Introduction to copulas

Welcome to the World of Copulas!

If you have ever wondered how actuaries model two or more risks that seem to happen at the same time—like a stock market crash happening simultaneously with a spike in insurance claims—you are in the right place. In this chapter, we explore copulas.

Don't worry if the name sounds a bit strange! At its heart, a copula is just a mathematical "glue" that sticks individual risks together to show how they behave as a group. By the end of these notes, you will understand how to separate the behavior of individual risks from the way they interact with each other.

1. The Big Idea: Marginals vs. Joint Distributions

Before we dive into the math, let’s use an analogy. Imagine you are making a sandwich.
• The Marginal Distributions are your ingredients: the bread, the cheese, and the ham. You know everything about them individually (their calories, their weight).
• The Copula is the "recipe" or the "glue" that decides how you stack them. It describes the dependence structure between the ingredients.

In actuarial work, we often know how individual risks behave (e.g., the probability of a car accident), but the tricky part is knowing how those risks move together. Copulas allow us to model the individual parts (the marginals) and the relationship (the copula) separately.

Quick Review: What you need to remember

1. Marginal Distribution: The probability distribution of a single random variable, ignored by all others.
2. Joint Distribution: The probability distribution that describes two or more variables happening at the same time.
3. Uniform Distribution: Copulas specifically work with variables that have been transformed into a Uniform \( (0,1) \) distribution.

2. What exactly is a Copula?

A copula is a joint cumulative distribution function (CDF) where every single marginal distribution is a Uniform distribution on the interval [0, 1].

Mathematically, for \( d \) variables, a copula is a function \( C \):
\( C(u_1, u_2, ..., u_d) = P(U_1 \le u_1, U_2 \le u_2, ..., U_d \le u_d) \)
where each \( U_i \) follows a \( U(0, 1) \) distribution.

Why do we use \( U(0,1) \)?

Did you know? Every continuous random variable \( X \) with a CDF \( F(X) \) can be turned into a Uniform \( (0,1) \) variable by simply plugging it into its own CDF! This is called the Probability Integral Transform. This trick is the "secret sauce" that lets copulas work with any type of data, whether it's normal, exponential, or something else entirely.

3. Sklar’s Theorem: The "Magic" Formula

Sklar’s Theorem is the most important concept in this chapter. It states that any joint distribution function can be written in terms of its marginals and a copula.

For two variables \( X \) and \( Y \), with joint CDF \( F(x, y) \) and marginals \( F_X(x) \) and \( F_Y(y) \):
\( F(x, y) = C(F_X(x), F_Y(y)) \)

What does this mean in plain English?
It means if you want to find the probability that \( X \le x \) AND \( Y \le y \), you just need to:
1. Find the individual probabilities (the marginals).
2. "Plug" them into the copula function \( C \).

Key Takeaway: Sklar's Theorem allows us to decompose a complex multivariate distribution into two independent parts: the individual behaviors and the dependence structure.

4. Fundamental Properties of Copulas

To be a valid copula, the function \( C(u, v) \) must follow these rules:

1. The Grounding Property: If the probability of one event is zero, the joint probability is zero.
\( C(u, 0) = 0 \) and \( C(0, v) = 0 \)
Example: If the chance of it raining is 0%, the chance of it raining AND you losing your umbrella is also 0%.

2. The Marginal Property: If one event is "certain" (probability = 1), the joint probability is just the probability of the other event.
\( C(u, 1) = u \) and \( C(1, v) = v \)

3. Monotonicity: The function must be non-decreasing in each variable. As the individual probabilities increase, the joint probability can't go down.

5. Three Special "Boundary" Copulas

There are three specific copulas you must know. They represent the extremes of how variables can relate to each other.

A. The Independence Copula (Product Copula)

This is used when the variables have no relationship at all. If \( X \) and \( Y \) are independent, their copula is simply the product of their probabilities:
\( C(u, v) = u \times v \)
Mnemonic: "Independent? Just multiply!"

B. The Upper Fréchet-Hoeffding Bound (Minimum Copula)

This represents perfect positive dependence (Comonotonicity). This happens when one variable is a strictly increasing function of the other.
\( C(u, v) = \min(u, v) \)
Analogy: Imagine a tandem bicycle. If the front person moves 1 meter, the back person must move exactly 1 meter. They are perfectly synced.

C. The Lower Fréchet-Hoeffding Bound

This represents perfect negative dependence (Countermonotonicity). When one variable goes up, the other goes down as much as possible.
\( C(u, v) = \max(u + v - 1, 0) \)
Analogy: A seesaw. When one side is at the very top, the other side must be at the very bottom.

6. Common Mistakes to Avoid

• Mixing up \( f(x) \) and \( F(x) \): Remember, copulas use the Cumulative Distribution Function (CDF), usually denoted by a capital \( F \), not the probability density function (small \( f \)).
• Thinking Correlation = Copula: Correlation (like Pearson's \( r \)) is just a single number. A copula is a whole function. Copulas capture much more detail, especially in the "tails" (extreme events).
• Ignoring the Bounds: Any copula you calculate must stay between the Lower and Upper Fréchet-Hoeffding bounds. If your result is larger than \( \min(u, v) \), you've made a mistake!

7. Summary and Final Check

Summary of Chapter 11:
• Definition: A copula is a joint CDF with Uniform marginals.
• Sklar's Theorem: It links joint distributions to marginals via \( F(x,y) = C(F_X(x), F_Y(y)) \).
• Why we care: It lets us model risks that are dependent on each other without forcing the marginals to be the same shape.
• The Bounds: \( \max(u+v-1, 0) \le C(u, v) \le \min(u, v) \).

Don't worry if the formulas for specific copula families (like Clayton or Gumbel) look intimidating when you see them later. Just remember that they are all just different "recipes" for how variables can be glued together!