Sei sulla pagina 1di 16

1.

Short answer problems Total points: 70


(a) Consider the following 1-dimensional optimization problem problem
min
1
x
2
+ 54x
s.t. x > 0.
i. 10 What is the global minimum of this optimization problem? You must check all
conditions for the point to be a global minimum
ii. 5 What will the new optimum be if the constraints were changed to x 0.5?
Solution: The gradient f

(x) = 2x
3
+ 54. Thus, the x
0
=
1
3
is the lone
stationary point. It is easy to check that f

(x
0
) > 0. Therefore, x
0
=
1
3
is the
global minimum. We could have come to the same conclusion by recognizing that
the function
1
x
2
+ 54x is convex on {x : x > 0}.
Since f

(x) 0 for all x 0.5, it follows that the global minimum is, in fact,
x
0
= 0.5.
(b) Consider the 2-dimensional function
f(x) = x
2
1
+ x
3
2
+ 3x
1
x
2
i. 5 Show that x = [1, 1]

is not a stationary point for this problem.


ii. 5 Is d = [1, 2]

an ascent direction at x = [1, 1]

?
iii. 5 Categorize x = [9/4, 3/2] as a possibly local max/min, denitely local min/max,
saddle point, or none of these.
Solution: The gradient
f(x) =
_
2x
1
+ 3x
2
3x
2
2
+ 3x
1
_
.
The gradient at x = [1, 1]

is given by
f( x) =
_
5
6
_
= 0.
Therefore, (1, 1) is not a stationary point. Since
f( x)

d = 5 1 + 6 (2) = 7
d is not an ascent direction at x.
The gradient at x = [9/4, 3/2] is 0, and the Hessian is given by
H =
_
2 3
3 6 x
2
_
=
_
2 3
3 9
_
Since the determinants of all the principal minors are positive, it follows that
H 0. Thus, x = [9/4, 3/2]

is denitely a local min.


2
(c) 5 For what range of and is the following quadratic optimization a convex opti-
mization problem:
min x

_
2 0
0 1 + 2
_
x,
s.t.
_
2 3

x 0.
Solution: Since this is a minimization problem, we require that eigenvalues of the
Hessian matrix be positive semidenite. Since the matrix is diagonal, it follows
that the eigenvalues of the Hessian are
{2(2 ), 2(1 + 2)}
And, the eigenvalues are non-negative provided [1/2, 2].
The feasible set is a half space for all , and therefore, convex for all .
(d) 10 Show that (
3
2
, 1) is the global optimum solution for the constrained nonlinear pro-
gram
max ln x
1
+ ln x
2
s.t. 2x
1
+ 3x
2
6,
x 0.
Solution: Since ln x is a concave function, it follows that the above nonlinear
program is a convex program. Therefore, a KKT point will be the global maximum.
The KKT conditions imply that (
3
2
, 1) is a KKT point if there exists 0 such
that
f(x) =
_
1
x
1
1
x
2
_
=
_
2
3
_
=
1
3
satises both equations.
(e) Consider the following two variable QP
max 2x
1
+ 3x
2
x
2
1
2x
2
2
,
s.t. x
1
+ x
2
3,
21 x 21,
and consider the feasible point x
1
= x
2
=
3
2
.
i. 5 Construct the equality constrained QP that you would solve to compute the
step d at the point x
ii. 10 Solve for the step d and the next iterate x + d.
Solution: Only the constraint x
1
+ x
2
= 3 is tight at the point x = [1.5, 1.5]

.
The gradient
f( x) =
_
2 2x
1
3 4x
2
_
=
_
1 3

3
The equality constrained QP is given by
max
_
1
3
_

d d

_
1 0
0 2
_
d
s.t. d
1
+ d
2
= 0.
The KKT point for this QP is given by
_
2 0
0 4
_
d =
_
1
3
_
v
_
1
1
_
d =
_
v1
2
v3
4
_
Substituting back into the constraint d
1
+ d
2
= 0, we get
v 1
2
+
v 3
4
=
3
4
v
5
4
= 0 v =
5
3
.
Thus, d
1
= d
2
=
1
3
.
We have to possibly check four constraints. Since d
1
> 0, we only have to check
x
1
2 constraint. This gives
3
2
. Since d
2
< 0, we have to check the constraint
x
2
2. This constraint also gives
3
2
. Thus, the step length = 1, and the
new point
x + d =
1
6
_
11
7
_
(f) 10 Consider the following simple knapsack problem
max 4x
1
+ 6x
2
+ 7x
3
+ 2x
2
s.t 7x
1
+ 12x
2
+ 8x
3
+ 2x
4
14,
x
i
{0, 1}, i = 1, 2,
x
i
0, i = 3, 4.
We want to solve this integer linear program (ILP) using gurobi withing MATLAB
by giving the command
[sol, val] = gurobi_mex(f,objtype,A,b,contypes,lb,ub,vtypes);
Write the value for each of the inputs.
Solution:
f = [4;6;7;2];
objtype = -1;
A = sparse([7,12,8,2]);
b = [14];
4
contypes = [<];
lb = zeros(4,1);
ub = [1;1;inf;inf];
vtypes = [B,B,C,C];
2. Integer programming for orderly liquidation Total points: 20
Consider a network of N nancial institutions.
Let y
ij
denote the obligation of bank i to bank j, i.e. the amount owed by bank i
to bank j.
Let w
i
denote the total revenues of bank i. This is the amount of money that is
available to cover the obligations to other banks.
Let
i
denote the threshold below which the bank is declared to be insolvent. Let
{
ij
: i = j} denote that payments that bank i pays bank j. (Clearly we will always
have that
ij
y
ij
.) Then bank i is said to be insolvent if
1
_
w
i
+

j=i

ji

i=j

ij

i
_
The regulator has access to a bailout fund with a total amount B that can be split
across the banks. The regulator also has the right to mandate the payments {
ij
:
i = j}.
(a) 15 Formulate a linear integer program that determines the bailout payments b =
(b
1
, . . . , b
N
)

that minimizes the number of insolvencies. Assume that bank i can


use its wealth w
i
, the capital injection b
i
, and payments
ji
to fund its payments
ij
.
(b) 5 Banks 1 and 2 are very closely linked. If bank 2 is solvent, we require that bank 1
must also be solvent. Modify the ILP in the previous step to include this constraint.
Solution: Let x
i
= 1(bank i solvent) denote the indicator variable that bank i is
solvent. We must have that when the bank is solvent the wealth of the bank i, after it
pays all its obligations, should at least be
i
; however, if the bank is insolvent then we
do not care about its net position. Thus, the rst set of constraints is given by
w
i
+ b
i
+

j=i

ji

i=j

ij

i
x
i
+ M
i
(1 x
i
) (1)
where M
i
is a lower bound on the bank i wealth. Clearly, M
i
= (w
i

j=i
y
ij
) is a
lower bound.
5
The second set of constraints is given by the limited liability constraints, i.e.
0
ij
y
ij
, i = j. (2)
The third set of constraints come from the constraints on the bailout fund, i.e.
b 0,
N

i=1
b
i
B (3)
The ILP is given by
max

N
i=1
x
i
,
s.t. (1) (3)
We can impose the extra constraint by simply requiring that
x
1
x
2
.
3. 15 Tracking error problem Total points: 15
Suppose {r
k
R
d
: k = 1, . . . , N} denote N IID samples of the quarterly rate of
return each of the d assets, and let {b
k
: k = 1, . . . , N} denote IID samples of the
quarterly rate of return on a benchmark. The price per share of the assets is given by
p =
_
p
1
p
2
. . . p
n

> 0.
Suppose we are interested in choosing the minimum cost portfolio of the d assets such
that P(|r

b| > ) 1 , i.e. solve the optimization problem


min p

x
s.t. P(|r

b| > ) 1
x 0
where P denotes the empirical distribution induced by the data.
Suppose a feasible position x is known, i.e. P(|r

b| > ) 1 . Use this data to


formulate the tracking problem as an integer linear program. Clearly indicate how you
would set any constants that you have to introduce to model the problem.
Solution: Let z
k
1
_
|r

k
x b
k
| >
_
. Then the constraint
N

k=1
z
k
(1 )N
models the tracking error constraint. Now, we can impose the constraint z
k
1
_
|r

b| >
_
by the constraint
|r

k
x b
k
| + M
k
z
k
6
where M
k
is a suitably chosen upper bound. Since x is feasible, it follows that we can
restrict ourselves to the set {x : p

x p = p

x, x 0} without any loss of generality.


Thus, we can set
M
k
= max
x:x0,p

x p
|r

k
x b
k
| = max
_
p max
1in
r
ki
p
i
b
k
, b
k
min
1in
r
ki
p
i
, |b
k
|
_
.
Thus, the optimization problem is given by
min p

x,
s.t.

N
k=1
z
k
(1 )N,
r

k
x b
k
+ M
k
z
k
, k = 1, . . . , N,
b
k
r

k
x + M
k
z
k
, k = 1, . . . , N,
x 0, z {0, 1}
N
.
4. Minimum Norm problem Total points: 15
Let
V =
_
1 2 2 1 4

and consider the optimization problem


min

5
i=1
(V
i
w
i
)
2
,
s.t. 31 w 31.
(a) 5 What is the unconstrained optimal solution, i.e. the solution when there are no
constraints on w.
(b) 5 Using the unconstrained solution that you computed in part (a), guess the tight
constraints and solve an equality constrained QP.
(c) 5 Is the solution to the equality constrained QP optimal for the inequality constrained
QP. Why/why not ?
Solution: The unconstrained solution is, clearly, V.
Therefore, the only constraint that is likely to be tight is w
5
= 3. So, we guess that
this constraint is tight.
Let denote the dual variable corresponding to the constraint w
5
= 3. Then the KKT
conditions are given by
2(V
i
w
i
) = 0, i = 1, 2, 3, 4,
and
2(V
5
w
5
) = 0 = 2(4 3) = 2 < 0
Since the dual variable has the correct sign, it follows that the solution of the equality
constrained problem is, in fact, optimal for the inequality constrained problem.
7
5. Dynamic programming formulation for utility maximization Total points: 55
Consider the following nonlinear version of the knapsack problem in the text:
max 11 ln(x
1
+ 1) + 8 ln(x
2
+ 1) + 6 ln(x
3
+ 1) + 4 ln(x
4
+ 1),
s.t. 7x
1
+ 5x
2
+ 4x
3
+ 3x
4
14,
x
j
0 and an integer, j = 1, . . . , 4.
(4)
(a) 10 Reformulate the nonlinear integer program (4) as a integer linear program.
Solution: Let k
i
=
W
c
i
, where W = 14 and c = [7, 5, 4, 3]. For j = 0, . . . , k
i
,
i = 1, . . . , 4, dene
ij
= p
i
ln(1 + j), where p = [11, 8, 6, 4]. Then the above
nonlinear integer program is equivalent to the linear integer program
max

4
i=1

k
i
j=0

ij
y
ij
s.t.

4
i=1

k
i
j=0
jc
i
y
ij
W,

k
i
j=0
y
ij
= 1, i = 1, . . . , 4,
y
ij
{0, 1}, j = 0, . . . , k
i
, i = 1, . . . , 4.
(b) Let
V
j
(s) = maximum value achievable using objects i j and a total budget s
i. 10 Write a recursion that relates V
j+1
(s) to V
j
(s).
ii. 5 Show how you would start this recursion.
iii. 5 Compute V
2
(14) and the optimal choice for the rst two objects.
Solution: The DP recursion is given by
V
j+1
(s) = max
0ksc
j+1

{w
j+1
log(1 + k) + V
j
(s c
j+1
k)}
where w
j
denotes the weight of objects of type j in the objective, and c
j
denotes
the per-unit cost of object j.
We can start this result by computing V
1
(s) for all values of s {0, 1 . . . , 14}. It
is easy to check that
V
1
(s) =
_
_
_
0 0 s 6,
11 log(2) 7 s 13
11 log(3) s = 14
k

1
(s) =
_
_
_
0 0 s 6,
1 7 s 13
2 s = 14
Since 0 k
2
14/5 = 2, we only have to compare three possibilities
V
2
(14) = max{V
1
(14), 8 log(2) + V
1
(9), 8 log(3) + V
1
(4)}
= max{11 log(3), 8 log(2) + 11 log(2), 8 log(3)} = 19 log(2)
Therefore, k

2
= 1, from the result for V
1
it follows that k

1
= 1.
8
(c) 10 Consider the following convex relaxation for (4):
max

4
i=1
p
i
ln(x
i
+ 1),
s.t.

4
i=1
c
i
x
i
W,
x
j
0, j = 1, . . . , 4,
where p = [11, 8, 6, 4], c = [7, 5, 4, 3] and W = 14. Let denote the Lagrange
multiplier corresponding to the knapsack constraint

4
i=1
c
i
x
i
W. Show that the
optimal x

satises:
4

i=1
c
i
x

i
= W, x

i
= max
_
p
i
c
i

1, 0
_
Solution: Since the constraints are linear the KKT conditions characterize the
global optimum of this problem.
The KKT conditions are given by
p
i
1 + x
i
= c
i
+
i
where is the dual variable corresponding to the budget constraint, and
i
0
are the dual variables corresponding to the x
i
0 constraint.
Since the log function is monotonically increasing, it follows that the budget con-
straint must be tight at any optimal solution. So, we guess that > 0.
Next, we consider two cases:
(a) x
i
> 0: In this case,
i
= 0 and x
i
=
p
i
c
i
1
(b) x
i
= 0: In this case,
i
0 implies that p
i
c
i
, or equivalently,
p
i
c
i
1 0.
Thus, x
i
= max{
p
i
c
i
1, 0}. We also need that x be primal feasible, i.e.

4
i=1
x
i
=
W.
(d) 15 Describe an algorithm to solve for that satises these two equations. The algo-
rithm should be general and not specic to the numbers in this problem.
Solution: We have to search over . Assume without loss of generality that
p
i
c
i
are in increasing order.
set k = 1
solve

n
j=k
c
j
_
p
i
c
i
1
_
= W set =

n
j=k
p
j
W+

n
j=k
c
j
while
_
>
p
k
c
k
_
set k = k + 1
9
set =

n
j=k
p
j
W+

n
j=k
c
j
set x

j
= 0 for all j k 1 and x

j
=
p
j
c
j
1 for all j k.
6. Option pricing on Binomial Trees Total points: 25
Consider a binomial tree with S
0
= 1, u =
1
d
= 2, the interest rate r = 0, and the
risk-neutral probability =
1
2
.
Consider the option that gives the buyer the right to
Stop and receive S
t
at each of the three nodes time t = 2
Continue and receive
V =
_

_
1
2
2
1
4
_

_
in each of the 5 nodes at time t = 4. The payo V
i
denote the payo in the node
corresponding to the stock price S
t
= S
0
u
5i
d
i1
, i = 1, . . . , 5.
Note that the buyer has the option of stopping in only the three nodes at time t = 2
otherwise she continues up to time t = 4.
(a) 15 Write out the dynamic programming recursion for this problem. Clearly indicate
the dierence in the DP recursion for nodes at time t = 2 and the nodes at time
t = 2.
Solution: The recursion in all nodes at time t = 2 is given by
V
2
(s) = max
_
s, V
3
(us) + (1 ) V
3
(ds)
_
For all states at times t {0, 1, 3} the recursion is given by
V
t
(s) = V
t+1
(us) + (1 ) V
t+1
(ds)
And, at time 4
V
4
= V
(b) 10 Solve the DP recursion to compute the price of the option, and the optimal action
in each of the states.
Solution: Using the above recursion, we get
V
3
=
_

_
0.5(1 + 2)
0.5(2 + 2)
0.5(2 1)
0.5(1 + 4)
_

_
=
_

_
0.5
2
0.5
1.5
_

_
10
IEOR E4007
G. Iyengar
IEOR E4007 Final Exam Solutions
1. State price deators and asset pricing
(a) The market is incomplete. Payo S

= [1 0 0 0]

is not attainable.
Correction criteria: 3pts for correct statement, and 2pts for the example.
(b) Denote (
1
,
2
,
3
,
4
) as the state price deator. From the prices we have

1
+
2
= 1

3
= 1

4
= 2
where
1
,
2
> 0. Theres no unique state price deator for this market.
Correction criteria: 5pts for equations, 2pts for the sign constraints, 3 points for
unique/non-unique.
(c) Upper bound should be sup
m
{
1
+ 2
2
+
3
} = 3. Similarly, lower bound is
inf
m
{
1
+ 2
2
+
3
} = 2.
Correction criteria: 3pts for setting up the problem for the bounds, and 2pts for
the correct numerical solution
2. Portfolio selection with trading costs and value-at-risk constraints
(a) Dene x
i
= x
+
i
x

i
, where x
+
i
, x

i
0. Then we dene binary variables:
u
i
=
_
1 x
+
i
w
0 x
+
i
= 0
v
i
=
_
1 x

i
w
0 x

i
= 0
Since the upper bound for x

i
is , the upper bound for x
+
i
can be computed as
x
+
i
W
y

k=i
x
k
W
y
+ (n 1). So we have
u
i
+ v
i
1
wu
i
x
+
i
[W
y
+ (n 1)]u
i
wv
i
x

i
v
i
Then the constraint of holding at least 2 assets from assets {1, ..., 10} can be
written as
10

i=1
(u
i
+ v
i
) 2
Correction criteria: -2pts for incorrect bounds, incremental points taken o for
errors in setting up equations, severe penalties for non-linear constraints
1
(b) We introduce a new binary variable w
45
. w
45
= 1 if x
4
and x
5
are both included
in the portfolio, w
45
= 0 otherwise. This can be done by setting
w
45
(u
4
+ v
4
) + (u
5
+ v
5
) 1, w
45
(u
4
+ v
4
), w
45
(u
5
+ v
5
)
To force x
4
= x
5
, we add constraint (note that the upper bound for |x
4
x
5
| is
W
y
+ n)
x
4
x
5
(1 w
45
)(W
y
+ n)
x
5
x
4
(1 w
45
)(W
y
+ n)
To force x
4
= x
10
, we add constraint
x
4
+ x
10
2(1 w
45
)(W
y
+ (n 1))
x
4
+ x
10
2(1 w
45
)
Correction criteria: -2pts for incorrect bounds, incremental points taken o for
errors in setting up equations, severe penalties for non-linear constraints
(c) For piecewise linear function (u), we introduce the following new binary variables:
a = 1{0 < u 1}, b = 1{1 < u M}
where M is the upper bound for u. In other words, we have 3 cases:
u = 0 a = 0, b = 0
0 < u 1 a = 1, b = 0
1 < u M a = 0, b = 1
Dene m, n as portion of u from the two intervals:
0 < u 1 m = u, n = 0
1 < u M m = 0, n = u 1
Thus u = m + (b + n), with constraints
0 m a,
0 n (M 1)b,
a + b 1.
The function (u) can be written as
(u) = 0.1a + 0.8b + 0.7m + 0.35n
2
Dene z
i
= |x
i
y
i
|. Then we can write the constraint as
z
i
x
i
y
i
z
i
y
i
x
i
z
i
= m
i
+ b
i
+ n
i
0 m
i
a
i
0 n
i
(M
i
1)b
i
a
i
+ b
i
1
0.1a
i
+ 0.8b
i
+ 0.7m
i
+ 0.35n
i

i
a
i
, b
i
binary
where M
i
is the upper bound for z
i
= |x
i
y
i
|. This can be computed by
M
i
= max{|W
y
+ (n 1) y
i
|, | + y
i
|}
Correction criteria: -2pts for incorrect bounds, incremental points taken o for
errors in setting up equations, severe penalties for non-linear constraints
(d) Introduce binary variables w
k
= 1
_

n
i=1
x
i
L
(k)
i
>
_
, k = 1, ..., N. Then we can
write value-at-risk constraint as
1
N
N

k=1
w
k
1
n

i=1
x
i
L
(k)
i
+ U
k
w
k
where U
k
is the upper bound for loss in sample L
(k)
:
U
k
=
n

i=1
_
_
L
(k)
i
_
+
[W
y
+ (n 1)]
_
L
(k)
i
_

_
.
Another possible bound is
U
k
= max
i
_
_
L
(k)
i
_
+
_
[W
y
+ (n 1)] min
i
_
_
L
(k)
i
_

_
(n 1)
Correction criteria: -2pts for incorrect bounds, incremental points taken o for
errors in setting up equations, severe penalties for non-linear constraints
(e) When theres no error above, put everything together. All the constraints other
than the budget constraint 1

x 1

n
i=1

_
|x
i
y
i
|
_
is explicitly modeled
above. This constraint can be modeled as by substituting the upper bounds
i
for (|x i y
i
|), i.e. the constraint to impose is
1

x 1

y
n

i=1

i
.
3
The constraint: either |x|
i
w or x
i
= 0 can be modeled by the following set of
constraints: for all i {1, . . . , n}
u
i
, v
i
, {0, 1}
u
i
+ v
i
1
wu
i
x
+
i
[W
y
+ (n 1)]u
i
wv
i
x

i
v
i
Correction criteria: -1pt for not formulating the budget constraint and -1pt for
not setting up the either-or constraint above.
3. Minimum variance hedging
(a) The equality QP is
minimize
5

i=1

i
(V
i
w
i
)
2
subject to
5

i=1

i
w
i
= B
This program is indeed
minimize (w
1
+ 1)
2
+ (w
2
2)
2
+ (w
3
2)
2
+ 2(w
4
+ 1)
2
+ (w
5
4)
2
subject to w
1
+ w
2
+ w
3
+ 2w
4
+ w
5
= 2
The gradient condition for this optimization problem gives

i
_
5

i=1

i
(V
i
w
i
)
2
_
=
i
2
i
(V
i
w
i
) =
i
w
i
=

2
+ V
i
.
Substituting this back in the constraint, we get
1
2
(1

) +

V = B =
2(B

V)
1
T

.
Note that for all B

V, the multiplier 0.
Substituting the value for B, we get
=
2(2 5)
6
= 1,
and
w =
_
1.5 1.5 1.5 1.5 3.5

Correction Criteria: incorrect Lagrangian formulation -5 pts, incorrectly computed


-5 pts, incorrectly computed w -5 pts.
4
(b) Since there is only one constraint in this problem and this constraint is tight at
the solution in part (a), all we have to do is to check if the dual multiplier 0.
We have shown above that 0 whenever the budget B

V = 5. Thus, the
solution obtained in part (a) is optimal.
Correction Criteria: not using the dual variable to prove it -5 pts, not explaining
why a negative dual variable implies the solution is optimal -3 pts.
(c) See CVX code below:
V = [-1 2 2 -1 4];
pi = [1 1 1 2 1];
B = 2;
cvx_begin
variables w(5)
minimize ((V-w)
*
diag(pi)
*
(V-w))
subject to
pi
*
w <= B
cvx_end
Correction Criteria: wrong cost function -5 to -8 pts depending on the error.
(d) In part (a), we established that the equality constrained problem gives the optimal
solution to the inequality constrained problem for all B 5, and moreover, the
optimal solution is
w
i
=

2
+ V
i
.
Therefore, the optimal value of the optimization problem is
f(B) = (1

)

2
4
=
(B

V)
2
1

=
1
6
(B

V)
2
.
Correction Criteria: not explaining that the curve is quadratic -6 pts, not adding
any information to the plot -8 pts.
4. Knapsack problem
(a) See MATLAB/Gurobi code below:
c = [11; 8; 6; 4];
objtype = -1;
A = sparse([7 5 4 3]);
b = W;
lb = 0;
ub = 3;
contypes = <;
5
vtypes = IIII;
[x,val,exitflag,output] =
gurobi_mex(c,objtype,A,b,contypes,lb,ub,vtypes,opts);
Correction criteria: -1pt for each incorrect MATLAB statement.
(b) First we compute upper bounds for x
1
, ..., x
4
. It is clear that x
1
{0, 1, 2},
x
2
{0, 1, 2}, x
3
{0, 1, 2, 3}, x
4
{0, 1, 2, 3, 4}. Then we introduce binary
variables x
ij
and set x
ij
= 1 if and only if x
i
= j. So we have
x
1
= x
11
+ 2x
12
x
2
= x
21
+ 2x
22
x
3
= x
31
+ 2x
32
+ 3x
33
x
4
= x
41
+ 2x
42
+ 3x
43
+ 4x
44
Since for variable x
i
, at most one binary variables x
ij
can be 1, so we have
x
11
+ x
12
1
x
21
+ x
22
1
x
31
+ x
32
+ x
33
1
x
41
+ x
42
+ x
43
+ x
44
1
Now we are able to replace the quadratic terms in objective function with linear
combinations of these binary variables:
x
2
1
= x
11
+ 4x
12
x
2
2
= x
21
+ 4x
22
x
2
3
= x
31
+ 4x
32
+ 9x
33
x
2
4
= x
41
+ 4x
42
+ 9x
43
+ 16x
44
Thus, the optimization problem is now given by
max 11(x
11
+ 2x
12
) 3.075(x
11
+ 4x
12
)
+ 8(x
21
+ 2x
22
) 1.65(x
21
+ 4x
22
) + 6(x
31
+ 2x
32
+ 3x
33
)
0.95(x
31
+ 4x
32
+ 9x
33
)
+ 4(x
41
+ 2x
42
+ 3x
43
+ 4x
44
) 0.45(x
41
+ 4x
42
+ 9x
43
+ 16x
44
),
s.t. 7(x
11
+ 2x
12
) + 5(x
21
+ 2x
22
)
+ 4(x
31
+ 2x
32
+ 3x
33
) + 3(x
41
+ 2x
42
+ 3x
43
+ 4x
44
) 14,
x
11
+ x
12
1,
x
21
+ x
22
1,
x
31
+ x
32
+ x
33
1,
x
41
+ x
42
+ x
43
+ x
44
1,
x
ij
{0, 1}.
6
Correction criteria: 5pt for correct objective function, 3pts for correct main con-
straint, 2pts for the bounds on the auxiliary variables.
5. Portfolio selection with multiple risk measures
See CVX code below:
for k = [1:m]
Q{k} = chol(Q{k}); % or Q{k} = sqrtm(Q{k});
end
cvx_begin
variables x(n) maxstd
maximize (mu
*
x - lambda
*
maxstd)
subject to
sum(x) == 1
Ax <= b
for k = [1:m]
maxstd >= norm(Q{k}
*
x, 2)
end
cvx_end
Correction Criteria: using min instead of max -5 pts., errors in the functions used -1
to -4 depending on the error, using a quadratic formulation -4 to -8 pts depending on
the error, not using multiple scenarios -5 to -8 pts.
6. Dark pool execution problem
(a) The revenue we get from selling the k-th block is v
j
(k), and the probability that
we can indeed sell this block is given by P(

S
j
k) = p
j
(k). So the total expected
revenue we get from selling k blocks in venue j is
f
j
(k) =
k

i=1
p
j
(i)v
j
(i)
Correction criteria: 15pts for the expression. Incremental points deducted for the
extent of errors.
(b) V
k+1
(s) is the maximum revenue we get from selling s blocks through venue k +
1, ..., 5. Now suppose we want to include venue k in our consideration. To decide
how many blocks in s should we allocate to venue k, we use the following recursion:
V
k
(s) = max
0js
{f
k
(s) + V
k+1
(s j)}
To initialize the recursion, we rst compute V
5
(s), 0 s B. This is straightfor-
ward since our only choice is to put everything into venue 5, we have V
5
(s) = f
5
(s).
7

Potrebbero piacerti anche