Markowitz Mean-Variance Portfolio Theory

Markowitz Mean-Variance Portfolio Theory

1. Portfolio Return Rates

An investment instrument that can be bought and sold is often called an asset. Suppose we purchase an asset for x0 dollars on one date and then later sell it for x1 dollars. We call the ratio

R = x1 x0

the return on the asset. The rate of return on the asset is given by

r = x1 - x0 = R - 1. x0

Therefore, x1 = Rx0 and x1 = (1 + r)x0 .

Sometimes it is possible to sell and asset that we do not own. This is called short selling. It works somewhat as follows. Suppose you wish to short (or short sell) a particular stock XXX. You begin by asking your stock broker if their firm is holding XXX in the total pool of stocks owned by all of their customers. If the brokerage does hold (or manage) some of stock XXX, you can ask them sell any number of stock XXX up to the number that they hold. This sale is credited against your account as a debt equal to the number of stock XXX they sell on your behalf. That is, your debt is not denominated in dollars, but rather in the number of stock XXX that you are shorting (i.e. your account is short by the given number of stock XXX). On your account asset sheet, this short sale appears as a negative number associated with the shorted asset. Remember, this negative number is not denominated in dollars, but rather in the number of stocks , or assets, shorted. Due to the sale of stock XXX you have received x0 dollars. Eventually, you must ask the brokerage to buy the same number of stock XXX back as you originally asked them to sell and return this stock to the pool of assets that they are holding for their customers. On the date at which you return stock XXX you ask your broker to re-purchase it at its current going value of x1 dollars and return it to the brokerage's asset pool. If x1 < x0, then you have made a profit on this transaction; otherwise, you have a loss. The return and rate of return on this transaction are given by

R = -x1 = x1 and r = (-x1) - (-x0) = x1 - x0 ,

-x0 x0

-x0

x0

respectively. Short selling can be very risky, and many brokerage firms

do not allow it. Nonetheless, it can be profitable.

1

2

Let us now consider constructing a portfolio consisting of n assets.

We have an initial budget of x0 dollars that we wish to assign to these assets. The amount that we assign to asset i is x0i = wix0 for i = 1, 2, . . . , n, where wi is a weighting factor for asset i. We allow the weights to take negative values, and when negative it means

that the asset is being shorted in our portfolio. To preserve the budget

constraint we require that the weights sum to 1:

n i=1

wi

=

1.

That

is,

n

n

the sum of the investments = wix0 = x0 wi = x0 .

i=1

i=1

Notice that by shorting some stocks we open up more funds for the purchase of other stocks, because when we short a stock we receive the dollar value of that stock today and we can turn around and re-invest those dollars elsewhere with the purchase of other assets.

If Ri denotes the return on asset i, then the total receipts from our portfolio is

n

n

x1 = Riwix0 = x0 Riwi,

i=1

i=1

and so the total return from the portfolio is

n

R = Riwi.

i=1

In addition, we have that the rate of return from asset i is ri = Ri - 1, i = 1, 2, . . . , n. Hence the rate of return on the portfolio is

n

n

n

n

r = R - 1 = ( Riwi) - ( wi) = (Ri - 1)wi = riwi .

i=1

i=1

i=1

i=1

2. The Basics of Markowitz Mean-Variance Portfolio Theory

In the Markowitz mean-variance portfolio theory, one models the rate of returns on assets as random variables. The goal is then to choose the portfolio weighting factors optimally. In the context of the Markowitz theory an optimal set of weights is one in which the portfolio achieves an acceptable baseline expected rate of return with minimal volatility. Here the variance of the rate of return of an instrument is taken as a surrogate for its volatility.

3

Let ri be the random variable associated with the rate of return for asset i, for i = 1, 2, . . . , n, and define the random vector

r1

z

=

r2 ...

.

rn

Set ?i = E(ri), m = (?1, ?2, . . . , ?n)T , and cov(z) = . If w =

(w1, w2, . . . , wn)T is a set of weights associated with a portfolio, then

the rate of return of this portfolio r =

n i=1

riwi

is

also

a

random

variable with mean mT w and variance wT w. If ?b is the acceptable

baseline expected rate of return, then in the Markowitz theory an opti-

mal portfolio is any portfolio solving the following quadratic program:

M minimize

1 2

wT

w

subject to mT w ?b, and eT w = 1 ,

where e always denotes the vector of ones, i.e., each of the components of e is the number 1. The KKT conditions for this quadratic program are

(1)

0 = w - m - e

(2)

?b mT w, eT w = 1, 0

(3)

T (mT w - ?b) = 0

for some , IR. Since the covariance matrix is symmetric and positive definite, we know that if (w, , ) is any triple satisfying the KKT conditions then w is necessarily a solution to M. Indeed, it is easily shown that if M is feasible, then a solution to M must always exist and so a KKT triple can always be found for M.

Proposition 2.1. Let A IRm?n, B IRm?t, E IRs?n, F IRs?t, M IRt?n, Q IRn?n and H IRt?t with Q and H symmetric, and let r IRm, and h IRs. Further assume that the symmetric

matrix

Q^ =

Q MT MH

is positive semi-definite. If the following quadratic program is feasible,

then it has finite optimal value and a solution attaining this optimal

value exists:

minimize

1 2

uT Qu + 2vT M u + vT Hv

subject to Au + Bv r

Eu + Fv = h

0u .

4

Proof. Let S be any square root of the matrix Q^, and set x = (uT , vT )T . Set = IRm+ ? IRs ? IRn+. Then the constraint region for the QP can be written as F = {x IRn ? IRt : T x }, where

A B T = E F .

I0

By assumption F = . Making the change of variable y = Sx in the QP yields the QP

minimize

1 2

y

2

subject to y SF = {Sx : x F} .

Since the set F is a closed nonempty polyhedral convex set, so is the set SF. Hence the solution y? to this QP is given by the point closest to the origin in the closed set SF which must exist since this set is closed and nonempty. Therefore, the solution set to the original QP is nonempty and is given by {x : y? = Sx, x F}.

Assume that is nonsingular and that w? is a solution to M (w?

exists by Proposition 2.1). We consider two cases.

? ?b < mT w?: In this case, the complementarity condition (3) implies = 0. Hence the KKT conditions reduce to the two equations 0 = w? - e and eT w? = 1. Multiplying the first through by -1 yields w? = -1e. Multiplying this equation through by e and using the fact that eT w? = 1 gives = (eT -1e)-1.

Therefore,

w? = (eT -1e)-1-1e.

It is important to note that this value of w gives the smallest possible variance over all portfolios since it solves the problem

Mmin-var : minimize

1 2

wT

w

subject to eT w = 1 .

Consequently, the return associated with the least variance so-

lution is

mT -1e

?min-var =

. eT -1e

We denote the set of weights associated with the minimum vari-

ance solution w? by wmin-var as well.

Finally observe that is the minimum variance weights wmin-var are feasible for M, that is, if mT wmin-var ?b, then wmin-var

must be the solution to M since it the solution to the prob-

lem Mmin-var. Therefore, when solving M one first computes wmin-var and checks to see if the inequality mT wmin-var ?b

5

holds. If it does hold, then wmin-var solves M and no further work is required. If it does not hold then you know that the constraint mT w = ?b at the solution to M.

? ?b = mT w?: Multiplying (1) through by -1 gives

(4)

w? = -1m + -1e .

Using this formula for w? and (2), we get the two equations

?b = mT -1m + mT -1e 1 = mT -1e + eT -1e ,

or equivalently, the 2 ? 2 matrix equation

(5)

mT -1m mT -1e mT -1e eT -1e

=

?b 1

.

Properties of positive definite matrices can be used to show that

the matrix

T=

mT -1m mT -1e mT -1e eT -1e

= [m e]T -1[m e]

is always positive semi-definite There are a number of ways to see this. The simplest is to exploit the factored form T = [m e]T -1[m e]. But a simple test is to check that

0 < = (mT -1m)(eT -1e) - (mT -1e)2.

This will always be the case whenever the vectors m and e are linearly independent.

If = 0, then it must be the case that m = e for some IR. In this case, if ?b/ = 1, then the problem M is necessarily infeasible. If ?b/ = 1, then wmin-var solves M which would have been detected by first computing wmin-var and then checking its feasibility for M.

If > 0, the system (5) can be solved to give

= eT v and = -mT v,

where

v = -1-1(?be - m) .

Plugging these values into (4) gives the optimal solution

-1e

-1m

-1e

w=

+

-

eT -1e

eT -1m eT -1e

-1e

-1m

= (1 - )

+

eT -1e eT -1m

= (1 - )wmin-var + wmk ,

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download