Financial Engineering and Risk Management

Mean Variance Optimization

Martin Haugh Garud Iyengar
Columbia University
Industrial Engineering and Operations Research
Overview
Assets and portfolios
Quantifying random asset and portfolio returns: mean and variance
Mean-variance optimal portfolios
Ecient frontier
Sharpe ratio and Sharpe optimal portfolios
Market portfolio
Capital Asset Pricing Model
2
Assets and portfolios
Asset anything we can purchase
Random price P(t)
Random gross return R(t) =
P(t+1)
P(t)
Random net return: r(t) = R(t) 1 =
P(t+1)P(t)
P(t)
Total wealth W > 0 distributed over d assets
w
i
= dollar amount in asset i: w
i
> 0 long, w
i
< 0 short
Net return on a position w
r
w
(t) =

d
i=1
R
i
(t)w
i

n
i=1
w
i
W
=

d
i=1
r
i
(t)w
i

d
i=1
w
i
=
d

i=1
r
i
(t)
w
i
W
..
x
i
portfolio vector x = (x
1
, . . . , x
d
): each component can be +ve/-ve
x
i
= fraction invested in asset i

d
i=1
x
i
= 1
3
How does one deal with randomness?
Random net return on the portfolio r
x
=

d
i=1
r
i
x
i
How does one quantify random returns ?
Maximize expected return E[r
x
]?
Should one worry about spread around the mean?
How does one quantify the spread?
4
Random returns on assets and portfolios
Parameters dening asset returns
Mean of asset returns:
i
= E[r
i
(t)]
Variance of asset returns:
2
i
= var(r
i
(t))
Covariance of asset returns:
ij
= cov
_
r
i
(t), r
j
(t)
_
=
ij

j
Correlation of asset returns
ij
= cor
_
r
i
(t), r
j
(t)
_
All parameters assumed to be constant over time.
Parameters dening portfolio returns
Expected return on a portfolio x = (x
1
, . . . , x
d
)

x
= E[r
x
(t)] =
d

i=1
E[r
i
(t)]x
i
=
d

i=1

i
x
i
Variance of the return on portfolio x:

2
x
= var(r
x
(t)) = var
_
d

i=1
r
i
x
i
_
=
d

i=1
d

j=1
cov
_
r
i
(t), r
j
(t)
_
x
i
x
j
5
Example
d = 2 assets with Normally distributed returns N(,
2
)
r
1
N(1, 0.1) r
2
N(2, 0.5) cor(r
1
, r
2
) = 0.25
Parameters

1
= 1
2
= 2

2
1
= var(r
1
) = 0.1
2
2
= var(r
2
) = 0.5

12
= cov(r
1
, r
2
) = cor(r
1
, r
2
)
1

2
= 0.25

0.05 = 0.0559
Portfolio: (x, 1 x)

x
=
d

i=1

i
x
i
= x + 2(1 x)

2
x
=
d

i,j=1

ij
x
i
x
j
=
d

i=1

2
i
x
2
i
+ 2

j>i

ij
x
i
x
j
= 0.1x
2
+ 0.5(1 x)
2
+ 2(0.0559)x(1 x)
6
Diversication reduces uncertainty
d assets each with
i
,
i
,
ij
= 0 for all i = j
Two dierent portfolios
x = (1, 0, . . . , 0)

: everything invested in asset 1

y =
1
d
(1, 1, . . . , 1)

: equal investment in all assets.

Expected returns of the two portfolios

x
= E[

d
i=1

i
x
i
] =
1
=

y
= E[

d
i=1

i
y
i
] =
1
d

d
i=1

i
=
Both have the same expected return!
Variance of returns of the two portfolios

2
x
= var(

d
i=1
r
i
x
i
) =
2

2
y
= var(

d
i=1
r
i
y
i
) =

d
i=1

2
(
1
d
)
2
=

2
d
Diversied portfolio has a much lower variance!
7
Markowitz mean-variance portfolio selection
Markowitz (1954) proposed that
Return of a portfolio Expected return
x
Risk of a portfolio volatility
x
Ecient frontier
0 20 40 60 80 100 120 140 160 180
60
40
20
0
20
40
60
80
100
120
volatility (%)
r
e
t
u
r
n

(
%
)
Ecient frontier max return
for a given risk
How does one characterize the
ecient frontier?
How does one compute e-
cient/optimal portfolios?
8
Mean variance formulations
Minimize risk ensuring return target return
min
x

2
x
s.t.
x
r
min
x

d
i=1

d
j=1

ij
x
i
x
j
s.t.

d
i=1

i
x
i
r

d
i=1
x
i
= 1.
Maximize return ensuring risk risk budget
max
x

x
s.t.
2
x

2
max
x

d
i=1

i
x
i
s.t.

d
i=1

d
j=1

ij
x
i
x
j

2
,

d
i=1
x
i
= 1.
Maximize a risk-adjusted return
max
x

x

2
x
max
x

d
i=1

i
x
i

d
i=1

d
j=1

ij
x
i
x
j
_
s.t.

d
i=1
x
i
= 1.
risk-aversion parameter
9
Financial Engineering and Risk Management
Ecient frontier
Martin Haugh Garud Iyengar
Columbia University
Industrial Engineering and Operations Research
Mean-variance for 2-asset market
d = 2 assets
Asset 1: mean return
1
and variance
2
1
Asset 2: mean return
2
and variance
2
2
Correlation between asset returns
Portfolio: (x, 1 x)

x
=
d

i=1

i
x
i
=
1
x +
2
(1 x)

2
x
=
d

i,j=1

ij
x
i
x
j
=
d

i=1

2
i
x
2
i
+ 2

j>i

ij
x
i
x
j
=
2
1
x
2
+
2
2
(1 x)
2
+ 2
1

2
x(1 x)
2
Mean-variance for 2-asset market
Minimize risk formulation for the mean-variance portfolio selection problem
min
x

2
1
x
2
+
2
2
(1 x)
2
+ 2
1

2
x(1 x)
s.t.
1
x +
2
(1 x) = r
Expected return constraint: x =
r
2

2
Variance:

2
r
=
2
1
_
r
2

2
_
2
+
2
2
_

1
r

2
_
2
+ 2
1

2
_
r
2

2
__

1
r

2
_
= ar
2
+ br + c
Explicit expression for the variance as a function of target return r.
3
Ecient frontier
0 2 4 6 8 10 12 14
6
4
2
0
2
4
6
8
10
12
volatility (%)
r
e
t
u
r
n

(
%
)

min
r
min
Efficient
Inefficient
Only the top half is ecient! why did we get the bottom?
How does one solve the d asset problem?
4
Computing the optimal portfolio
Mean-variance portfolio selection problem

2
(r) = min
x

d
i=1

d
j=1

ij
x
i
x
j
s.t.

d
i=1

i
x
i
= r

d
i=1
x
i
= 1.
Form the Lagrangian with Lagrange multipliers u and v
L =
d

i=1
d

j=1

ij
x
i
x
j
v
_
d

i=1

i
x
i
r
_
u
_
d

i=1
x
i
1
_
Setting
L
x
i
= 0 for i = 1, . . . , d gives d equations
2
d

j=1

ij
x
j
v
i
u = 0, for all i = 1, . . . d ()
Can solve the d + 2 equations in d + 2 variables: x
1
, . . . , x
d
, u and v.
Theorem. A portfolio x is mean-variance optimal if, and only if, it is feasible and
there exists u and v satisfying ().
5
Computing the optimal portfolio
Matrix formulation
_

_
2
11
2
12
. . . 2
1d

1
1
2
21
2
22
. . . 2
2d

2
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
d1
2
d2
. . . 2
dd

d
1

1

2
. . .
d
0 0
1 1 . . . 1 0 0
_

_
. .
A
_

_
x
1
x
2
.
.
.
x
d
v
u
_

_
=
_

_
0
0
.
.
.
0
r
1
_

_
..
b
Therefore
_

_
x
1
x
2
.
.
.
x
d
v
u
_

_
= A
1
b
6
Two fund theorem
Fix two dierent target returns: r
1
= r
2
Suppose
x
(1)
= (x
(1)
1
, . . . , x
(1)
d
)

optimal for r
1
: Lagrange multipliers (v
1
, u
1
)
x
(2)
= (x
(2)
1
, . . . , x
(2)
d
)

optimal for r
2
: Lagrange multipliers (v
2
, u
2
)
Consider any other return r
Choose =
rr
1
r
2
r
1
Consider the position: y = (1 )x
(1)
+ x
(2)
y is a portfolio
d

i=1
y
i
= (1 )
d

i=1
x
(1)
i
+
d

i=1
x
(2)
i
= (1 ) + = 1
y is feasible for target return r
d

i=1

i
y
i
= (1 )
d

i=1

i
x
(1)
i
+
d

i=1

i
x
(2)
i
= (1 )r
1
+ r
2
= r
7
Two fund theorem (contd)
Set v = (1 )v
1
+ v
2
and u = (1 )u
1
+ u
2
.
2
d

j=1

ij
y
j
v
i
u =
d

j=1
2
ij
((1 )x
(1)
j
+ x
(2)
j
)

i
((1 )v
1
+ v
2
) ((1 )u
1
+ u
2
)
= (1 )
_
2
d

j=1

ij
x
(1)
j
v
1

i
u
1
_
+
_
2
d

j=1

ij
x
(2)
j
v
2

i
u
2
_
= 0
y is optimal for target return r!
Theorem All ecient portfolios can be constructed by diversifying between any
two ecient portfolios with dierent expected returns.
Why are there so many funds in the market?
8
Ecient frontier
The optimal portfolio for target return r
y

=
_
r
2
r
r
2
r
1
_
x
(1)
+
_
r r
1
r
2
r
1
_
x
(2)
= r
_
x
(2)
x
(1)
r
2
r
1
_
. .
g
+
_
r
2
x
(1)
r
1
x
(2)
r
2
r
1
_
. .
h
y

i
= rg
i
+ h
i
, i = 1, . . . , d.
Therefore

2
(r) =
d

i=1
d

j=1

ij
(rg
i
+ h
i
)(rg
j
+ h
j
)
= r
2
_
d

i=1
d

j=1

ij
g
i
g
j
_
+ 2r
_
d

i=1
d

j=1

ij
g
i
h
j
_
+
_
d

i=1
d

j=1

ij
h
i
h
j
_
The d-asset frontier has the same structure as the 2-asset frontier.
9
Ecient frontier
0 2 4 6 8 10 12 14
6
4
2
0
2
4
6
8
10
12
volatility (%)
r
e
t
u
r
n

(
%
)

min
r
min
Efficient
Inefficient
10
Financial Engineering and Risk Management
Mean-variance with a risk-free asset
Martin Haugh Garud Iyengar
Columbia University
Industrial Engineering and Operations Research
Mean Variance with a risk-free asset
New asset: pays net return r
f
with no risk (deterministic return)
x
0
= fraction invested in the risk-free asset
Mean-variance problem: x
0
does not contribute to risk.
max
_
r
f
x
0
+

d
i=1

i
x
i
_

d
i=1

d
j=1

ij
x
i
x
j
_
s. t. x
0
+

d
i=1
x
i
= 1.
Only meaningful for r r
f
Substituting x
0
= 1

d
i=1
x
i
we get
max r
f
+

d
i=1
(
i
r
f
)x
i

d
i=1

d
j=1

ij
x
i
x
j
_

i
=
i
r
f
= excess return of asset i
2
Mean-variance optimal portfolio
Taking derivatives we get

i
2
d

j=1

ij
x
j
= 0, i = 1, . . . , d.
Matrix formulation
2
_

11

12
. . .
1d

21

22
. . .
2d
.
.
.
.
.
.
.
.
.
.
.
.

d1

d2
. . .
dd
_

_
. .
V
_

_
x
1
x
2
.
.
.
x
d
_

_
=
_

_

1
.
.
.

d
_

_
. .

x() =
1
2
V
1

The family of frontier portfolios as a function of :
_
_
1
d

i=1
x
i
(), x()
_
: 0
_
3
One-fund theorem
The positions in the risky assets in the frontier portfolio
x =
1
2
V
1

do not add up to 1.
Dene a portfolio of risky assets by dividing x by the sum of its components.
s

=
_
1

d
i=1
x
i
_
x =
_
1
1
2

d
i=1
(V
1
)
i
_
_
1
2
V
1

_
The portfolio s

is independent of ! Since

d
i=1
x
i
= 1 x
0
, x = (1 x
0
)s

.
Family of frontier portfolios = {(x
0
, (1 x
0
)s

) : x
0
R}
Theorem All ecient portfolios in a market with a risk-free asset can be
constructed by diversifying between the risk-less asset and the single portfolio s

.
4
Ecient frontier with risk-free asset
Return and risk of portfolio s

s
=

d
i=1

i
s

i
,

s
=
_

d
i=1

j=1

2
ij
s

i
s

j
Return on a generic frontier portfolio: x
0
in risk-free and (1 x
0
) in s

x
= x
0
r
f
+ (1 x
0
)

s

x
= (1 x
0
)

s
Ecient Frontier
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
(0, r
f
) risk-free asset
volatility (%)
r
e
t
u
r
n

(
%
)
(

s
,

s
) s

Straight line with an intercept r

f
at = 0 and slope
m =

s
r
f

s
How does this relate to the fron-
tier with only risky assets?
Does the portfolio s

have an
economic interpretation?
5
Ecient frontier with risk-free asset
0 1 2 3 4 5 6 7 8 9
1
2
3
4
5
6
7
8
9
volatility (%)
r
e
t
u
r
n

(
%
)

s
,

s
) s

r
f
s

must be an ecient risky portfolio

The ecient frontier with a risk-free
asset must be tangent to the ecient
frontier with only risky assets.
The portfolio s

maximizes the angle or equivalently

tan() =

d
i=1

i
x
i
r
f
_

d
i=1

d
j=1

ij
x
i
x
j
=
expected excess return
volatility
6
Sharpe Ratio
Denition. The Sharpe ratio of a portfolio or an asset is the ratio of the
expected excess return to the volatility. The Sharpe optimal portfolio is a
portfolio that maximizes the Sharpe ratio.
The portfolio s

is a Sharpe optimal portfolio

s

= argmax
_
x:

d
i=1
x
i
=1
_
_

x
r
f

x
_
Investors diversify between the risk-free asset and the Sharpe optimal portfolio.
The investment in the various risky assets are in xed proportions ...
prices/returns should be correlated! This insight leads to the Capital Asset
Pricing Model.
7
Financial Engineering and Risk Management
Capital Asset Pricing Model
Martin Haugh Garud Iyengar
Columbia University
Industrial Engineering and Operations Research
Market Portfolio
Denition. Let C
i
, i = 1, . . . , d, denote the market capitalization of the d
assets. Then the market portfolio x
(m)
is dened as follows.
x
(m)
i
=
C
i

d
j=1
C
j
, i = 1, . . . , d.
Let
m
denote the expected net rate of return on the market portfolio, and let

m
denote the volatility of the market portfolio.
Suppose all investors in the market are mean-variance optimizers. Then all of
them invest in the Sharpe optimal portfolio s

. Let
w
(k)
= wealth of the k-th investor
x
(k)
0
= fraction of wealth of the k-investor in the risk-free asset
Then
C
i
=

k
w
(k)
(1 x
(k)
0
)s

i
The market portfolio x
(m)
= Sharpe optimal portfolio s

!
2
Capital Market Line
Capital market line is another name for the ecient frontier with risk-free asset
Recall: Ecient frontier = line though the points (0, r
f
) and (
m
,
m
)
Slope of the capital market line
m
CML
=

m
r
f

m
= maximum achievable Sharpe ratio
m
CML
is frequently called the price of risk. It is used to compare projects.
Example. Suppose the price of a share of an oil pipeline venture is $875. It is
expected to yield $1000 in one year, but the volatility = 40%. The current
interest rate r
f
= 5%, the expected rate of return on the market portfolio

m
= 17% and the volatility of the market
m
= 12%. Is the oil pipeline worth
considering?
r
oil
=
1000
875
1 = 14% r = r
f
+
_

m
r
f

m
_
= 45%
Not worth considering!
3
Inferring asset returns from market returns
An asset is a portfolio: asset j x
(j)
= (0, . . . , 1, . . . , 0)

, 1 in the j-th position.

Diversify between x
(j)
and market portfolio x
(m)
: x
(j)
+ (1 )x
(m)
return

=
j
+ (1 )
m
volatility

=
_

2
j
+ (1 )
2

2
m
+ 2
jm
(1 )
0 1 2 3 4 5 6 7 8 9
1
2
3
4
5
6
7
8
9
volatility (%)
r
e
t
u
r
n

(
%
)


(
m
,
m
) x
(m)
risk-free asset (0, r
f
)
Efficient frontier with riskfree
Efficient frontier w/o riskfree
Efficient frontier of market + asset 2
4
All three curves are tangent at (
m
, r
m
)
Slope of the capital market line
m
CML
=

m
r
f

m
Slope of the frontier generated by asset j and market portfolio x
(m)
d

=
d

d
d

d
=

j

2
j
(1)
2
m
+(1)
jm

jm

2
j
+(1)
2

2
m
+2
jm
(1)
d

=0
=

j

jm

2
m

m
Equating slopes at = 0 we get the following result:

j
r
f
=
_

jm

2
m
_
. .
beta of asset j
(
m
r
f
)
This pricing formula is called the Capital Asset Pricing Model (CAPM).
5
Connecting CAPM to regression
Regress the excess return r
j
r
f
of asset j on the excess market return r
m
r
f
(r
j
r
f
) = + (r
m
r
f
) +
j
Parameter estimates
coecient
j
=
cov(r
j
r
f
,r
m
r
f
)
var(r
m
r
f
)
=

jm

2
m
intercept
j
= (E[r
j
] r
f
) (E[r
m
] r
f
) = (
j
r
f
) (
m
r
f
).
residuals
j
and (r
m
r
f
) are uncorrelated, i.e. cor(
j
, r
m
r
f
) = 0.
CAPM implies that
j
= 0 for all assets.
Eective relation: r
j
r
f
=
j
(r
m
r
f
) +
j
Decomposition of risk
var(r
j
r
f
) =
2
j
var(r
m
r
f
) +var()

2
j
=
2
j

2
m
. .
market risk
+ var()
. .
residual risk
Only compensated for taking on market risk and not residual risk
6
Security Market Line
Plot of the historical returns on an asset vs r
f
+ (
m
r
f
)
2 1.5 1 0.5 0 0.5
8
6
4
2
0
2
4
1
2
3
4
5
6
7
8
Security market line
beta
r
e
t
u
r
n

(
%
)
The assets are labeled in the order they appears in the spreadsheet.
All assets should lie on the security line if CAPM holds. So why the discrepancy?
7
Assumptions underlying CAPM
All investors have identical information about the uncertain returns.
All investors are mean-variance optimizers (or the returns are Normal)
The markets are in equilibrium.
Leveraging deviations from the security market line
Jensens index or alpha
= (
j
r
f
)
j
(
m
r
f
)
hold long if positive, short otherwise
Sharpe ratio of a stock
s
j
=

j
r
f

j
hold long if > m
MCM
, short otherwise.
8
CAPM as a pricing formula
Suppose the payo from an investment in 1yr is X. What is the fair price for this
investment.
Let r
X
=
X
P
1 denote the net rate of return on X. The beta of X is given by

X
=
cov(r
X
, r
m
)

2
m
=
1
P
cov(X, r
m
)

2
m
Suppose CAPM holds. Then
X
= E[r
X
] must lie on the security market line, i.e.

X
= r
f
+
X
(r
m
r
f
)
E[X]
P
1 = r
f
+
1
P
cov(X, r
m
)
var(r
m
)
(
m
r
f
)
Rearranging terms:
P =
E[X]
1 + r
f

cov(X, r
m
)
(1 + r
f
)var(r
m
)
(
m
r
f
)
9