Portfolio Optimization

Unit 7: Discrete-Time Mean-Variance Portfolio Selection

Duan LI & Xiangyu Cui
India Institute of Technology Kharagpur
May 26 - 30, 2014

2
Literature Review and Research Need
The mean-variance formulation by Markowitz provides a fundamental
basis for portfolio selection in single period.
The literature in multi-period portfolio selection has been dominated
by the results of maximizing expected utility functions of the terminal
wealth until 2000.
Difficulty in elicitating utility function for investors.
Trade-off information between the risk and the expected return is lost
in the utility approach.
The concept of the Markowitzs mean-variance formulation needs to
be extended to the multi-period portfolio selection.
Reference: Li, D. and W. L. Ng (2000): Optimal dynamic portfolio
selection: Multiperiod mean variance formulation. Mathematical
Finance, 10, 387-406.

3
Mean-Variance Formulation for Multi-Period Portfolio Selection

We consider a capital market with (n + 1) risky securities, with random rates of returns.
An investor joins the market at time 0 with an initial wealth x0 .
The investor can allocate his wealth among the (n + 1) assets. The
wealth can be reallocated among the (n + 1) assets at the beginning
of each of the following (T 1) consecutive time periods.
The vector
return of the risky securities at time period t:
 0 1of random


et = et , et , ..., ent ,
where eit is the random return for security i at time period t.
et , t = 0, 1, . . ., T - 1: statistically independent.






0
1
n
Expected return vector: E (et ) = E et , E et , ..., E (et )

4
Covariance matrix of the random return:

t,00

..
Cov (et ) = ...
.
t,0n

t,0n
.. .
.
. . . t,nn

xt : the wealth of the investor at the beginning of the t-th period.

uit , i = 1, 2, . . ., n: the amount invested in the ith risky asset at the
beginning of the t-th time period.
The amount invested in the 0th P
risky asset at the beginning of the
t-th time period is equal to xt ni=1 uit .

5
Mean-variance formulation for multi-period portfolio selection with security
0 being taken as a reference:
(P 1()) :

max E(xT )
s.t. V ar(xT )
n
X
xt+1 =
eit uit +
=

i=1
e0t xt

xt

n
X

uit

i=1

e0t

+ Pt ut , t = 0, 1, 2, ..., T 1

and
(P 2()) :

min V ar(xT )
s.t. E(xT )
xt+1 = e0t xt + Pt ut , t = 0, 1, 2, ..., T 1

where
Pt =


1
2
n
Pt , Pt , ..., Pt

(e1t

e0t ), (e2t

e0t ), ..., (ent


0
et )

6
Notice that E(et (et ) ) = Cov(et ) + E(et )E(et ).
Assume

E(et (et ) )

E((e0t )2 ) E(e0t e1t ) . . . E(e0t ent )

E(e1t e0t ) E((e1t )2 ) . . . E(e1t ent )

...................................
E(ent e0t ) E(ent e1t ) . . . E((ent )2 )
E(Pt Pt ) > 0

> 0, t = 0, 1, . . . , T 1

t = 0, 1, . . . , T 1

E((e0t )2 ) E(e0t Pt )E 1 (Pt Pt )E(e0t Pt ) > 0, t = 0, 1, . . . , T 1

7
Enable an investor to specify a risk level which he can afford when he
is seeking to maximize his expected terminal wealth or to specify an
expected terminal wealth he would like to achieve when he is seeking
to minimize the corresponding risk.
It is easier and more direct for investors to provide this kind of subjective information than for them to construct a utility function in term
of the terminal wealth.
Multi-period portfolio policy:

= {0 , 1 , 2 , ..., T 1 }
1 1
1
0
1
T 1

20 21
2T 1

.. , .. , ..., ..

. .
.

n0
n1
nT 1

8
: feedback policy.

u1t
u2t

1t
2t

(xt )
(xt )
..
.

.. =
.
unt
nt (xt )

A multi-period portfolio policy, , is efficient if there exists no other

multi-period portfolio policy, , such that E (xT ) | E (xT ) | and
V ar (xT ) | V ar (xT ) | with at least one equality strictly.

9
An equivalent formulation to either (P 1()) or (P 2()) in generating
efficient multi-period portfolio policies:
(E (w)) :

max E (xT ) wV ar (xT )

s.t. xt+1 = e0t xt + Pt ut , t = 0, 1, ..., T 1

where w [0, ) is the trade-off between maximizing the expected

final wealth and minimizing the risk.
If solves (E(w)), then solves (P 1()) with = V ar (xT ) |
and solves (P 2()) with = E (xT ) | .
w=

E(xT )
V ar(xT )

at the optimal solution of (E(w)).

Problem formulation (E(w)) is preferable to be adopted where an

investor is able to specify his desirable trade-off between the expected
terminal wealth and the associated risk.

10
Analytical Solution to the Multi-period MV Formulation
Let
Bt

A1t

A2t

Bt1

Bt2

E Pt E 1 (Pt Pt ) E (Pt )


0
E et

E (Pt ) E 1 (Pt Pt ) E e0t Pt

0 2
E (et )

1


0

0
E et Pt E (Pt Pt ) E et Pt
QT 1
1
k=t+1 Ak
Bt QT 1
2 k=t+1 A2k
!2
QT 1
1
k=t+1 Ak
Bt
QT 1
2 k=t+1 A2k
TY
1
t=0

A1t

11

T
1
X
t=0

TY
1

TY
1

A1k

k=t+1

Bt1

A2t

t=0

a =
b =
c =

2
2

a
2 ab2

The optimal multi-period portfolio policy for problem (E(w)):

1
0
ut = E
Pt Pt E et Pt xt
!
TY
1


1

A1k

1
+
(bx0 +
)
E
Pt Pt E (Pt )
2
2
2wa
Ak
k=t+1

uT 1

E 1

t = 0, 1, . . . , T 2

0
PT 1 PT 1 E eT 1 PT 1 xT 1

12
+

1
1

(bx0 +
)E
PT 1 PT 1 E (PT 1 )
2
2wa

The optimal multi-period portfolio policy for problems (P 1()) and

(P 2()):

1
0
ut = E
Pt Pt E et Pt xt
!
TY
1


1

A1k

1
(bx0 +
)
E
Pt Pt E (Pt )
+
2

2
2w a
Ak
k=t+1

uT 1

=
+

where

w =

t = 0, 1, . . . , T 2




1

0
E
PT 1 PT 1 E eT 1 PT 1 xT 1


1

(bx0 +
)E
PT 1 PT 1 E (PT 1 )

2
2w a
(

2 a(cx20 )
2
2a[(+b)x0 ]

when (P 1()) is solved

when (P 2()) is solved

13
The mean-variance efficient frontier for problems (P 1()), (P 2())
and (E(w)):
a
2
2
[E
(x
)

(
+
b)x
]
+
cx
V ar (xT ) =
T
0
0
2
for E (xT ) ( + b)x0
With the analytical solution, the implementation of optimal multiperiod portfolio policy for problem (P 1()), (P 2()), or (E(w)) is
straightforward.
The optimal multi-period portfolio policy consists of two terms and
exhibits a separation property between the investors risk attitude and
his current wealth.
The second term in ut is dependent on the investors risk attitude
and is independent of his current wealth. It can be calculated
off-line before the real investment process starts.
The first term in ut is dependent on the current wealth and is
independent of the investors risk attitude. It is calculated on-line
at every time period when the current wealth is observed.

14
Derivation of the Analytical Solution

All three problems (P 1()), (P 2()) and (E (w)) are difficult to be

solved directly due to their nonseparability in the sense of dynamic
programming.
Variance minimization has been a notorious problem in stochastic control.
Let I t be an information set available at time t and I t1 I t , t.
A key observation is that while the expectation operator satisfies the
smoothing property:
E[E( | I j ) | I k ] = E( | I k ), j > k,
the variance operation does not:
V ar[V ar( | I j ) | I k ] 6= V ar( | I k ), j > k.

15
A solution scheme adopted is to embed problem (E(w)) into a tractable
auxiliary problem that is separable, investigate the relationship between the
solution sets of problem (E (w)) and the auxiliary problem, and search for
the solution of the auxiliary problem that attains the optimum point of problem (E (w)).
Q

Define E (w) to be the set of optimal solutions of problem (E (w))

with given w, i.e.
E (w) = {| is a maximizer of (E (w))}

(1)

Define

=
=

2
e
U E xT , E (xT )
E (xT ) wV ar (xT )



2
2
wE xT + wE (xT ) + E (xT )

2
e
U is a convex function of E xT and E (xT ).

(2)

16
Auxiliary problem:
(A (, w))

max E wx2T + xT
s.t. xt+1 = e0t xt + Pt ut ,

t = 0, 1, 2, ..., T 1

Prominent features of problem (A (, w)):

Problems (A (, w)) and (E (w)) have the same sets of admissible
controls (The feasibility sets are the same).
(A (, w)) is of a separable structure in the sense of dynamic programming.
The objective function of (A (, w)) is of a quadratic form while the
system dynamic is of a linear form.

17
Define A (, w) to be the set of the optimal solutions of problem
(A (, w)) for given and w, i.e.,
A (, w) = {| is a maximizer of (A (, w))}

(3)

Denote
d(, w)

=
=

2
e
U E xT , E (xT )
|
E (xT )
1 + 2wE (xT ) |

(4)

Theorem 1 For any E (w), A (d ( , w) , w).

The implication of Theorem 1 is that the solution set for problem (E (w))
is a subset of the solution set for problem (A (, w)). We can embed
the nontractable primal problem (E (w)) into a tractable auxiliary problem (A (, w)) with a quadratic utility function.

18
Proof of Theorem 1
By contradiction, assume that
/ A (d ( , w) , w). Then there exists a such
that








2
2


E
x
E
x
T
T
 > [w, d ( , w)]

[w, d ( , w)]


E (xT ) 
E (xT ) 

(5)

Notice (4) and

e E x2 , E (xT )
U
T


= w
(6)
E x2T


e is a convex function of E x2 and E (xT ), the following property is satisfied,
Since U
T








e E x2 , E (xT ) | U
e E x2 , E (xT ) |
U

T
T








E x2
2


E
x
T
T

+ [w, d ( , w)]

(7)



E (xT ) 
E (xT ) 

Combining (5) and (7) yields

2
2
e
e
U E xT , E (xT ) | > U E xT , E (xT ) |
that contradicts the assumption of E (w).

(8)

19
Optimal solution of the auxiliary problem (A (, w)):
ut (xt ; ) = Kt xt + vt ()

t = 0, 1, . . . , T 1

(9)

where

Kt

vt () =

(10)
w

1
0
E
Pt Pt E et Pt
(11)
!
TY
1
1

Ak
1

E
(P
P
t
t ) E (Pt ) , t = 0, 1, .., T 2(12)
2
2
Ak
k=t+1

with the following boundary condition,

1

PT 1 PT 1 E (PT 1 )
vT 1 () = E
2

(13)

Substituting (9) into the equation of wealth dynamics yields the dynamics of the wealth under portfolio policy ut (xt ; ),

0
xt+1 () = et Pt Kt xt () + Pt vt ()
(14)

20
Taking expectation on both sides of (14) and noticing the statistical
independence between (e0t , Pt ) and xt :
!
TY
1
A1k

1
Bt
(15)
E (xt+1 ()) = At E (xt ()) +
2
2
Ak
k=t+1

Taking square on both sides of (14)

h
2
i

x2t+1 () =
e0t 2e0t Pt Kt + Kt Pt Pt Kt x2t ()

0
+2 et Pt Kt xt () Pt vt ()

+vt () Pt Pt vt () , t = 0, 1, . . . , T 1

(16)

Taking expectation on both sides of the above equation and simplifying

the resulted expression lead to the following:
E

x2t+1

() =

A2t E

x2t

() +
4

TY
1

A1k
A2k
k=t+1

!2

Bt

(17)

21

Solving two recursive equations (15) and (17) yields explicit expressions
for the expected values of the terminal wealth and the square of the terminal
wealth under portfolio policy ut (xt ; ),
E (xT ()) = x0 +

(18)

2
(19)
E
() =
+
2
The variance of the terminal wealth under portfolio policy ut (xt ; ):


2
V ar (xT ()) = E xT () E 2 (xT ())
x2T

x20

a( bx0 )2 + cx20

(20)

The expected terminal wealth E (xT ()) is an increasing linear function of .

The variance of the terminal wealth, V ar (xT ()), is a quadratic
function of .

22




2
e
From (18) and (20), we can express U E xT , E (xT ) as a function
of ,




2
e
U E xT , E (xT )
=

x0 + w[a( bx0 )2 + cx20 ]

(21)

e is a concave function of . Differentiating (21) with respect to

U
yields,
e
dU
= 2wa( bx0 )
(22)
d
Optimal satisfies

= bx0 +
(23)
2wa
Substituting the optimal in (23) into (9) yields the optimal multi-period
portfolio policy for (E(w)) specified in (1) and (1).

23
Substituting (23) into (18) and (20) yields the expression for the expected value and the variance of the terminal wealth on the efficient frontier
in terms of w,
2
(24)
E (xT (w)) = ( + b)x0 +
2wa
2
2
+
cx
(25)
V ar (xT (w)) =
0
4aw2
Given a problem (P 1()) or (P 2()), we can first calculate the associated
w in terms of or using (24) or (25) and then compute the corresponding optimal using (23). Substituting the optimal into (9) yields the
optimal multi-period portfolio policy for (P 1()) or (P 2()).
The mean-variance efficient frontier given in (1) can be obtained by
eliminating the parameter w in (24) and (25).

24
Investment with One Riskless Asset
Let the 0th security be riskless. We consider now a capital market
with n risky assets and a riskless asset.


0
0 i
In this case et equals to a constant st and cov et , et = 0.
The parameters now take the following forms,
Bt

A2t

Bt2

E Pt E 1 Pt Pt E (Pt ) , A1t = st (1 Bt )
Bt
s2t (1 Bt ) , Bt1 = QT 1
2 k=t+1 sk
Bt

4(

QT 1

2
k=t+1 sk )

TY
1
1
[1
(1 Bt )],
2
t=0

TY
1

st (1 Bt ) ,

TY
1

t=0

s2t (1 Bt ) ,

t=0

T 1
TY
1
TY
1
1 Y
(1 Bt ) [1
(1 Bt )], b = 2
st , c = 0
4 t=0
t=0
t=0

25

The optimal parameter for problem (E(w)):

=2

TY
1

st x0 +
w

t=0

Q

T 1
t=0

(1 Bt )

The optimal portfolio policy for problem (E(w)) in the investment

situations with a riskless asset:

1
ut = st E
Pt Pt E (Pt ) xt
!
TY
1
TY
1
1
1
Q
]
E 1 (Pt Pt ) E (Pt )
+ [
sk x0 +
T 1
sk
2w
(1 Bk )
k=0

k=0

t = 0, 1, . . . , T 2

k=t+1

26

uT 1

=
+

sT 1 E
[

TY
1
k=0

sk x0 +

PT 1 PT 1
2w

Q

PT 1 PT 1

1
T 1
k=0

E (PT 1 ) xT 1

(1 Bk )

E (PT 1 )

]

ut , t = 1, 2, . . . , T 1, is proportional to E 1 (Pt Pt )E(Pt ).

Each investor will spread his wealth among risky securities in the
same relative proportions. On the other side, the ratio of the investment in risky assets to the investment in the riskless asset is determined at each time period by observing the realized value of his wealth
and based on the investors attitude towards risk. An extension of
the well known separation theorem in single-period portfolio selection.

27
The expected terminal wealth and the variance of the terminal wealth
under the optimal portfolio policy ut in the investment situations with
a riskless asset:
E (xT )
=

QT 1

1 t=0 (1 Bt )
Q

st x0 +
T 1
2w
t=0
t=0 (1 Bt )


QT 1
1 t=0 (1 Bt )
V ar (xT ) =
QT 1
2
4w
t=0 (1 Bt )
TY
1

The optimal portfolio policy for (P 1()) and (P 2()) in the investment
situations with a riskless asset:

1
ut = st E
Pt Pt E (Pt ) xt
+

TY
1
k=0

sk x0 +

2w

Q

T 1
k=0

(1 Bk )

]

28
TY
1

uT 1

=
+

k=t+1

1
sk

sT 1 E 1
[

TY
1

kt x0 +

k=0

E 1 (Pt Pt ) E (Pt )
t = 0, 1, . . . , T 2

PT 1 PT 1 E (PT 1 ) xT 1

2w

PT 1 PT 1

where

w =

Q

QT 1

T 1
k=0

(1 Bk )

E (PT 1 )

t=0 (1Bt ))
QT 1
t=0 (1Bt )
QT 1
1
(
t=0 (1Bt ))
QT 1
Q 1
2( t=0 st x0 )( T
t=0 (1Bt ))

1
2

(1

]

for (P 1())
for (P 2())

29
The analytical expression of the mean-variance efficient frontier for
situations with a riskless asset:
QT 1
t=0 (1 Bt )
V ar (xT ) =
QT 1
1 t=0 (1 Bt )
!2
TY
1

E (xT ) x0
st
t=0

for E (xT )

TY
1

st x0

t=0

When setting T = 1 in our formulation, problems (P 1()) and (P 2())

are reduced to the single-period mean-variance formulation and the
efficient frontier are exactly the same as given in this paper and in Eq.
(35) of Merton (1962).

30
Illustrative Cases

Example 1 An investor has one unit wealth in the very beginning of the
planning horizon of T = 4. The investor is trying to find the best allocation
of his wealth among three risky securities, A, B, and C in order to maximize
E(x4 ) while keeping his risk not exceeding 2, i.e. = 2.
The expected returns for risky securities, A, B, and C are E(eA
t ) = 1.162,
C
E(eB
t ) = 1.246, and E(et ) = 1.228, t = 0, 1, 2, 3.

0.0146 0.0187

The covariance of et = [eA , eB , eC ] is Cov(et ) = 0.0187 0.0854

0.0145 0.0104
t = 0, 1, 2, 3.

0.0145

0.0104 ,

0.0289

31
Take security A as the reference asset.
E(Pt ) = [0.084, 0.066] , t = 0, 1, 2, 3.
E(Pt Pt ) =

0.0697 0.0027
0.0027 0.0189

, t = 0, 1, 2, 3.

E(eA Pt ) = [0.1017, 0.0766], t = 0, 1, 2, 3.

Bt = 0.3566, A1t = 0.7424, A2t = 0.8711, t = 0, 1, 2, 3, = 0.3038,
= 0.4077, a = 0.0376, b = 3.2933, and c = 0.0754.
The mean-variance efficient frontier:
V ar(x4 ) = 0.2262[E(x4 ) 1.6465]2 + 0.0754
where E(x4 ) 1.6465.

32
Optimal w = 0.75773.
The associated optimal portfolio policy:

where Kt

5.1094
14.0004
in the first

ut = Kt xt + vt

1.6238
4.3548
=
, t = 0, 1, 2, 3, v0 =
, v1 =
4.2907
11.9327





5.9948
7.0335
, v2 =
, and v3 =
. The investment
16.4263
19.2726
P i

security, asset A, at period t is equal to xt ut .

The corresponding expected terminal wealth and the risk level are given
by E(x4 ) = 4.5632 and V ar(x4 ) = 2, respectively.

33
Example 2 Consider now a modified version of Example 1. In addition
to the three risky assets, A, B, and C, there exists a riskless asset with a
sure return rate of 1.04. Suppose this time that the investor seeks an efficient portfolio policy with a desired trade-off between the expected terminal
wealth and risk,

E(x4 )
V ar(x4 )

= 2. More directly, the investor would like to

maximize E(x4 ) - 2V ar(x4 ).

E(Pt ) = [0.122, 0.206, 0.188] , t = 0, 1, 2, 3.

0.0295 0.0438
E(Pt Pt ) = 0.0438 0.1278
0.0374 0.0491

0.0374
0.0491 , t = 0, 1, 2, 3.
0.0642

Bt = E(Pt )E 1 (Pt Pt )E(Pt ) = 0.593817, t = 0, 1, 2, 3.

34
Mean-variance efficient frontier:
V ar(x4 ) = 0.02798[E(x4 ) 1.1699]2
where E(x4 ) 1.1699.
The associated optimal portfolio policy:

ut = Kt xt + vt

3.5440
0.4004
where Kt = 0.6496 , t = 0, 1, 2, 3, v0 = 5.7494 , v1 =
2.3133
20.4751

3.9865
3.8332
3.6858
5.9794 , v2 = 6.2185 , and v3 = 6.4673 . The invest23.0317

22.1459
21.2941
P i
ment in the riskless asset at period t is equal to xt ut .

The corresponding expected terminal wealth and the risk level are E(x4 )
= 10.1043 and V ar(x4 ) = 2.2336, respectively.