Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convex Regularization of Local Volatility Models from Option Prices: Convergence Analysis and
Rates
Adriano De Cezaro e-mail: decezaro@impa.br
Joint work withJ. P. Zubelli (IMPA)andO. Scherzer (Univ.
Vienna)
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Plan
1 Asset Models and Option Pricing 2 The Formulation of the Inverse Problem
3 Properties ofF and ill-posedness of the inverse problem 4 State of the art from the literature
5 Non-quadratic regularization of calibration problem Convergence
Convergence rates 6 Exponential Families 7 Conclusions
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Derivative Contracts
European Call Option: a forward contract that gives the holder the right, but not the obligation, to buy one unit of an underlying asset for an agreedstrike price K on thematuritydateT.
Its payoff is given by h(XT) =
XT−K ifXT>K, 0 ifXT≤K.
Fundamental QuestionHow to price such an obligation given today’s information?
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Derivative Contracts
European Call Option: a forward contract that gives the holder the right, but not the obligation, to buy one unit of an underlying asset for an agreedstrike price K on thematuritydateT.
Its payoff is given by h(XT) =
XT−K ifXT>K, 0 ifXT≤K.
Fundamental QuestionHow to price such an obligation given today’s information?
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Main Contributions
L. Bachelier (Paris) P. Samuelson F. Black M. Scholes R. Merton
recognized by the Nobel prize in Economics award to R. Merton and M. Scholes
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Black-Scholes Market Model
We consider the Black-Scholes equation (see Black-Scholes-1973):
Ut+1
2σ2(t,X)X2UXX+ (r−q)XUX =rU, (1) Ut,S(t=T,X) =max(X−K,0) European Call (2) Here,X =X(t)denotes the spot price,K is called strike,T the maturity,ris the interesting rate,qthe dividends andσ(t,X)is the local volatility.
Note 1U=U(t,x;σ,r)fort≤T. Note 2Final Value Problem
RemarkConstant volatility model - Nobel Prize: Robert Merton and Myron Scholes
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Black-Scholes Market Model
We consider the Black-Scholes equation (see Black-Scholes-1973):
Ut+1
2σ2(t,X)X2UXX+ (r−q)XUX =rU, (1) Ut,S(t=T,X) =max(X−K,0) European Call (2) Here,X =X(t)denotes the spot price,K is called strike,T the maturity,ris the interesting rate,qthe dividends andσ(t,X)is the local volatility.
Note 1U=U(t,x;σ,r)fort≤T.
Note 2Final Value Problem
RemarkConstant volatility model - Nobel Prize: Robert Merton and Myron Scholes
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Black-Scholes Market Model
We consider the Black-Scholes equation (see Black-Scholes-1973):
Ut+1
2σ2(t,X)X2UXX+ (r−q)XUX =rU, (1) Ut,S(t=T,X) =max(X−K,0) European Call (2) Here,X =X(t)denotes the spot price,K is called strike,T the maturity,ris the interesting rate,qthe dividends andσ(t,X)is the local volatility.
Note 1U=U(t,x;σ,r)fort≤T. Note 2Final Value Problem
RemarkConstant volatility model - Nobel Prize: Robert Merton and Myron Scholes
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Smile effect: Empirical remark that calls having different strikes, but otherwise identical, have different implied volatilities.
Before the 1987 crash the graph ofI(K)(fixt,X,T) had aUshape with a minimum close toK =X.
Since 1987 it is a decreasing function in the range
95%<K/X<105%then (forK >>X) it bends upwards. Limitations ...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Smile effect: Empirical remark that calls having different strikes, but otherwise identical, have different implied volatilities. Before the 1987 crash the graph ofI(K)(fixt,X,T) had aUshape with a minimum close toK =X.
Since 1987 it is a decreasing function in the range
95%<K/X<105%then (forK >>X) it bends upwards. Limitations ...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Smile effect: Empirical remark that calls having different strikes, but otherwise identical, have different implied volatilities. Before the 1987 crash the graph ofI(K)(fixt,X,T) had aUshape with a minimum close toK =X.
Since 1987 it is a decreasing function in the range
95%<K/X<105%then (forK >>X) it bends upwards.
Limitations ...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Smile effect: Empirical remark that calls having different strikes, but otherwise identical, have different implied volatilities. Before the 1987 crash the graph ofI(K)(fixt,X,T) had aUshape with a minimum close toK =X.
Since 1987 it is a decreasing function in the range
95%<K/X<105%then (forK >>X) it bends upwards.
Limitations ...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Local Volatility Models
IdeaAssume that the volatility is given by σ=σ(t,x) i.e.: it depends on time and the asset price.
The Direct Problem
Givenσ=σ(t,x)and the payoff information, determine U=U(t,x,T,K;σ)
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
The Inverse Problem
Given a set of observed prices
{U=U(t,x,T,K;σ)}(T,K)∈S find the volatilityσ=σ(t,x).
The setS is taken typically as[T1,T2]×[K1,K2]. In PracticeVery limited and scarce data
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
The Smile Curve and Dupire’s Equation
Assuming that there exists a local volatility functionσ=σ(S,t) Dupire(1994) showed that the call price satisfies
∂TU−12σ2(K,T)K2∂2KU+rS∂KU=0 , S>0 ,t<T
U(K,T =0) = (S−K)+, (3)
Theoreticalway of evaluating the local volatility
σ(K,T) = s
2
∂TU+rK∂KU K2∂2KU
(4) In practiceTo estimateσfrom (3), limited amount of discrete data and thus interpolate.Numerical instabilities!Even to keep the argument positive is hard.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
The Smile Curve and Dupire’s Equation
Assuming that there exists a local volatility functionσ=σ(S,t) Dupire(1994) showed that the call price satisfies
∂TU−12σ2(K,T)K2∂2KU+rS∂KU=0 , S>0 ,t<T
U(K,T =0) = (S−K)+, (3)
Theoreticalway of evaluating the local volatility σ(K,T) =
s 2
∂TU+rK∂KU K2∂2KU
(4) In practiceTo estimateσfrom (3), limited amount of discrete data and thus interpolate.
Numerical instabilities!Even to keep the argument positive is hard.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
The Smile Curve and Dupire’s Equation
Assuming that there exists a local volatility functionσ=σ(S,t) Dupire(1994) showed that the call price satisfies
∂TU−12σ2(K,T)K2∂2KU+rS∂KU=0 , S>0 ,t<T
U(K,T =0) = (S−K)+, (3)
Theoreticalway of evaluating the local volatility σ(K,T) =
s 2
∂TU+rK∂KU K2∂2KU
(4) In practiceTo estimateσfrom (3), limited amount of discrete data and thus interpolate.Numerical instabilities!Even to keep the argument positive is hard.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Early Results
Bouchoev-Isakov Uniqueness and “Stability”
Consider the case oftime-independent volatilityand working with y=log(K/S(0)) τ=T−t. SupposeU(y,τ)anda(y) =σ(K) satisfies (3) with
U(y,0) =S(0)(1−ey)+,y ∈R (5) U(y,τ∗) =U∗(y),y∈I (6) whereIis a sub-interval ofR. Then, we have
Uniqueness of the volatility
Stability of the volatility in the H ¨olderλ-norm w.r.t. to variations of the data on the 2+λ-norm. (i.e., one needs TWO extra
derivatives of the data).
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Early Results
More precisely
Theorem (Bouchouev & Isakov)
Let U1and U2be solutions of (3-6) with a=a1and a=a2, resp., and the corresponding final data in Eq. (6) given by U1∗and U2∗. Let I0be an open interval with I⊃I06=0/. Then,
1 If U1∗=U2∗on I and a1(y) =a2(y)on I0then a1(y) =a2(y)on I.
2 If, in addition, a1(y) =a2(y)on I0∪(R\I)and I is bounded, then
∃C=C(|a1|λ(I),|a2|λ(I),I,I0,τ∗)s.t.
|a1−a2|λ(I)≤ |U1∗−U2∗|2+λ(I)
RemarkThe assumption thatais known onI0⊂Imakes the problem overdetermined.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Early Results
More precisely
Theorem (Bouchouev & Isakov)
Let U1and U2be solutions of (3-6) with a=a1and a=a2, resp., and the corresponding final data in Eq. (6) given by U1∗and U2∗. Let I0be an open interval with I⊃I06=0/. Then,
1 If U1∗=U2∗on I and a1(y) =a2(y)on I0then a1(y) =a2(y)on I.
2 If, in addition, a1(y) =a2(y)on I0∪(R\I)and I is bounded, then
∃C=C(|a1|λ(I),|a2|λ(I),I,I0,τ∗)s.t.
|a1−a2|λ(I)≤ |U1∗−U2∗|2+λ(I)
RemarkThe assumption thatais known onI0⊂Imakes the problem overdetermined.
Adriano De Cezaro e-mail: decezaro@impa.br Convex Regularization of Local Volatility Models from Option Prices: Convergence Analysis and Rates
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Joint work with J.P. Zubelli & O. Scherzer
Start from
−Ut+1
2σ2(T,K)K2UKK−(r−q)KUK−qU=0 (7) Ut,X(T=t,K) = (X−K)+, for K>0, (8) SettingK=Xey,τ=T−t,u(τ,y) =exp(
RT t
q(s)ds)U(.,T,K), b(τ) =q(τ)−r(τ)and
a(τ,y) =1
2σ2(τ;K) (9)
yields
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
−uτ+a(τ,y)(uyy−uy) +b(τ)uy =0 (10) u(0,y) =X(1−ey)+. (11) The parameter-to-solution mapF is defined by
F(a) =u(a), (12)
whereu(a)is the solution of (10) fora∈D(F).
We assume noise datauδsatisfies the inequality
||u∗−uδ||L2(Ω)≤δ. (13)
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Properties ofF and ill-posedness of the inverse problem
i) There exists ap∗>2 such thatF :D(F)→Wp2,1(Ω)is continuous and compact for 2≤p<p∗. Moreover,F is weakly (sequentially) continuous and thus weakly closed.
ii) F is Gateaux differentiable w.r.t. a∈D(F)in directionshsuch thata+h∈D(F), and the derivativeF0(a)extends as a linear operator toH1(Ω).
iii) F0(a)is injective and compact. MoreoverR(F0(a)∗) =H1(Ω).
Proof:see DC & Scherzer & Zubelli (2009) and Egger & Engl (2005) or Crepey (2003).
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Ill-posedness in more details
D(F):={a∈a0+H1 :a≤a,a0≤a}. ak *a˜ withuk =F(ak)→u˜=F(a˜)
similar pricesuk andu˜ linked with completely different volatilities
ask for regularization!!!! Tikhonov regularization
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Ill-posedness in more details
D(F):={a∈a0+H1 :a≤a,a0≤a}. ak *a˜ withuk =F(ak)→u˜=F(a˜)
similar pricesuk andu˜ linked with completely different volatilities ask for regularization!!!!
Tikhonov regularization
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Ill-posedness in more details
D(F):={a∈a0+H1 :a≤a,a0≤a}. ak *a˜ withuk =F(ak)→u˜=F(a˜)
similar pricesuk andu˜ linked with completely different volatilities ask for regularization!!!!
Tikhonov regularization
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Standard Tikhonov regularizationresidual norm +βtimes penalization penalization=|| · ||2.
source wise conditionequivalentwhat we assume know about the inverse solution
convergence rates to Tikhonov for quadratic penalization
||aδ−a†||=O(√δ)
Our approach: convex regularizationβf(·)withf convex, positive, etc...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Standard Tikhonov regularizationresidual norm +βtimes penalization penalization=|| · ||2.
source wise conditionequivalentwhat we assume know about the inverse solution
convergence rates to Tikhonov for quadratic penalization
||aδ−a†||=O(√δ)
Our approach: convex regularizationβf(·)withf convex, positive, etc...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Standard Tikhonov regularizationresidual norm +βtimes penalization penalization=|| · ||2.
source wise conditionequivalentwhat we assume know about the inverse solution
convergence rates to Tikhonov for quadratic penalization
||aδ−a†||=O(√δ)
Our approach: convex regularizationβf(·)withf convex, positive, etc...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Standard Tikhonov regularizationresidual norm +βtimes penalization penalization=|| · ||2.
source wise conditionequivalentwhat we assume know about the inverse solution
convergence rates to Tikhonov for quadratic penalization
||aδ−a†||=O(√δ)
Our approach: convex regularizationβf(·)withf convex, positive, etc...
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
State of the art from the literature
Let the standard choice of regularization parameterβ=β(δ)∼δ. (i) Iff(·) =|| · ||2H1(Ω), then we have the convergence rate results
(Egger & Engl(2005))
||aδβ−a†||H1(Ω)=O(√δ)and||F(aδβ)−uδ||L2(Ω)=O(δ), (14)
assuming the source-wise representationa∗−a†=F0(a†)∗w.
(ii) In (Hoffman & Kramer (2005)), with
f(a(τ)) =RI{a(τ)lnaa¯(τ)(τ)+a¯(τ)−a(τ)}dτthe convergence rate results
||aδβ−a†||L1([0,T])=O(√δ) (15)
using a source-wise representationlnaa†∗ =F0(a†)∗w.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
iii) Withf(·) =|| · ||2L2([0,T]), (Hoffman et al. (2006)) obtained the same rates of (Hoffman & Kramer (2005) (volatility
time-dependent only) assuming the source wise representation ξ†=F0(a†)∗w∈∂f(a†) (16) in terms of Bregman distances.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Our approach
Minimize the functional
Fβ,uδ(a):=1
2||F(a)−uδ||2L2(Ω)+βf(a), (17) wheref :dom(f)⊂B1→[0,∞]is a convex, proper and lower
semi-continuous stabilization functional.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Convergence
Theorem (Existence, Stability, Convergence)
Suppose that F , f ,D(F)as above,β>0and(13)holds. Then There exists a minimizer ofFβ,uδ.
If(uk)→u in L2(Ω), then every sequence(ak)with ak ∈argmin
Fβ,uk(a):a∈D(F)
has a subsequence which weak converges. The limit of every w-convergent subsequence(ak0)of(ak)is a minimizera of˜ Fβ,u, and f(ak0)
converges to f(a˜).
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Theorem (Semi-convergence)
If there exists a solution of (12)inD(F), then there exists an f -minimizing solution of (12).
Assume that (12)has a solution inD(F)(which implies the existence of an f -minimizing solution) and thatβ:(0,∞)→(0,∞) satisfies
β(δ)→0and δ2
β(δ) →0, asδ→0. (18) Moreover,(δk)→0, and that uk :=uδk satisfiesku¯−ukk ≤δk. Setβk :=β(δk). Then, every sequence(ak)of elements
minimizingFβk,uk, has a subsequence(ak0)that is w-convergent.
The limit a†of any w-convergent subsequence(ak0)is an f -minimizing solution of (12), and f(ak)→f(a†).
Proof:see DC & Scherzer & Zubelli (2009).
Adriano De Cezaro e-mail: decezaro@impa.br Convex Regularization of Local Volatility Models from Option Prices: Convergence Analysis and Rates
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Lemma
Letζ†∈∂f(a†). Then There exists a function w†∈L2(Ω)and a function r∈H1(Ω)such that
ζ†=F0(a†)∗w†+r (19) holds. Furthermore,krkH1(Ω)can to be taken arbitrarily small.
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Definition
Let1≤q<∞. Moreover, letU be a subset of H˜ 1. The Bregman distance Dζ(·,a˜)of f :H1→R∪ {+∞}ata˜∈DB(f)andζ∈∂f is said to be q-coercive with constant c>0if
Dζ(a,a˜)≥cka−a˜kq˜
U ∀a∈D(f). (20)
Outline Asset Models and Option Pricing The Formulation of the Inverse Problem Properties ofFand ill-posedness of the inverse problem State of the art from the literature Non-quadratic regularization of calibration problem Exponential Families Conclusions
Convergence Convergence rates
Convergence rates
Lemma
Letζ†∈∂f(a†)satisfy(19)with w†and r such that c Ckw†kL1(Ω)+krkL2(Ω)
:=β1∈[0,1),
and the Bregman distance with respect to f is1−coercive (as in the Definition 6.1) withU˜ :=H1(Ω). Then,
hζ†,a†−ai ≤β1Dζ†(a,a†) +β2kF(a)−F(a†)kL2(Ω), (21) for a∈Mβmax(ρ), whereβmax,ρ>0satisfy the relationρ>βmaxf(a†).