $1/\sqrt{2}$ if each has unit volatility: diversification, the thing every textbook promises.
2. Comonotone - $B=A$ (perfect positive dependence). There is no diversification at all; the book is one asset in disguise.
3. Countermonotone - $B=-A$ (perfect negative dependence). The book is riskless; the two positions always cancel.

Same two marginals. Same dollars. The $95\%$ VaR of the book goes from $0$ to $1.64$ depending on which of these is true. That gap is the whole subject.

Three steps.

1. Marginals are not enough. Knowing each position’s loss distribution leaves the portfolio loss distribution undetermined. Dependence is additional information.
2. Dependence is a function of the ranks, not the levels. What matters is not how big each loss is but how their percentiles line up. This is why rank correlation, not Pearson correlation, is the natural measure, and why a copula - a function on the unit square of percentiles - is the right object.
3. Diversification is a copula statement. “Diversification reduces risk” is true for the independence copula and false for the comonotone one. There is no diversification number that holds for all dependence structures with the same marginals - only a range (the Fréchet bounds).

---

2. Mathematical Ground Truth & Derivations

Setup. Let two positions have returns $A,B$ with identical standard-normal marginals, $A,B\sim N(0,1)$. The equal-weighted portfolio loss is $L=-(A+B)/2$ and its $95\%$ VaR is the $0.95$-quantile of $L$.

Case 1 - independence. $\mathrm{Var}(A+B)=2$ and $\mathrm{Var}(L)=\frac{1}{2}$, so $L\sim N(0,\frac{1}{2})$ and

$${\mathrm{VaR}}_{0.95}(L)=z_{0.95}/\sqrt{2}=1.6449/1.4142=\mathrm{1.1631.}$$

Case 2 - comonotone ($B=A$). $L=-A$, so $L\sim N(0,1)$ and ${\mathrm{VaR}}_{0.95}=1.6449$.

Case 3 - countermonotone ($B=-A$). $L=0$ exactly; ${\mathrm{VaR}}_{0.95}=0$.

So with the same marginals the book VaR spans $[0,1.6449]$. Nothing about $A$ or $B$ individually told us which one applies.

The general statement (Fréchet). For any joint distribution $F$ of $d$ positions with margins $F_i$, writing $u_i=F_i(x_i)$, the copula $C$ that links them is bounded:

$$\max (\sum _{i=1}^d{u}_i+1-d,0)\le C(u_1,\dots ,u_d)\le \min (u_1,\dots ,u_d).$$

The lower bound is the countermonotone copula $W$ (positions perfectly negatively aligned), the upper bound the comonotone copula $M$ (perfectly positively aligned; $M(u)=\min u_i$). Every portfolio risk measure therefore lies in a band whose width is exactly the dependence uncertainty:

$$\text{risk}(W)\le \text{risk}(C)\le \text{risk}(M)\text{(for risk measures monotone in the concordance order, e.g. ES; plain VaR is not generally so).}$$

Why “ranks, not levels”. The copula is invariant under strictly increasing transformations of the margins: if you replace a loss in dollars by its logarithm, or a return by its rank, the copula does not change. Pearson correlation does change. That is the first clue that the correct dependence input is a rank quantity.

The portfolio loss is a copula functional. For a generic $d$-asset book with loss $L={\sum }_i{w}_i{\ell }_i$, everything about $\mathbb{E}[L]$ is marginal, but everything about the shape of $L$‘s distribution - skew, kurtosis, and every quantile above the mean - is a functional of the copula. That is why two books with identical marginals can have VaR $1.16$ and $1.64$.

---

3. Computational Implementation - same marginals, three risks

Stdlib only. We sample the three dependence structures explicitly and measure the book’s $95\%$ VaR and Expected Shortfall; the simulation is checked against the analytic values of §2.

The simulation reproduces the analytic VaR to three decimals. The three books are indistinguishable position-by-position and differ in portfolio VaR by a factor of $\infty$ (from $0$ to $1.64$). No amount of marginal analysis could have told the risk manager which book she held - that information lives entirely in the copula.

---

4. Failure Modes & First-Principles Breakdowns

1. “VaR is additive.” Only under the comonotone copula. Adding position VaRs is the upper Fréchet bound, not an estimate; using it is a deliberate worst-case assumption, not a neutral default.
2. Pearson correlation as a dependence model. Two variables can have Pearson $\rho =0$ and be strongly dependent (e.g. $B=A^2$), and Pearson $\rho$ is not invariant under monotone rescaling - so it cannot be a copula parameter. Rank correlations (Kendall, Spearman) are the invariant measures and are the correct calibration input (page 02).
3. The “diversification number” illusion. Reporting a single diversification benefit implicitly assumes independence. Independence is one copula among a continuum; a risk report that does not state its dependence assumption has hidden the largest assumption it makes.
4. Tail vs centre. In the centre of the distribution, independence is a benign assumption. In the tail - where VaR and ES live - the same marginals can produce joint losses an order of magnitude apart. Dependence matters most exactly where you are least able to check it. This is the seed of 04 · Tail Dependence.

---

5. References

• McNeil, Frey & Embrechts (2015), Quantitative Risk Management
• Embrechts, McNeil & Straumann (2002), Correlation and Dependence in Risk Management: Properties and Pitfalls
• Nelsen, Roger B. (2006), An Introduction to Copulas, 2nd ed.
• Embrechts, Klüppelberg & Mikosch (1997), Modelling Extremal Events
• Hull (11th ed.), Options, Futures, and Other Derivatives

---

6. Connected Graph Bridges

• Base: Probability & Measure Theory · Statistics & Inference
• Continue: 02 · Sklar’s Theorem & Copulas · Index Hub
• Related risk: VaR & Expected Shortfall (the subadditivity question this page makes concrete) · Parametric, Historical & Monte Carlo VaR