Page 1
Calibration and portfolio optimization issuesin asset price modelling driven by Levy processes
Jos e E. Figueroa-L opez
University of California, Santa Barabara
Department of Statistics and Applied Probability
ORIE Colloquium . Cornell University
January 23rd, 2007
Page 2
Outline
I. Asset price modelling driven by Levy processes
II. Calibration issues
III. Portfolio optimization problems
1
Page 3
Asset price modelling driven by L evy processes
What is a L evy process?
A random quantity Xt evolving in time in such a way that
• The increment, Xt+∆t −Xt, during a time period [t, t + ∆t] is both
(1) independent of the “past” Xs : s ≤ t and
(2) with distribution law depending only on the time span ∆t
• The process can exhibit sudden changes in magnitude (jumps), but these
occur at unpredictable times (no fixed jump times).
Conclusion: Levy process is the most natural generalization of Brownian motion,
where only continuity of paths is relaxed, while preserving the statistical qual-
ities of the increments.
2
Page 4
A natural extension of the Black-Scholes model:
Geometric Levy models for an asset price process:
dSt = St−(bdt + d Zt︸︷︷︸Levy process
) ⇐⇒ St = S0 exp( Xt︸︷︷︸Levy process
),
Why a L evy Market?
• More flexible modelling of return distribution; e.g. heavy-tails, asymmetry,
and high kurtosis
• More consistency with the actual asset price evolution, which are made up
of discrete trades, and are exposed to sudden changes
3
Page 5
Challenges and drawbacks of L evy-based modelling
Statistical issues
• Intractable or not explicit marginal probability density functions
• Too many parametric models around!! Normal, stable, hyperbolic distrib-
utions, tempered or truncated stable, etc. Which is better?
• Hard to guarantee convergence to the maximum likelihood estimators.
Conclusion: Computational demanding and numerically instable estimation
by traditional statistical methods.
Option pricing, hedging, and portfolio optimization:
(1) Incompleteness
(2) Infinitely many arbitrage free option prices.
(3) Lack of numerical schemes to find replication strategies and optimal (utility-
based) strategies.
4
Page 6
Statistical properties of L evy processes
Characterization and parameters
• The law of the process is determined by the distribution of X1
• Three parameters, two reals σ2, µ, and a measure ν(dx), such that
(1) The counting process that counts the number of jumps with size falling
on an interval [a, b] is a Poisson process with intensity ν([a, b]).
(2) Xt = µt+σWt +limε↓0 Xεt − tmε, where Xε is a compound
Poisson process with mean EXεt = tmε.
(3) The jump evolution of the process is independent from the Brownian part.
A typical assumption: The Levy measure ν is determined by a “density func-
tion” s in the sense that
ν([a, b]) =∫ b
a
s(x) dx.
5
Page 7
Remarks about L evy density functions:
Necessary and sufficient conditions to be a L evy density:
∫ 1
−1
s(x)x2dx < ∞ and∫
|x|>1
s(x)dx < ∞.
Consequences:
• A Levy density is not a probability density function:
∫ ∞
−∞s(x)dx = ∞⇒ s(x)
|x|→0−→ ∞⇐⇒ inf. many jumps of
arbitrarily small size
• It is easier to specify a Levy process via the Levy density of ν than via
the probability density of X1.
6
Page 8
Nonparametric estimation of the L evy density
Problem: Estimate directly the Levy density s assuming only qualitative informa-
tion on s.
Objectives
• Avoid model biases: “Let the data speak”.
• Computational efficiency, and suitability to deal with high-frequency data.
7
Page 9
Basic ideas of the model selection paradigm
• Approximation of the (infinite-dimensional) non-parametric model by finite-
dimensional linear models:
s(x) ≈ β1ϕ1(x) + · · ·+ βnϕn,
where the ϕ’s are known functions. The space
S := β1ϕ1(x) + · · ·+ βnϕn : β1, . . . , βn reals
is called an (approximating) linear model.
• Typical examples: Histograms, splines, trigonometric polynomials, wavelets.
• Two problems to solve:
Projection estimation: Estimating the element ofS that is “closest” to s.
Model selection: Determining a good approximating model from a col-
lection of linear models.
8
Page 10
Implementation of the model selection approach
• Assumptions:
– Suppose that there exists a reference measure η such that, on a given es-
timation window D ⊂ R\0, s is bounded and∫
Ds2(x)η(dx) < ∞.
– Suppose one has a good estimator β(ϕ) of the inner product
β(ϕ) :=∫
D
ϕ(x)s(x)η(dx),
for any ϕ s.t.∫
Dϕ2(x)η(dx) < ∞.
• Implication: η introduce a natural distance between functions; namely,
‖s− p‖2η :=∫
D
(s(x)− p(x))2η(dx).
9
Page 11
• Implementation:
– Take approximating linear models of the form
S := β1ϕ1(x) + · · ·+ βnϕn : β1, . . . , βn reals
with functions ϕ′is such that∫
Dϕ2
i (x) η(dx) < ∞. Without loose of
generality, one can always take ϕ1, . . . , ϕn orthonormal on D with re-
spect to η.
– The “closest” member of S to the function s is when
βi = β(ϕi) =∫
D
ϕi(x)s(x)η(dx).
– Natural projection estimator of s onS , with respect to the distance induce
by η, is
s(x) := β(ϕ1)ϕ1(x) + · · ·+ β(ϕn)ϕn(x).
10
Page 12
• Data-driven Model Selection:
– Problem: Choose between two or more finite-dimensional linear models
– Goal: Accomplish a good trade off between the bias of the approximat-
ing model and the standard error of the estimation inside the model.
– A sensible solution: Choose the model that minimize an unbiased estima-
tor of the risk:
E[‖s− s‖2η
]= ‖s‖2η + E
[−‖s‖2η]+ β(ϕ2
S ).
where
ϕ2S (x) :=
∑
i
ϕ2i (x).
Choose the model that minimizes the observable statistic
−∑
i
β2(ϕi) + β(ϕ2S ).
11
Page 13
A Numerical Illustration
Gamma Levy process with parameters α and β
• X(t) D∼ Gamma(αt, β), β - scale parameter and α - shape parameter
• No Brownian part and Levy density s(x) = αx e−x/β , x > 0:
Pure-jump increasing process with infinite jump-activity.
• Jump-behavioral interpretation:
α - overall jump activity and β - frequency of big-jumps
12
Page 14
Model selection methods for the Gamma L evy process
I. Estimation outside the origin:
• On a window of estimation D = [a, b] away from the origin, s(x) =αx e−x/β is bonded and square integrable with respect to η(dx) = dx.
• Recall that
E∑
u≤T
χ(c,d)
(∆Xu) = T
∫χ(c,d)(x)s(x)dx.
Then,
E∑
u≤T
ϕ(∆Xu) = T
∫
D
ϕ(x)s(x)dx.
A natural estimator for β(ϕ) :=∫
Dϕ(x)s(x)dx is
β(ϕ) :=1T
∑
u≤T
ϕ(∆Xu).
13
Page 15
• Example: Take histogram estimators on regular partitions.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
5
10
15
20
25
30
35
Penalized Estimation Projection
x
s(x
)
Sample Path Information: Gamma Process with α = 1 and β = 1(2000 jumps on [ 0 , 365 ])
Method of Estimation: Regular Histograms
c=2 Estimation window = [0.02, 1.0]
Best partition = 51 intervals
• How good is the estimator? Fit the model αx e−x/β (using least-squares)
to the histogram estimator: αMS = 0.93 and βMS = 1.055(vs. αMLE = 1.01 and βMLE = 0.94)
14
Page 16
• Estimation based on discrete sampling:
– Problem: Estimate
β (ϕ) :=1T
∑
t≤T
ϕ(∆Xt),
based on n equally spaced observations on [0, T ]: XT/n
, . . . , XT .
– A natural solution: Approximate by
βn (ϕ) :=1T
n∑
k=1
ϕ(Xtk
−Xtk−1
),
where tk = T kn , for k = 1, . . . , n.
15
Page 17
– Example: Gamma Levy process with α = 1 and β = 1:
Time span MS-LSF MLE
1 1.01 1.46 .997 .995
.5 1.03 1.09 .972 .978
.1 .944 .995 1.179 .837
.01 .969 .924 1.01 .98
(Based on 36500 jumps on [0, 365]≈ 100 jumps/ unit time)
16
Page 18
– Sampling distribution
Means and standard errors of αMS and βMS based on 1000 repeti-
tions
∆t MS-LSF MLE
.1 0.81 (0.06) 1.40 (0.50) 1.001 (0.01) 0.99 (0.05)
.01 0.92 (0.08) 1.12 (0.31) 1.007 (0.07) 0.99 (0.08)
.001 0.93 (0.08) 1.13 (0.34) 1.007 (0.07) 0.99 (0.08)
(simulations based on equally-spaced observations, with time span ∆t,
during the time interval [0, 365])
17
Page 19
Model selection methods for Tempered Stable Process
Model: [Rosinski 2005] No Brownian part with Levy density
s(x) =1
|x|α+1q(|x|), 0 < α < 2,
where q(x) and q(−x) (x > 0) are bounded “completely monotone func-
tions” (decreasing, convex; e.g. exp(−x)) vanishing for large x.
Appealing Property:
• h−1/αXhtD≈ Stable distribution with parameter α, for h small.
Stable-like high-frequency returns
• h−1/2XhtD≈ Normal distribution with parameter α, for h large.
Normal-like low-frequency returns
Goal: Estimate q.
18
Page 20
Projection estimator:
• Approximating linear model for q:
S := β1ϕ1(x) + · · ·+ βdϕd(x),
where the ϕ’s are orthonormal functions defined on D = [0, h].
• Estimator of the projection of q on S :
q(x) = β(ϕ1)ϕ1(x) + · · ·+ β(ϕd)ϕd(x),
where
β (ϕ) :=1T
∑
t≤T
ϕ(∆Xt) with ϕ(x) := |x|α+1ϕ(x)
and α being a “good” estimate of α; e.g. using methods for stable processes.
19
Page 21
• Intuition: If ϕ(x) ≈ xα+1ϕ(x) then
E [∑
t≤T
ϕ(∆Xt)] ≈ T
∫
D
ϕ(x)q(x)dx.
• Numerical illustration: One-sided TS, s(x) = axα+1 e−x/b, x > 0
– Histogram approximations of q(x) = ae−x/b on D = [0, 3]– Paths generated from 36500 jumps on [0, 365] with a = b = 1 and
α = .1.
– Jumps approximated by equally-spaced increments (time span ∆t)
– α estimated by Zolotarev method for stable distributions (valid for h ≈0) based on the moments of log |Xt|.
∆t Penalized Projection - Least-Squares Fit Misspecified Gamma MLE
.01 1.03 (0.15) 0.97 (0.14) 0.09 (0.0002) 1.2 (0.08) 0.89 (0.079)
20
Page 22
Summary
We developed estimation and model selection schemes for the Levy density of a
Levy process:
• Flexible: it can be used histograms, splines, wavelets, etc.
• Model free
• Easily implementable
• Reliable and robust: Oracle inequality and adaptivity; i.e. asymptotically com-
parable to minimax estimators on classes of smooth Levy densities
21
Page 23
The problem of portfolio optimization
• Formulation of the problem
• Solution using convex duality: advantages and drawbacks
• Improvements in jump-diffusion models driven by Levy processes
• Conclusions
22
Page 24
Formulation of the problem
Set-up: A frictionless market consisting of a risky asset with price St and a risk-
free bond with value Bt defined on a probability space (Ω,F ,P).
Goal: Dynamically allocate an initial endowment w so that to maximize the agent’s
expected final utility during a fixed time horizon [0, T ].
State-dependent utility U(w,ω) : R+ × Ω → R: [Follmer & Leukert, 2000]
Increasing and concave in w and constant for w ≥ H(ω)︸ ︷︷ ︸“Contingent Claim”
.
Problem: MaximizeE U(VT, ω) over all self-financing trading strategies such
that the associated wealth process Vtt≥0 satisfies
V0 ≤ w︸ ︷︷ ︸Budget Constraint
and V· ≥ 0.︸ ︷︷ ︸Admissibility condition.
23
Page 25
Motivation:
Optimal Hedging of Contingent Claims.
Setting: B−1t St is a semimartingale and the classM of Equivalent
Martingale Measures (EMM) is not empty. Then,
• Hedging away the risk of a contingent claim H is feasible: [Kramkov 97]
Exists admissible Vtt≤T with V0 = w s.t. VT≥ H,
for all
w ≥ w0 := supQ∈M
EQ[B−1
TH
].
• An initial endowment w smaller than w0 entails short-falling risk (sometimes
VT
< H), motivating the problem of minimizing the expected shortfall:
Minimize EL
((H − V
T)+
)s.t. V0 ≤ w and Vt ≥ 0,
for a Loss Function L : R+ → R (L(0) = 0, increasing, and convex).
24
Page 26
• The shortfall minimization problem is equivalent to a utility maximization prob-
lem with (state-dependent) utility
U(z; ω) := L(H(ω))− L((H(ω)− z)+).
0 2 4 6 8 10 12 140
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
WEALTH
UT
ILIT
Y
State−Dependent Utility Function
H
U( ⋅,ω)
25
Page 27
Convex Duality
• Basic idea: Upper bound a maximization problem with constraints, using a
convex minimization problem without constraints:
Primal Problem︷ ︸︸ ︷p∗ := max f(x)
s.t. h(x) ≤ 0=⇒
• f(x) ≤L(x,λ)︷ ︸︸ ︷
f(x)− λh(x), λ ≥ 0
• p∗ ≤ L(λ) := maxx L(x, λ)
Convex
• p∗ ≤ d∗ := minλ≥0
L(λ)︸ ︷︷ ︸
Dual Problem
.
• We say that strong duality holds if p∗ = d∗. This holds if for instance the
primal problem is “convex” and h(x) < 0 (Slater condition).
• Applications: Duality plays an important role in convex optimization:
(i) Perturbation and sensibility analysis; (ii) Interior point methods.
26
Page 28
Convex duality in portfolio optimization problems[Karatzas et. al. 91, Schachermayer-Kramkov 99, Follmer-Leukert 2000]
Primal problem:
p∗(w) := supE U(VT, ω)
such that V0 ≤ w and V· ≥ 0,
Assumption: w < supQ∈M EQ[B−1
TH
]< ∞.
The dual domain Γ: Nonnegative supermartingales ξtt≥0 such that
• 0 ≤ ξ0 ≤ 1 and ξtB−1t Vtt≥0 is a supermaringale for all admissible
wealth processes Vtt≥0.
Consequences:
• E ξ
TB−1
TV
T
≤ ξ0B−10 V0 ≤w if V0 ≤ w.
• The pricing formula Eξ
TB−1
TH
is free of arbitrage.
• “ M⊂ Γ ” : The density processes of EMMs belong to Γ.
27
Page 29
Road towards weak duality:
• For ξ ∈ Γ and any admissible wealth process V· with V0 ≤ w:
E U (VT , ω)≤E U (VT , ω) − λ(E
ξT B−1
TVT
− w)
= EU (V
T, ω)− λ ξ
TB−1
TV
T
+ λw
≤ E
supv≥0
U (v, ω)− λ ξT B−1
Tv
+ λw
= E
U(λξTB−1
T, ω)
+ λw,
where U(λ, ω) := supv≥0 U (v, ω)− λv [Convex Dual Function].
• Let Γ ⊂ Γ a subclass. Then,
p∗(w)= supE U(VT, ω) ≤ inf
ξ∈ΓE
U(λξ
TB−1
T, ω)
︸ ︷︷ ︸d∗Γ(λ)
+λw.
28
Page 30
Dual problem associated to Γ ⊂ Γ:
d∗Γ(λ) := inf
ξ∈ΓE
U(λξT B−1
T, ω)
.
“Weak duality”: p∗(w) ≤ d∗Γ(λ) + λw.
0 0.5 1 1.5 2 2.5 3 3.5 40.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Dual domain
Convex Dual Function
Slope is −H
U(H)
U’(H−)U’(0+)
U(0)
Strictly Convex Part
29
Page 31
Questions:
1. Does strong duality (p∗(w) = d∗Γ(λ) + λw) hold for some λ?
2. Is the dual problem infξ∈Γ E
U(λξTB−1
T, ω)
attainable?
3. Is the primal problem supV :V0≤w E U(VT , ω) attainable?
Solution when Γ = Γ: [Follmer-Leukert (2000) & Xu (2004)]
Yes, if−∞ < EU(0, ω) ≤ EU(H,ω) < ∞. Moreover, we have
a dual characterization of the optimal final wealth:
V ∗T
= I(λ∗ ξ∗
TB−1
T
),
where I(λ, ω) := inf z : U ′(z, ω) < λ .
Our problem:
• Can one specify further ξ∗ for a given market model and a utility function?
• Can one narrow down the dual domain Γ ⊂ Γ where to search ξ∗?
30
Page 32
Portfolio optimization for geometric Levy models[Kallsen (2000), Nualart et. al. (2004), Kunita (2003), Figueroa & Ma (2007)]
Market model: Geometric Levy model and a constant interest rate bond:
(1) dSt = b St−dt + St− dZs, (2) dBt = rBtdt, B0 = 1,
(3) Zt = bt + σWt +∫ t
0
∫z (N(dt, dz)− ν(dz)ds)︸ ︷︷ ︸eN(dt,dz)
,
(4) F = Ftt≥0, where Ft := FZt ∧ P− null set.
Dual domain Γ:
S :=
Xt :=∫ t
0
G(s)dWs +∫ t
0
∫F (s, z)N(ds, dz) : F ≥ − 1
,
Γ := ξ· := ξ0 +∫ ·
0
ξs− d(Xs −As) s.t. ξ ∈ Γ, X ∈ S, A increasing.
31
Page 33
Remark:
ξ = ξ0E(X −A) nonnegative with X ∈ S and A predictable increasing
belongs to Γ if and only if a.s.
b− r + α +∫
|z|≥1
zν(dz) + σG(t) +∫
RzF (t, z)ν(dz) ≤ at,
for almost every t ≤ τ(ω), where a· is the density of the absolutely contin-
uous part of A and τ is the time ξ hits 0.
Dual problem associated with Γ: There exist ξ∗ ∈ Γ and λ∗ > 0 such that
(1) The problem infλ>0
d∗
Γ(λ) + λw
is attained at λ∗
(2) The dual problem d∗Γ(λ∗) := infξ∈Γ E
U(λ∗ξ
TB−1
T, ω)
is attained
at ξ∗ ∈ Γ
(3) infλ>0
d∗
Γ(λ) + λw
= E
[U
(W∗Γ, ω
)], whereW∗
Γ= I
(λ∗ ξ∗
TB−1
T
)
(4) Eξ∗
TB−1
TW∗
Γ
= w.
32
Page 34
Strategy to prove strong duality:
In view of the weak duality property (p∗(w) ≤ d∗(λ∗)), ifW∗Γ
is replicable
with an initial endowment of w, thenW∗Γ
will be the optimal final wealth.
Road towards replicability:
(1) By perturbating the “parameters” of ξ∗, obtain a “variational equality”:
Eξ∗
TB−1
TI
(λ∗B−1
Tξ∗
T
)Y
T
= 0,
for any (G, F , a) in a suitable convex set Λ∗, where
Yt :=∫ t
0
G(s)dWG∗s +
∫ t
0
∫
R0
F (s, z) NF∗(ds, dz)+∫ t
0
asds,
with
WG∗t := Wt −
∫ t
0
G∗(s)ds,
NF∗(dt, dz) := N(dt, dz)− (1 + F ∗(s, z))ν(dz)ds.
33
Page 35
(2) Prove that A∗ in ξ∗ = E (X∗ −A∗) is identically zero, and hence, ξ∗
is a (local) martingale. By the Girsanov Theorem, WG∗ is a standard Brown-
ian Motion and NF∗ is a martingale measure with respect to the (finitely
additive) probability measure dQ∗ := ξ∗T dP (up to stopping times).
(3) Establish the decomposition
M2(Q∗) = L(R)⊕ L (Λ∗0) ,
where
– L(R) is the set of all M ∈M2(Q∗) which can be written as Mt =∫ t
0βsdRs, with dRt := S−1
t dSt, and S is the discounted price process.
– L (Λ∗0) is the stable subspace ofM2(Q∗) generated by the integrals∫ ·
0
G(s)dWG∗s +
∫ ·
0
∫F (s, z) NF∗(ds, dz)
with (G, F ) in an appropriate class Λ∗0.
34
Page 36
(4) Show that theQ∗-martingale,
X∗t := EQ∗
B−1
TI
(λ∗ ξ∗
TB−1
T
) |Ft
admits the representation
X∗t = w +
∫ t
0
βsdRt.
Then, the portfolio with initial value w and trading strategy β∗t := Btβt
is admissible and its final wealth isW∗Γ
= I(λ∗ ξ∗
TB−1
T
).
Conclusions
• The method developed here is more explicit in the sense that the dual do-
main enjoys a concrete parametrization.
• Due to the explicit parametrization, discrete time approximations in particu-
lar cases might be possible to implement.
35
Page 37
• It can accommodate much more general jump-diffusion models driven by Levy
processes such as
dSi(t) = Si(t−)bitdt +
d∑
j=1
σijt dW j
t +∫
Rd
h(t, z)N(dt, dz).
• Also suitable to solve optimum portfolio problems with consumption. The prob-
lem is then to maximize the utility coming from both consumption and final
wealth, utility determined as follows
E
U1(VT ) +
∫ T
0
U2(t, c(t))dt
,
under a budget constraint and a no bankruptcy condition (now the wealth is
determined by dVt := rVtdt + βtdSt − ctdt).
36