6-ConceptsProbability Page 198 Wednesday, February 4, 2004 3:00 PM
198 The Mathematics of Financial Modeling and Investment Management
demonstrated that joint normal distributions produce a linear regression
function. Consider the joint normal distribution
-^1 ---
1
f()v = [( 2 π)
n
ΣΣΣΣ]
2
exp – ---(v – μμμμ)
T
ΣΣΣΣ
- 1
(v – μμμμ)
2
where parameters are those defined in an earlier section in this chapter.
Let’s partition the parameters as follows:
x μx σxx, σz,x
v= , μ= , Σ=
z μz σx,z Σz
where μx, μz are respectively a scalar and a p-vector of expected values,
σx,x, σx,z, σz,x, and Σz are respectively a scalar, p-vectors and a p×p
2
matrix of variances and covariances and σxx= σ
2
, x , σzi,zi = σzi. It can
be demonstrated that the variable (X|Z = z) is normally distributed with
the following parameters:
(XZ = z) ∼N[μ –^1
x – (Σ
– (^1) σ
z z,x)' (μz – z), σxx, – σx,zΣz σz,x +]
From the above expression we can conclude that the conditional
expectation is linear in the conditioning variables. Let’s call
α= μx – (Σ– z^1 σz,x)' μz and β = Σ–z^1 σz,x
We can therefore write
g()z = EX[ Z = z]= α β′+ z
If the matrix Σis diagonal, the random variables (X,Z 1 ,...,Zp) are
independent, such that σz,x= 0 and β= Σz
- 1
σz,x = 0 and therefore the
regression function is a constant that does not depend on the condition-
ing variables. If the matrix Σz is diagonal but σx,z, σz,xdo not vanish,
then the linear regression takes the following form
p σ p σ
xz, i xz, i
g()z = EX[ Z = z]= μx – ∑-----------μ + ∑-----------zi
i= 1 σ^2
zi
zi i=^1 σ^2 zi