Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic...
Transcript of Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic...
![Page 1: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/1.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Variance Swap ModelsTheory and Evidence
Damir Filipovic
Swiss Finance InstituteEcole Polytechnique Federale de Lausanne
(joint with Elise Gourier and Loriano Mancini)
6th World Congress of the Bachelier Finance SocietyToronto, 26 June 2010
![Page 2: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/2.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
English EPFL > CDM > Swiss Finance Ins... > EVENTS > Swissquote Confer...
Swissquote Conference on Interest Rate and Credit Risk
Date: Thursday 28th and Friday 29th October 2010
Venue: Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland
Scope: Interest rate and credit risk constitute the major risk sources for banks,insurance companies, and other financial institutions. This conference brings togetherdifferent perspectives and tools for the valuation and managing of these risks. Itaddresses academics and practitioners, and shall foster the interaction betweenindividuals and across institutions.
Speakers include:
Tomas Björk (Stockholm School of Economics)Giovanni Cesari (UBS)Pierre Collin-Dufresne (Columbia University)Rama Cont (University Paris 6)Jean-Pierre Danthine (Swiss National Bank)Mark Davis (Imperial College)David Lando (Copenhagen Business School)Alex Lipton (Bank of America Merrill Lynch and Imperial College)Dilip Madan (University of Maryland)Fabio Mercurio (Bloomberg)Antoon Pelsser (Maastricht University)Kenneth Singleton (Stanford University)
Scientific Committee: Tomas Björk (Stockholm School of Economics), Mark Davis(Imperial College), Damir Filipovic (EPFL), David Lando (Copenhagen BusinessSchool), Wolfgang Runggaldier (University of Padova), Kenneth Singleton (StanfordUniversity)
Organising Committee at EPFL: Rüdiger Fahlenbrach, Damir Filipovic, JulienHugonnier, Semyon Malamud, Loriano Mancini, Erwan Morellec, Claudia Ravanelli,Gregor Svindland, Anders Trolle
Participation is free, but places at this conference are limited. Please register byemailing [email protected]
Site map • © 2010 EPFL , CDM SFI - Odyssea - 1015 Lausanne, tel. +41 21 693 24 [email protected]
Swissquote Conference http://sfi.epfl.ch/Jahia/site/sfi/pid/85028/cache/off/?print=true&matrix=6647
1 of 2 26.06.2010 00:16
![Page 3: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/3.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Outline
1 Variance Swaps
2 Quadratic Term Structures
3 Quadratic (Pearson) Diffusions
4 Model Estimation
![Page 4: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/4.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Realized Variance
• Filtered probability space (Ω,F , (Ft)t≥0,Q)
• Price process (e.g. S&P 500 index): semimartingale S
• Annualized realized variance on t = t0 < · · · < tk = T :
RV2t,T =
k
n
k∑i=1
(log
Sti
Sti−1
)2
where n = number of trading days per year
• Approximate by quadratic variation, for k →∞:
k∑i=1
(log
Sti
Sti−1
)2
→ [log S ]T − [log S ]t
• Justified for daily sampling
![Page 5: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/5.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Variance Swaps
• A variance swap initiated at t with maturity T pays
RV2t,T −VS2
t,T
• VS2t,T = variance swap rate fixed at t
• Assume deterministic risk-free rate r :
VS2t,T =
1
T − tEQ [[log S ]T − [log S ]t | Ft ]
• Provides hedging instrument against volatility increases,which often coincide with drops of stock prices
![Page 6: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/6.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
S-Characteristics
• Assume
dSt
St−= rt dt +σt dWt +
∫R
(e x − 1) (µ(dt, dx)− νt(dx)dt)
• In particular,
∆ log St =
∫R
x µ(dt, dx)
• Hence
[log S ]T − [log S ]t =
∫ T
tσ2s ds +
∫ T
t
∫R
x2 µ(ds, dx)
![Page 7: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/7.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Model-Free Replication
• Britten–Jones and Neuberger [2], Jiang and Tian [11],Carr and Wu [4], a.o. showed:
[log S ]T − [log S ]t
=
∫ Ft
0
2
K 2(K − ST )+ dK +
∫ ∞Ft
2
K 2(ST − K )+ dK
+ 2
∫ T
t
(1
Fs−− 1
Ft
)dFs
− 2
∫ T
t
∫R
(e x − 1− x − x2
2
)µ(ds, dx)
• Ft = St/P(t,T ) = T -futures price of S
![Page 8: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/8.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Model-Free Valuation
• Taking Q-expectation:
VS2t,T =
2
T − t
∫ ∞0
Θt(K ,T )
P(t,T )K 2dK + ε
with error term
ε = − 2
T − tEQ
[∫ T
t
∫R
(e x − 1− x − x2
2
)νs(dx)ds | Ft
]• Θt(K ,T ) = out-of-the-money option price
![Page 9: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/9.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
VIX
• Chicago Board Options Exchange Volatility Index (VIX)
VIXt =√
VS2t,t+30 days × 100 [%]
calculated as weighted blend of options on S&P 500 index
• Introduced in 1993, revised in 2003
• Industry benchmark for market volatility
![Page 10: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/10.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Term-Structure Models• OTC variance swaps at many different maturities available⇒ design and estimate term-structure models of varianceswap rates and risk premiums!
Figure: Variance swap rates√
VS2t,t+τ on the S&P 500 index from
Jan 4, 1996 to Apr 2, 2007. Source: Bloomberg
![Page 11: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/11.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Forward Variance
• Define the forward variance
f (t,T ) = EQ
[σ2T +
∫R
x2 νT (dx) | Ft
]• Then the variance swap rates equal
VS2t,T =
1
T − tEQ
[∫ T
tσ2s ds +
∫ T
t
∫R
x2 µ(ds, dx) | Ft
]=
1
T − t
∫ T
tf (t, s) ds
• The spot variance is
vt = limT↓t
VS2t,T = f (t, t)
• Note the analogy to yields vs. forward rates
![Page 12: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/12.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Term-Structure Models: Program
• Exogenous (factor) model of f (t,T ), or vt , under P and Q• Define index Q-dynamics . . .
dSt
St−= rt dt+σt dWt+
∫ (e δt(ξ) − 1
)(µ(dt, dξ)− νt(dξ)dt)
• . . . such that spot variance satisfies
vt = σ2t +
∫δt(ξ)2 νt(dξ)
and
d [v , log S ]t ≤ 0 (“leverage effect”)
• E.g. Buehler [3], Egloff et al. [9], Cont and Kokholm [7],a.o.
![Page 13: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/13.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Factor Model
• Forward variance factor model
f (t,T ) = g(T − t,Xt)
• State space X ⊆ R open
• Jump-diffusion state process
dXt = b(Xt)dt+Σ(Xt)dWt+
∫Rξ (µ(dt, dξ)−ν(Xt−, dξ)dt)
• b, Σ, γ, g nice enough . . . in particular linear growth
b(x)2 + Σ(x)2 +
∫Rξ2 ν(x , dξ) ≤ K (1 + x2)
![Page 14: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/14.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Term Structure
Theorem 2.1.The forward variance model admits a quadratic term structure:
g(t, x) = φ(t) + ψ(t)x + π(t)x2
if (and essentially only if) the state process X is quadratic:
b(x) = b + βx
Σ2(x) = a + αx + Ax2
ν(x , dξ) = n(dξ) + ν(dξ)x + N(dξ)x2.
Moreover, . . .
![Page 15: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/15.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Term Structure
Theorem 2.1 (cont’d).
. . . the functions φ, ψ, π satisfy the linear ODE
d
dt
φψπ
=
0 b a +∫R ξ
2 n(dξ)0 β 2b + α +
∫R ξ
2 ν(dξ)0 0 2β + A +
∫R ξ
2 N(dξ)
φψπ
with initials φ0, ψ0, π0 determined by the spot variancefunction
g0(x) ≡ g(0, x) = φ0 + ψ0x + π0x2.
![Page 16: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/16.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Proof
• Kolmogorov backward equation:
∂g(t, x)
∂t= b(x)
∂g(t, x)
∂x+
1
2Σ2(x)
∂2g(t, x)
∂x2
+
∫R
(g(t, x + ξ)− g(t, x)− ∂g(t, x)
∂xξ
)ν(x , dξ)
• Reads here:
φ′(t) + ψ′(t)x + π′(t)x2 = b(x) (ψ(t) + 2π(t)x) + Σ2(x)π(t)
+
∫Rπ(t)ξ2 ν(x , dξ)
= ψ(t)P1(x) + π(t)P2(x)
![Page 17: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/17.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Proof
• Hence
P1(x) = b(x)
P2(x) = 2b(x)x + Σ2(x) +
∫Rξ2 ν(x , dξ)
are quadratic polynomials in x
• Necessity follows for diffusion case (ν(x , dξ) = 0)
• Separate terms in 1, x , x2 yields the result
![Page 18: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/18.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Processes
Theorem 2.2.The process X is quadratic if and only if
E [X nt | X0 = x ] =
n∑k=0
Mkn(t)xk
for all n ≥ 0. Moreover, the (n + 1)× (n + 1)-matrix M solvesthe ODE
d
dtM(t) = BM(t)
M(0) = Id
That is, M(t) = eBt , where . . .
![Page 19: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/19.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Processes
Theorem 2.2 (cont’d).
. . . the matrix B is upper triangular, and reads for the diffusioncase (for simplicity):
B =
0 b 2 a2 0 · · · 0
0 β 2(b + α
2
)3 · 2 a
2 0...
0 0 2(β + A
2
)3(b + 2α2
) . . . 0
0 0 0 3(β + 2A
2
) . . . n(n − 1) a2... 0
. . . n(b + (n − 1)α2
)0 . . . 0 n
(β + (n − 1)A2
)
• Zhou [13], Forman and Sørensen [10], Cuchiero et al. [8]
![Page 20: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/20.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Eigenpolynomials and Moments
• Spectral decomposition
B = S diag (λ0, . . . , λn) S−1
with eigenvalues λk = Bkk , k = 0, . . . , n
• Gives eigenpolynomials pk(x) =∑k
j=0 Sjkx j s.t.
E [pk(Xt) | X0 = x ] = eλk tpk(x), k = 0, . . . , n
• Stationary moments given by first row of S−1:
E[X kt
]= S−1
0k , k = 0, . . . , n
• Very efficiently computable! (e.g. with Mathematica)
![Page 21: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/21.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Quadratic Interest Rate vs.Variance Swap Models
• An interest rate factor model r = r(X ) admits a quadraticterm structure if
E[e−
∫ t0 r(Xs)ds | X0 = x
]= eΦ(t)+Ψ(t)x+Π(t)x2
• Ahn, Dittmar, and Gallant [1], Leippold and Wu [12], a.o.
• Chen, Filipovic, and Poor [5]: the only (!) consistentjump-diffusion state process X is Gaussian
dXt = (b + βXt) dt + Σ dWt
• Quadratic variance swap term-structure models are muchmore flexible!
![Page 22: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/22.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Model Identification
• The quadratic property of X is invariant w.r.t. affinetransformations
X 7→ c + γX
• Can be offset with affine transformation of the quadraticforward variance function:
φ+ψx + πx2 (φ+ ψc + πc2
)+ (ψγ + 2πγ) x + πγ2x2
• Need canonical representation of X for econometric modelidentification!
![Page 23: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/23.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Canonical Representation
• Forman and Sørensen [10]
Theorem 3.1.Denote by D = α2 − 4aA the discriminant of the diffusionfunction of the quadratic diffusion process
dXt = (b + βXt) dt +√
a + αXt + AX 2t dWt .
Suppose A > 0. Then X falls in one of the following threeequivalence classes . . .
![Page 24: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/24.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Class 1: D < 0
• State space X = R• Canonical representative:
dXt = (b + βXt)dt +√
1 + AX 2t dWt
for b ≥ 0 and β ∈ R• If β < 0: ∃ stationary density
µ(x) ∝(1 + Ax2
)β/A−1exp
[2b√
Aarctan[
√Ax ]
](Pearson’s type IV, or skew t-distribution)
• For A→ 0: Gaussian limit with stationary density
µ(x) ∝ exp[2bx + βx2
]
![Page 25: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/25.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Class 2: D = 0
• State space X = (0,∞)
• Canonical representative:
dXt = (b + βXt)dt +√
AX 2t dWt
for b ≥ 0 and β ∈ R• If b > 0 and β < 0: ∃ stationary density
µ(x) ∝ x2β/A−2 exp
[− 2b
Ax
](inverse Gamma distribution)
• Also called GARCH diffusion
![Page 26: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/26.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Class 3: D > 0
• State space X = (0,∞)
• Canonical representative:
dXt = (b + βXt)dt +√
Xt + AX 2t dWt
for b ≥ 12 and β ∈ R
• If β < 0: ∃ stationary density
µ(x) ∝ x2b−1(1 + Ax)2β/A−2b−1
(scaled F -distribution)
• For A→ 0: affine limit case with stationary density
µ(x) ∝ x2b−1 exp[2βx ]
(Gamma distribution)
![Page 27: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/27.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Measure Change
• Aim: equivalent change of measure Q ∼ P preservingquadratic property of X :
dXt =(
bQ + βQXt
)dt
+√
a + αXt + AX 2t
(dWt +
`+ λXt√a + αXt + AX 2
t
dt
)︸ ︷︷ ︸
=dWQt
• With return variance risk premium parameters
` = b − bQ, λ = β − βQ
• Problem: Novikov’s condition fails in general
• But equivalent measure change works here, due to . . .
![Page 28: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/28.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
. . . Measure Change Theorem
Theorem 3.2 (Cheridito, Filipovic, and Yor [6]).
Let b, σ, λ be locally bounded functions on X , and
Xt = X0 +
∫ t
0b(Xs) ds +
∫ t
0Σ(Xs) dWs .
Assume that the martingale problem for
Af (x) = (b(x) + Σ(x)λ(x)) f ′(x) +1
2Σ(x)2f ′′(x)
is well posed in X . Then stochastic exponential
Et(λ(X)> •W ) = exp
(∫ t
0λ(Xs)>dWs −
1
2
∫ t
0‖λ(Xs)‖2 ds
)is a martingale.
![Page 29: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/29.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Application: Martingality of S
• Recall quadratic state process
dXt = (b + βXt) dt +√
a + αXt + AX 2t dW 1
t
• Let ρ : X → [−1, 1] be Lipschitz and s.t. for some x∗ > 0:
ρ(x)
≤ 0, x ≥ x∗,
≥ 0, x ≤ −x∗
• Model the discounted S&P 500 index process as
dSt
St=√
g0(Xt)
(ρ(Xt) dW 1
t +√
1− ρ2(Xt) dW 2t
)• It satisfies (with high probability) the “leverage effect”
d [g0(X ), log S ]t = g ′0(Xt)√
g0(Xt)ρ(Xt) ≤ 0
![Page 30: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/30.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Application: Martingality of S
• Question: is S a true Q-martingale? (vital for pricing!)
• Yes! Write S as stochastic exponential
St = S0 Et(λ(X)> •W
)with
λ(x) =√
g0(x)
(ρ(x)√
1− ρ2(x)
)• . . . and note that the martingale problem for
Af (x) =(
b + βx +√
a + αx + Ax2√
g0(x)ρ(x))
f ′(x)
+1
2
(a + αx + Ax2
)f ′′(x)
is well posed in X (Yamada–Watanabe)
![Page 31: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/31.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Generalized Method of Moments
• Model parameters
θ = (a(= 0, 1), α(= 0, 1),A, b, β, `, λ, φ0, ψ0, π0)
• Observations: 5-vector of VS rates
Yt = (VS2t,t+τ1
, . . . ,VS2t,t+τ5
)>, t = 1, . . . ,T
• Define (martingale increments) vector-valued function
ht = h(Yt ,Yt−1, θ)
such that
1
T
T∑t=1
ht ≈ Eθ0 [h(G (Xt),G (Xt−1), θ0)] = 0
where G (x) = (g(τ1, x), . . . , g(τ5, x))>
• Notice: eigenpolynomials of Xt in closed form: Forman and Sørensen [10] provide explicit optimal
martingale estimating functions for Xt . Problem: Xt is not observed
![Page 32: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/32.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Extended Kalman Filter
• De facto standard estimation method
• Pros:• Faster and more stable than GMM (simulation study)• Uses all data• Filters latent factor Xt (useful for pricing)
• Cons: asymptotically inconsistent (but here finite sample!)
• Approximate state transition equation (QML):
Xt+1 ∼ N (Eθ[Xt+1 | Xt ], varθ[Xt+1 | Xt ])
• Linearize observation equation:
Yt+1 = G (Xt+1|t, θ) + G ′(Xt+1|t, θ)(Xt+1 − Xt+1|t) + εt+1
where Xt+1|t = predicted state
![Page 33: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/33.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Data
Figure: Variance swap rates√
VS2t,t+τ on the S&P 500 index from
Jan 4, 1996 to Apr 2, 2007. Source: Bloomberg
![Page 34: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/34.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Summary Statistics
Panel A: Variance swap ratesMaturity Mean Std Skew Kurt Q22 ADF
2 20.76 6.80 0.87 4.09 49,806.04 −3.843 20.90 6.54 0.78 3.87 52,308.54 −3.656 21.48 6.32 0.78 3.93 54,570.82 −3.45
12 22.25 6.06 0.62 3.19 56,549.61 −3.0724 22.86 5.90 0.55 2.75 57,460.86 −2.86
Panel B: Calculated VIX2 20.68 6.14 0.71 3.38 51,405.71 −3.633 20.71 5.80 0.63 3.20 51,697.76 −3.406 20.78 5.23 0.47 2.79 55,149.03 −3.22
Table: Panel A: Summary statistics of the variance swap rates on theS&P 500 index at different maturities (in months) from January 4,1996 to April 2, 2007, for a total of 2832 observations.
![Page 35: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/35.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Summary Statistics cont’d
• Panel A cont’d: The table reports mean, standarddeviation (Std), skewness (Skew), kurtosis (Kurt); theLjung–Box portmanteau test for up to 22nd orderautocorrelation, Q22, 10% critical value is 30.81; theaugmented Dickey–Fuller test for unit root involving 22augmentation lags, a constant term and time trend, ADF,10% critical value is −3.16.
• Panel B: summary statistics of the two-, three- andsix-month VIX calculated using SPX options and applyingthe revised CBOE VIX methodology.
![Page 36: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/36.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Principal Component Analysis
• PCA of variance swap curve τ 7→√
VS2t,t+τ
• One major factor (level), explains 96% of variance
• Second factor (slope), explains 3% of variance
Figure: First two variance swap curve loadings
![Page 37: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/37.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Model Specifications
• Full specification: no restrictions
dXt = (b + βXt) dt +√
a + αXt + AXt dWt
f (t, t) = φ0 + ψ0Xt + π0X 2t
• Nested restricted specifications:• A = 0: affine X -dynamics• π = 0: linear spot variance function• ψ2
0 = 4φ0π0: spot variance function has exactly one zero• φ0 = ψ0 = 0: sv function has exactly one zero at x = 0
• Report likelihood ratio
LR = 2 (log LFull − log LREST) ∼ χ2# rest
![Page 38: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/38.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Estimation Results: Overview
Type Full A = 0 π = 0 ψ20 = 4φ0π0 φ0 = ψ0 = 0
a 0 0 0 0 0α 1 1 1 0 0A 0.402 0 1.620 0.462 0.320b 2.005 0.529 4.083 0.979 0.098β -0.742 -0.264 -0.839 -1.034 -0.849` -0.023 -0.003 -1.273 0.300 0.032λ -0.243 0.020 -1.503 -0.590 -0.458φ0 0.016 0.000 0.693 0.002 0ψ0 -0.002 0.007 0.008 0.010 0π0 0.002 0.009 0.003 0.013 1.640LR 0 386 620 264 286
![Page 39: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/39.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Estimation Results
• Full model of class 3 gives best fit
dXt = (b + βXt) dt +√
Xt + AX 2t dWt
• All nested restricted specifications strongly rejected, inparticular the affine ones (“A = 0”, “π = 0”)
• Class 3 combines affine behavior for small Xt andquadratic behavior for large Xt
• Quadratic terms allow for extreme movements and humpshaped VS term structure
• Drawback: spot variance function bounded away from zero(≥ 0.12)
![Page 40: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/40.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
In-Sample Analysis: Predicted VS
Figure: Very good fit of predicted (filtered) variance swap rates vs.data for 6 months maturity
![Page 41: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/41.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
In-Sample Analysis: Humps
Figure: Hump shaped VS term structure on 15-Dec-1998. Quadraticmodel (left), linear model (π = 0) right.
![Page 42: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/42.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Out-of-Sample Analysis: VIXFutures Pricing
Figure: Challenging exercise for the model: VIX futures on Feb 15,2007. Maturities 30, 60, 90, 120, 150, 180, 270, 420 days.Systematic pricing error for VIX due to jumps. Acceptable result.
![Page 43: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/43.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Nonlinear Phenomena
Figure: Simulation of a spot volatility trajectory over 100 years. Inblack the linearized model. We see volatility spikes and clustering.Do we need jumps after all?
![Page 44: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/44.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
Conclusion
• Need for variance swap term-structure models
• Quadratic term structure (closed form) led to quadraticfactor process
• Quadratic models are much more flexible than linear-affinemodels
• Data imply strong statistical evidence for quadratic terms
• Quadratic models capture nonlinear phenomena: rareevents, volatility clustering
![Page 45: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/45.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
References I
D. H. Ahn, R. F. Dittmar, and A. R. Gallant.Quadratic term structure models: Theory and evidence.The Review of Financial Studies, 15:243–288, 2002.
M. Britten-Jones and A. Neuberger.Option prices, implied price processes and stochasticvolatility.Journal of Finance, 55(2):839–866, 2000.
H. Buehler.Consistent variance curve models.Finance Stoch., 10(2):178–203, 2006.
P. Carr and L. Wu.Variance risk premiums.The Review of Financial Studies, 22(3):1311–1341, 2009.
![Page 46: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/46.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
References II
L. Chen, D. Filipovic, and H. V. Poor.Quadratic term structure models for risk-free anddefaultable rates.Math. Finance, 14(4):515–536, 2004.
P. Cheridito, D. Filipovic, and M. Yor.Equivalent and absolutely continuous measure changes forjump-diffusion processes.Ann. Appl. Probab., 15(3):1713–1732, 2005.
R. Cont and T. Kokholm.A Consistent Pricing Model for Index Options andVolatility Derivatives.SSRN eLibrary, 2009.
![Page 47: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/47.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
References III
C. Cuchiero, M. Keller-Ressel, and J. Teichmann.Polynomial processes and their applications inmathematical finance.Working Paper, 2008.
D. Egloff, M. Leippold, and L. Wu.The term structure of varaince swap rates and optimalvariance swap investments.Journal of Financial annd Quantitative Analysis,forthcoming, 2010.
J. L. Forman and M. Sørensen.The Pearson diffusions: a class of statistically tractablediffusion processes.Scand. J. Statist., 35(3):438–465, 2008.
![Page 48: Quadratic Variance Swap Models - Theory and …...(Pearson) Di usions Model Estimation Quadratic Interest Rate vs. Variance Swap Models An interest rate factor model r= (X) admits](https://reader034.fdocuments.us/reader034/viewer/2022042713/5faf9cfa03383b12b00da5ca/html5/thumbnails/48.jpg)
QuadraticVariance Swap
Models
D. Filipovic
VarianceSwaps
QuadraticTermStructures
Quadratic(Pearson)Diffusions
ModelEstimation
References IV
G. J. Jiang and Y. S. Tian.The model-free implied volatility and its informationcontent.The Review of Financial Studies, 18(4):1305–1342, 2005.
M. Leippold and L. Wu.Asset pricing under the quadratic class.Journal of Financial and Quantitative Analysis,37:271–295, 2002.
H. Zhou.Ito conditional moment generator and the estimation ofshort-rate processes.J. Financial Econometrics, 1:250–271, 2003.