A Risqué Case of using Correlation to measure Dependence

Posted by: tanya Comments: 0

Data science is science deployed in a heuristic manner. The manner in which it is deployed gives different answers to the same data set. This essay will focus primarily on one of the most popular yet misinterpreted measures of dependence viz. correlation.

As per the Markowitz portfolio theory, risk is minimized when one invests in asset classes that show a low level of correlation in relation to each other thereby helping to smoothen out volatility across asset returns*. Co integration on the other hand would assist in determining long-term trends in these markets with or without correlation. Thus, while correlation talks in relation to short term movements, co integration highlights how the presence of co integration makes difficult a risk diversification exercise.

Stochastic dependence refers to the relationship in distribution that may exist between two or more random variables. For random variables to be dependent, they will fail to satisfy the property of stochastic independence. Thus if X1 and X2are random variables, then they are independent if and only if, for all x1 and x2 ε R,

P(X1≤ x1, X2≤ x2)=P(X1≤ x1 ) P(X2≤ x2)

When the above definition doesn’t hold, the random variables are said to be dependent. Since the above property of stochastic dependence is a rather broad definition, an all encompassing measure would have to be of a non-parametric nature and most measures are only able to capture a portion of this dependence. Most Copulas simplify this non-parametric question to more manageable proportions by assuming a parametric model to describe the dependence structure. Let’s examine some basic properties for dependence measures to examine which measures viz. Pearson’s or linear correlation, Spearman’s rank correlation and tail dependence respectively, meet these conditions.

Property 1: Comonotonicity and Countermonotonicity

A monotonic function is a function between ordered sets (a sequence that is distinguished from other sequences of the same element by the order of the elements) that preserves the given order.

Comonotonicity refers to perfect positive dependence between the components of a random vector, essentially saying that they can be represented as increasing functions of a single random variable.

δ(X,Y)=1X,Y

X=f(Z),Y=g(Z) a.s., where f and g are two increasing real-valued functions

Applicability in our case would imply that if, in particular, the sum of the components X1+X2+…Xn is the riskiest if the joint probability distribution of the random vector (X1+X2+…Xn) is comonotonic. Furthermore, the α-quantile of the sum equals of the sum of the α-quantiles of its components, hence comonotonic random variables are quantile-additive. This property is an extension of the concept of (perfect) positive correlation to variables with arbitrary distributions.

Countermonotonicity implies X and – Y are comonotonic.

δ(X,Y)=-1X,Y

X=f(Z),-Y=g(Z) a.s., where f and g are two increasing real-valued functions


Property 2: Symmetry
δ(X,Y)=δ(Y,X)

Property 3: Boundedness
-1≤δ(X,Y)≤1

Property 4:
For h an strictly monotonic function on the range of X:

δ(h(X),Y)=f(x)={δ(X,Y)if h increasing,

{-δ(X,Y)if h decreasing.

Property 5:

δ(X,Y)= 0X,Y are independent

Property 4 and 5 are contradictory

Established risk measures like Pearson’s product moment correlation coefficient and Spearman’s rank order correlation coefficient are controlled by small movements around the mean and thus fail to describe dependence between extreme events. Facing asset selection and allocation with respect to portfolio management extreme events are primarily represented by jump risks and default risks. The target thus is to diversify away extreme risks by minimizing extreme dependence between assets within a portfolio.

*The underlying assumption is of asset returns being explained by first and second moments of dispersion

Leave a Reply

Your email address will not be published. Required fields are marked *