Volatility Dynamics and Option Prices
Loriano Mancini
Swiss Finance Institute andEcole Polytechnique Federale de Lausanne
Spring School and Workshop on Volatility Dynamics and OptionPrices, and Econometrics of Intraday Data, 10–13 April 2017
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
2/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Dynamics 3/69
Asset Returns
I Many asset classes
I Equities, bonds, commodities, currencies, etc.
I Significantly different economic features
I Empirical regularities of asset returns or stylized facts
I Surprisingly similar
I Model-free phenomena (essentially)
I Impose constraints on models of asset returns
I Relevant for asset and risk management, derivative pricing, etc.
Volatility Dynamics 4/69
Market Returns: S&P500 Index
1982 1987 1993 1998 2004 2009 2015−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
Figure: Daily log-returns, log(St/St−1), of the S&P500 equity marketindex St . Note: i) Stochastic volatility, ii) volatility clustering, and iii)large market returns, are evident. Other asset returns exhibit similarphenomena.
Volatility Dynamics 5/69
Asset Returns: Stylized facts
I Time variation of asset returns or volatility
I Stochastic (i.e., changes randomly over time)
I Persistent (i.e., temporal dependence or volatility clustering)
I Price discontinuities (i.e., jumps)
I Heavy tails of return distributions
I Leverage effect, i.e., Cov[asset return, volatility change] < 0
I . . .
Volatility Dynamics 6/69
Two Modeling Frameworks for Asset Returns and Volatility
I Discrete-time models
I Inference :)
I Prototype model: GARCH model
I Engle (1982), Bollerslev (1986)
I Continuous-time models
I “Closed-forms” :)
I Prototype model: Square-root model
I Heston (1993)
Volatility Dynamics 7/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Dynamics 8/69
GARCH model: Classic Specification
I Asset log-return rt = log(St/St−1)
rt = µ+ εt
where εt = σt zt , and zt is an i.i.d. shock D(0, 1)
I GARCH(1,1) conditional variance σ2t = Var[rt |Ft−1]
σ2t = β0 + β1σ
2t−1 + β2ε
2t−1
where Ft−1 is the information set at time t − 1
I Plus other conditions to ensure σ2t > 0, stationary, etc.
⇒ σ2t is stochastic, persistent, and known at time t − 1
Volatility Dynamics 9/69
GARCH model: Maximum Likelihood Estimation
I Joint density of log-returns r1, . . . , rT from p.d.f. f (rt |Ft−1)
I Log-likelihood function, assuming rt |Ft−1 ∼ N (µ, σ2t )
log f (r1, . . . , rT ) =T∑
t=1
log f (rt |Ft−1)
= −1
2
T∑t=1
[log(2π) + log σ2
t + (rt − µ)2/σ2t
]
I Maximize log-likelihood w.r.t. µ, β0, β1, β2
Volatility Dynamics 10/69
Market Returns and GARCH Volatility
1982 1987 1993 1998 2004 2009 2015
−0.2
−0.1
0
0.1
1982 1987 1993 1998 2004 2009 20150
0.2
0.4
0.6
0.8
1
Figure: Upper graph: Daily S&P500 log-returns. Lower graph: EstimatedGARCH volatility σt annualized. Note: GARCH volatility tracks marketreturn variability.
Volatility Dynamics 11/69
Market Returns and GARCH Innovations
1982 1987 1993 1998 2004 2009 2015
−0.2
−0.1
0
0.1
1982 1987 1993 1998 2004 2009 2015−10
−5
0
5
Figure: Upper graph: Daily S&P500 log-returns. Lower graph: EstimatedGARCH innovations zt = (rt − µ)/σt . Note: GARCH innovations exhibitmuch less volatility clustering than market returns.
Volatility Dynamics 12/69
Temporal Dependence
0 5 10 15 20 25 30 35 40 45−0.1
0
0.1
0.2
0.3
Lag
AC
F r t2
0 5 10 15 20 25 30 35 40 45−0.1
0
0.1
0.2
0.3
Lag
AC
F z
t2
Figure: Upper graph: Sample autocorrelations of squared daily S&P500log-returns. Lower graph: Sample autocorrelations of squared GARCHinnovations (zt)2, where zt = (rt − µ)/σt . Note: Squared market returnsexhibit strong temporal dependence (volatility clustering), squaredGARCH innovations do not.
Volatility Dynamics 13/69
GARCH Models: Some Common Specifications
I Return innovation zt (driving rt = µ+ σt zt) non-normal
I e.g., Bollerslev (1987), Engle and Gonzalez-Rivera (1991)
I Asymmetric GARCH (to capture “leverage” effects)
σ2t = β0 + β1σ
2t−1 + β2ε
2t−1 + β3It−1ε
2t−1
where It−1 = 1 when εt−1 < 0, and It−1 = 0 otherwise
I Glosten, Jagannathan, and Runkle (1993)
I Long-run and short-run component GARCH
σ2t = qt + β1(σ2
t−1 − qt−1) + β2 η1,t−1
qt = γ0 + γ1qt−1 + γ2 η2,t−1
where η1,t−1 and η2,t−1 are zero-mean innovations
I Christoffersen et al. (2008)Volatility Dynamics 14/69
GARCH Models Nowadays
I Widely used to estimate and forecast conditional variances
I Largely applied for risk management, derivative pricing, etc.
I Multivariate GARCH models well-developed
I e.g., Bauwens, Laurent, and Rombouts (2006)
Volatility Dynamics 15/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Dynamics 16/69
Continuous-time Stochastic Volatility Models
I Instantaneous variance vt of asset returns is unobservable
I Call for filtering, inference not straightforward
I Heston model, or square-root diffusion, prototype model
dSt/St = µ dt +√
vt dW1,t
dvt = κ(θ − vt)dt + σ√
vt dW2,t
where dW1,t and dW2,t Brownian increments (correlation ρ),speed of mean reversion κ > 0, long run mean θ >0, etc.
I Exercise: Dynamics of√
vt (use Ito’s Lemma)?
I Affine models, drift and variance of vt are affine in vt
I Characteristic function of log(ST/St) in closed-form
E[e iu log(ST /St )|Ft ] = eC(T−t)+D(T−t)vt
Volatility Dynamics 17/69
Exercise: Moment Generating Function of Log-returnsI Given the Heston model in the previous page
I Use Ito’s Lemma to get the dynamics of xt = log(St)
Goal: Derive the moment generating function E[eu log(ST /St )|Ft ]I Define the martingale process ζt = E[euxT |Ft ]I Guess the functional form ζt = euxt +C(T−t)+D(T−t)vt
I Use Ito’s Lemma to get the dynamics of ζt
dζt =∂ζt
∂tdt+
∂ζt
∂xtdxt+
1
2
∂2ζt
∂x2t
(dxt)2+∂ζt
∂vtdvt+
1
2
∂2ζt
∂v 2t
(dvt)2+∂2ζt
∂xt∂vt(dxt dvt)
I Plug in all partial derivatives, and dxt and dvt termsI Drift(ζt) = 0, ∀dt and ∀vt dt, because ζt is a martingaleI Setting dt terms = 0 and vt dt terms = 0 gives two ODEs,
one for D(T − t) and one for C (T − t)
D ′ =1
2σ2D2 + (uρσ − κ)D +
1
2u2 − 1
2u
C ′ = uµ+ Dκθ
with initial conditions D(0) = 0 and C (0) = 0Volatility Dynamics 18/69
Continuous-time Model: Common Specification
I Remaining in the class of affine models
I Two-factor affine jump diffusion (to increase “vol of vol”):
dSt/St− = µ dt +√
vt dW1,t + (eJt − 1)dNt − νt dt
dvt = κv (mt − vt)dt + σv√
vt dW2,t
dmt = κm(θm −mt)dt + σm√
mt dW3,t
where jump intensity νt is affine, plus other conditions
I From Transform Analysis of Duffie, Pan, and Singleton (2000)
E[
exp
(−∫ T
tγ · Xs ds
)eu·XT |Ft
]= eA(T−t)+B(T−t)·Xt
where Xt = [log(St), vt ,mt ] is the state vector, andA(T − t) and B(T − t) solve ODEs
Volatility Dynamics 19/69
Volatility Dynamics from Two-factor Affine Model
97 98 99 00 01 02 03 04 05 06 07 08 09 100
10
20
30
40
50
60
70
80
v1/2
m1/2
Figure: Estimated instantaneous volatility,√
vt , and its stochastic longrun mean,
√mt , from S&P500 data, using a two-factor affine jump
diffusion model. Volatility annualized and in percentage units.
Volatility Dynamics 20/69
Continuous-time Stochastic Volatility Models Nowadays
I Multi-factor models
I e.g., Gruber, Tebaldi, and Trojani (2015)
I “Unspanned” risks
I e.g., Andersen, Fusari, and Todorov (2015)
I Non-affine models
I e.g., Filipovic, Gourier, and Mancini (2016)
I . . .
Volatility Dynamics 21/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Option Pricing 22/69
DerivativesI No doubt things can go wrong with derivatives (see subprime)
I Derivative markets serve the fundamental purpose of risksharing by meeting hedging and speculative demands
I Option markets are large and fast growingI e.g., S&P500 options is a multi-trillion market
1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
Num
ber
of options
×104
0
2
4
6
8
10
12Number of option contracts by year
Calls
Puts
Figure: Number of out-of-the-money puts and calls on S&P500.Option Pricing 23/69
European Call and Put Options
I Options are financial instruments
I Call / Put option
I Written on an underlying asset S (e.g., S&P500 index)
I Gives the right, but not the obligation, to buy / sell theunderlying asset at the strike price K
I European options can only be exercised at maturity T
I Option price determined at time t < TI Paid by the option buyer to the option seller (or writer)
Option Pricing 24/69
What determines an option price?
I Put option on market index pays out when market drops, i.e.,an insurance contract against market drop
80 85 90 95 100 105 110 115 1200
2
4
6
8
10
12
14
16
18
20
(K −
ST)+
ST
Figure: Put option payoff (K − ST )+ with strike price K = 100.
I Put option price at time t < T depends on
I Objective probability of market drop, and
I Investors’ preferences for protection, risk aversion, etc.⇒ Investors’ preferences lead to risk premiaOption Pricing 25/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Option Pricing 26/69
Equilibrium Model: Intuition
I One period exchange economy with representative agent
I One consumption good, one risky asset ψ
I One representative agent has endowment wt , can buy ξ unitsof the risky asset, consumes ct , discounts future utility u(cT )at β, and at terminal time T > t receives the payoff of the riskasset and consumes everything
I Representative agent maximizes the expected utility
maxξ
u(ct) + EP[βu(cT )|Ft ]
subject to ct = wt − ξψt , with cT = ξψT
I First order condition
0 = u′(ct)(−ψt) + EP[βu′(cT )ψT |Ft ]Option Pricing 27/69
Fundamental Asset Pricing EquationI First order condition ⇒ Fundamental Asset Pricing Equation
ψt = EP[β
u′(cT )
u′(ct)ψT |Ft
]= EP [Mt,T ψT |Ft ]
I When ψ is the risk free bond, ψt = e−r(T−t) and ψT = 1
e−r(T−t) = EP[β
u′(cT )
u′(ct)|Ft
]I When ψ is the risky asset, taking the ratio, and solving for ψt
ψt = e−r(T−t)EP[
βu′(cT )/u′(ct)
EP [βu′(cT )/u′(ct)|Ft ]ψT |Ft
]= e−r(T−t)
∫dPt,T
βu′(cT )/u′(ct)∫dPt,Tβu′(cT )/u′(ct)
ψT
= e−r(T−t)
∫dQt,T ψT = EQ
[e−r(T−t)ψT |Ft
]Option Pricing 28/69
Risk Neutral Pricing
I Risk neutral measure, or pricing measure
dQt,T = dPt,Tβu′(cT )/u′(ct)
EP[βu′(cT )/u′(ct)|Ft ]= dPt,T
Mt,T
EP[Mt,T |Ft ]
I Risky assets can be priced as if investors were risk neutral
ψt = EQ[e−r(T−t)ψT |Ft
]I Stochastic discount factor Mt,T
Mt,T = βu′(cT )
u′(ct)= e−r(T−t) dQt,T
dPt,T
I Exercise: Derive Mt,T when u(c) = c1−γ/(1− γ) has CRRA γ
I How does Mt,T look like “in practice”?Option Pricing 29/69
Estimated Stochastic Discount Factor from SPX data
2002/10
2004/02
2005/07
2006/11
2008/04
2009/08
0.5
1
1.5
0
1
2
3
4
5
6
ST / S
t
SD
F
Figure: Empirical pricing kernel projected on one-year S&P500 grossreturns ST/St , Mt,T = βu′(cT )/u′(ct) = e−r(T−t)dQt,T/dPt,T , basedon Barone-Adesi, Engle, and Mancini (2008). Note: Mt,T as a functionof ST/St , estimated each day t, changes shape over time.
Option Pricing 30/69
Equilibrium Model: Recap
I Three key quantities
I Objective probability measure Pt,T
I Risk neutral measure Qt,T
I Investors’ preferences βu′(cT )/u′(ct) = e−r(T−t)dQt,T/dPt,T
Note: Any two determine the third
I Objective probability P needs to be risk-adjusted for optionpricing, which is under Q ⇒ Change of measure from P to Q
I Active research on which variables enter the SDF
Mt,T = βu′(cT )/u′(ct)
i.e., on priced risk factors
Option Pricing 31/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Option Pricing 32/69
GARCH Option Pricing
I Asset return dynamic under P
log(St/St−1) = r + λσt −1
2σ2
t + εt
σ2t = β0 + β1σ
2t−1 + β2ε
2t−1
where εt |Ft−1 ∼ N (0, σ2t ), and risk premium λ > 0
I Time varying excess return, as EP[St/St−1|Ft−1] = er+λσt
I How do we change measure from P to Q?
One approach is Locally Risk Neutral Valuation RelationshipI Duan (1995)
There are (infinitely many) other ways of changing measure
Option Pricing 33/69
Locally Risk Neutral Valuation Relationship
I Definition: Risk neutral measure Q satisfies LRNVR when
I Q is absolutely continuous with respect to PI EQ[St/St−1|Ft−1] = er
I VarP[log(St/St−1)|Ft−1] = VarQ[log(St/St−1)|Ft−1]
I log(St/St−1)|Ft−1 ∼ N under Q
I Comments on the LRNVR
X To avoid arbitrage, Q must be absolutely continuous withrespect to P, i.e., for any event A, P(A) = 0⇒ Q(A) = 0
X Under Q, any traded asset (stock, bond, option, etc.) earnsthe risk free rate in expectation (i.e., risk neutral pricing)
Same GARCH variance under P and Q from t − 1 to t
Same class of log-normal return distributions under P and QOption Pricing 34/69
Risk Neutral GARCH Dynamic
I Asset return dynamic under Q, from LRNVR
log(St/St−1) = r − 1
2σ2
t + (εt + λσt)
σ2t = β0 + β1σ
2t−1 + β2ε
2t−1
where εt |Ft−1 ∼ N (−λσt , σ2t ), due to change of measure
I From P to Q, p.d.f. f (εt |Ft−1) “rigidly shifts to left”
I Under Q, (εt + λσt) is the zero-mean return shock
I Change of measure in continuous time model is “similar”,e.g., P-Brownian motion gets some drift from P to Q
I Exercise: If P-dynamic is dSt/St = (r + λ√
vt)dt +√
vtdW Pt ,
what is the Q-dynamic of St?
Option Pricing 35/69
Pricing Options: Two Methods
I Using risk neutral GARCH dynamics, there are two methodsto price options:
I Monte Carlo simulationI Simulate many price paths from t to T , S (i)
T , i = 1, . . . ,mI Compute, say, time-t put price, e−r(T−t)
∑mi=1(K − S
(i)T )+/m
I e.g., Duan (1995), Barone-Adesi, Engle, and Mancini (2008)
I Fourier inversionI Specify (affine) GARCH modelI Derive moment generating function, EQ[eu log(ST/St )|Ft ]I Price options via Fourier inversion
I e.g., Fang and Oosterlee (2008)
I e.g., Heston and Nandi (2000), Majewski et al. (2015)
Option Pricing 36/69
Risk Neutral Parameters
I Calibration (classic approach): Minimize the distance betweenmodel-based and market-based option prices by changingmodel parameters
I Given N market option prices Pmarket(Kj ,Tj ), j = 1, . . . ,NI For a given set of model parameters θ = β0, β1, β2, λ,
compute model-based prices Pmodel (Kj ,Tj ; θ), j = 1, . . . ,NI Change model parameters θ as to
minθ
N∑j=1
(Pmodel (Kj ,Tj ; θ)− Pmarket(Kj ,Tj )
weightj
)2
I Likelihood based method (recent approach): Maximize theweighted sum of log-likelihood of asset returns andlog-likelihood of option pricing errors
I e.g., Badescu, Cui, and Ortega (2016)
Option Pricing 37/69
Implied Volatility Surface from SPX Options
0.40.5
0.60.7
0.80.9
11.1
1.21.3
0
50
100
150
200
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
moneyness = Strike/Futures
Impl
ied
vola
tility
Figure: Implied volatilities σIV (K ,T ) for various strike prices K andmaturities T from SPX options on 31-Dec-2004. Recall, σIV (K ,T ) issuch that PBlack-Scholes(K ,T ;σIV (K ,T )) = Pmarket(K ,T ). Note: Twofactors in volatility help fitting term structure of σIV ; large negative pricejumps help fitting short-term skew of σIV .
Option Pricing 38/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Derivatives 39/69
Volatility Derivatives
I “Volatility has become an asset class”— (Any) investment bank’s brochure, 1990s
I “The next financial crisis will come from volatility markets”— (Anonymous) variance swap trader, last year
Volatility Derivatives 40/69
Volatility Derivatives
I “Volatility has become an asset class”— (Any) investment bank’s brochure, 1990s
I “The next financial crisis will come from volatility markets”— (Anonymous) variance swap trader, last year
Volatility Derivatives 40/69
Volatility Derivatives
I Many volatility derivatives
I Volatility swaps, futures/options on VIX, VIX ETF, etc.
I Variance Swaps
I Basic contract
Volatility Derivatives 41/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Derivatives 42/69
Variance Swap (time 0)
0 50 100 150 200 250−30
−20
−10
0
10
20
30
Time (days)
Pay
off
0 50 100 150 200 250
1000
1050
1100
1150
1200
Time (days)
Sto
ck P
rice
Volatility Derivatives 43/69
Variance Swap (time T > 0)
0 50 100 150 200 250−30
−20
−10
0
10
20
30
Time (days)
Pay
off
0 50 100 150 200 250
1000
1050
1100
1150
1200
Time (days)
Sto
ck P
rice
Volatility Derivatives 44/69
Variance Swaps Term Structure
I Variance Swaps
I actively traded over-the-counter
I for different time horizons ⇒ term structureI previous slide: only one time horizon of 250 days
I Develop term structure models for variance swaps
Volatility Derivatives 45/69
Variance Swap Payoff
Variance Swap payoff = Floating leg − Fix leg
= QVt,t+τ −VSt,t+τ
I Floating leg: Quadratic variation over horizon [t, t + τ ]
I Fix leg: Variance Swap rate, fixed at time t
VSt,t+τ = EQ[QVt,t+τ |Ft ]
depend on time horizon τ ⇒ term structure
I Because variance swap is a forward contractEQ[e−r(T−t)(QVt,t+τ −VSt,t+τ )|Ft ] = 0
Volatility Derivatives 46/69
Variance Swap Term Structure: Data
97 98 99 00 01 02 03 04 05 06 07 08 09 10 236
12
24
10
20
30
40
50
60
70
80
Maturity months
Year
Var
ianc
e S
wap
Rat
e %
Figure: Variance swap rates,√VSt,t+τ × 100, on the S&P500 index
from 4-Jan-1996 to 2-Sep-2010, daily quotes. Source: Bloomberg.
Volatility Derivatives 47/69
Summary Statistics
Variance Swap ratesτ Mean Std Skew Kurt
2 22.14 8.18 1.53 7.083 22.32 7.81 1.32 6.056 22.87 7.40 1.10 4.97
12 23.44 6.88 0.80 3.7724 23.93 6.48 0.57 2.92
Realized Variances2 18.90 12.40 4.31 28.403 19.06 12.04 3.80 21.816 19.46 11.33 2.93 13.17
12 20.13 10.47 1.97 6.8624 20.60 8.81 1.09 3.48
Table: Daily data from 4-Jan-1996 to 2-Sep-2010, volatility percentageunits. Note: Variance risk premium tends to increase with time horizon τ .
Volatility Derivatives 48/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Derivatives 49/69
Modelling Variance Swaps: SV Approach
I From stochastic volatility (SV) literature
I Specify price index dynamics (judiciously)
dSt
St−= rt dt + σt dWt +
∫Rξ (χ(dt, dξ)− νt(dξ)dt)
where σt diffusive volatility, χ jump measure, νt compensator
I Quadratic variation over time horizon [t, t + τ ]
QVt,t+τ =1
τ
(∫ t+τ
tσ2
s ds +
∫ t+τ
t
∫R
(log(1 + ξ))2χ(ds, dξ)
)
I Variance swap rate (possibly in closed-form)
VSt,t+τ = EQ [QVt,t+τ |Ft
]Volatility Derivatives 50/69
Modelling Variance Swaps: TS Approach
I From term structure (TS) literature on bond yields
I Observed term structure: Bond yield curve
y(t,T ) = − 1
T − tlogEQ
[exp
(−∫ T
trs ds
)|Ft
]⇒ Spot interest rate rt = y(t, t), often affine in latent factors
I Observed term structure: Variance swap curve
VS(t,T ) =1
T − tEQ[∫ T
tvs ds|Ft
]⇒ Spot variance vt = VS(t, t), flexible function of latent factors
I Link to SV approach: vt = σ2t +
∫R(log(1 + ξ))2νt(dξ)
Volatility Derivatives 51/69
Quadratic Variance Swap Model
I Spot variance vt quadratic in Xt (subsumes affine models)
vt = φ+ ψ>Xt + X>t πXt
from Filipovic, Gourier, and Mancini (2016)
I State process Xt is m-dimensional quadratic diffusion:
dXt = µ(Xt)dt + Σ(Xt)dWt
where µ(x) linear in x , and Σ(x)Σ(x)′ quadratic in x
I Quadratic term structure model for variance swap
VS(t,T ) = Φ(T − t) + Ψ(T − t)>Xt + X>t Π(T − t)Xt
where Φ, Ψ and Π satisfy a linear system of ODEsVolatility Derivatives 52/69
Model Estimation
I State process Xt is not observable: Require filtering
I Maximum likelihood with (unscented) Kalman filter
I Quadratic model in state-space form:
Xt = Φ0 + ΦX Xt−1 + ωt , ωt ∼ N (0,Qt) (Transition)
Zt = H(Xt) + ut , ut ∼ N (0,Ωt) (Measurement)
I Daily observations Zt = [VSt,T1 , . . . ,VSt,T5 ]′, t = 1, . . . ,T
I Measurement equation nonlinear (quadratic)I Standard Kalman filter not applicable
Volatility Derivatives 53/69
Unscented Kalman Filter (UKF)KF: If measurement equation were linear H(X ) = H0 + H1 X
I Prediction step of Xt and Zt :
Xt|t−1 = Φ0 + ΦX Xt−1, with Pt|t−1 = Cov[Xt − Xt|t−1]
Zt|t−1 = H0 + H1Xt|t−1, with Ft|t−1 = Cov[Zt − Zt|t−1]
I Update step of Xt :
Xt = Xt|t−1 + Wt(Zt − Zt|t−1), with Pt = Cov[Xt − Xt ]
Kalman gain Wt such that ∂trace(Pt)/∂Wt = 0
UKF: Because measurement equation is nonlinear H(X )
Approximate p.d.f. of X with “sigma points” χ(i)t|t−1, ωi
I Zt|t−1 =∑
i ωi H(χ(i)t|t−1), and Ft|t−1, using “sigma points”
I Log-likelihood based on (unscented) Kalman filter
T∑t=1
−1
2
[5 log(2π) + log |Ft|t−1|+ (Zt − Zt|t−1)′F−1
t|t−1(Zt − Zt|t−1)]
Volatility Derivatives 54/69
Filtered Trajectory from Bivariate Quadratic Model
1998 2000 2002 2004 2006 2008 2010−1
0
1
2
3
4
5
6
X1−(λ0+β12X2)
λ1+β11
Figure: Filtered trajectory of bivariate state vector Xt . Quadratic modelfitted to variance swap rates on the S&P500.
I X1t more volatile, mimics trajectories of short-term VS rates
I X2t more persistent, mimics long-term movements in VS rates
Volatility Derivatives 55/69
Fitting Variance Swap Rates
1998 2000 2002 2004 2006 2008 201010
20
30
40
50
60
70
80T = 2 months
Model−based VS ratesActual VS rates
1998 2000 2002 2004 2006 2008 201010
15
20
25
30
35
40
45
50T = 24 months
Model−based VS ratesActual VS rates
Figure: Model-based and actual (i.e., observed) variance swap rates onthe S&P500. Bivariate quadratic model fitted to variance swap rates.
Volatility Derivatives 56/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Derivatives 57/69
Optimal Portfolio Problem
Maximize expected utility from terminal wealth VT of a powerutility investor with constant relative risk aversion (CRRA) η
maxnt ,wt ,φt ,0≤t≤T
EP
[V 1−η
T
1− η
]
By dynamically and optimally investing:
I nt = (n1t , n2t) relative notional exposures to each on-the-runτi -variance swap, i = 1, 2
I wt fraction of wealth invested in stock index
I φt fraction of wealth invested in index option
I and risk-free bond
Solution and implementation via HJB equation (a lot of work)
I Filipovic, Gourier, and Mancini (2016)Volatility Derivatives 58/69
Optimal Investment in VS: Short-Long Strategy
1998 2000 2002 2004 2006 2008 2010−5
−4
−3
−2
−1
0
1
2
n
1t
n2t
I Short position in 2-year VS (earn variance risk premium),long position in 3-month VS (hedge volatility risk)
I Periodic patterns in nt
I Based on bivariate quadratic model
Volatility Derivatives 59/69
Optimal Investment in Stock Index and Put Option
1998 2000 2002 2004 2006 2008 20100.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
wt
1998 2000 2002 2004 2006 2008 2010−5
0
5
10
15
20x 10
−3
φ t
I Positive optimal weight wt in stock index
I Positive, tiny optimal weight φt in put optionVolatility Derivatives 60/69
Wealth Trajectory with Optimal Investment
1998 2000 2002 2004 2006 2008 201050
100
150
200
250
300
Optimal portfolioProxy portfolioS&P500
I Smooth wealth growth with little volatilityI Merely a diversification effect? Not really. . .
I Suited for risk-averse investors (CRRA η = 5)
Volatility Derivatives 61/69
Wealth Trajectory with Optimal Investment: Log-investor
1998 2000 2002 2004 2006 2008 201050
100
150
200
250
300
350
400
Optimal portfolioProxy portfolioS&P500
I Wealth fluctuations larger than S&P500 indexI To seek risk premiums (increased exposures to risk factors)
I Suited for less risk-averse investors (CRRA η = 1)
Volatility Derivatives 62/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
Volatility Derivatives 63/69
VIX Index and Variance Swap RateI VIX: CBOE 30 days (“fear”) volatility index, highly watched
I Assuming no jumps in the index price S , using theQ-dynamic of the forward price1 Fs = EQ[ST |Fs ]
dFs/Fs =√
vs dW Qs
d log Fs = −1
2vs ds +
√vsdW Q
s
1
T − t
∫ T
tvs ds = − 2
T − tlog
FT
Ft+
2
T − t
∫ T
tdFs/Fs
1
T − tEQ[∫ T
tvs ds|Ft
]=
2
T − tEQ[− log
FT
Ft|Ft
]= VIX 2
t
I VIX 2t is (T − t) days VS rate when S does not jump
1And using the identity
− logFT
Ft= 1− FT
Ft+
∫ Ft
0
(K − FT )+
K 2dK +
∫ ∞
Ft
(FT − K)+
K 2dK
Volatility Derivatives 64/69
Difference between VS rates and VIX-type indices
97 98 99 00 01 02 03 04 05 06 07 08 09 10−5
0
5
10
15
Year
VS
min
us V
IX %
2−month3−month6−month
Figure: VSt,t+τ −VIXt,t+τ for τ = 2, 3, 6 months, daily. VIX-typeindices calculated using SPX options and revised VIX methodology fromCBOE. Volatility percentage units. Note: VSt,t+τ −VIXt,t+τ can beinterpreted as profit/loss from selling VS and hedging with SPX optionsand futures, in frictionless markets, when index price S jumps.
Volatility Derivatives 65/69
Wall Street Journal, 22 October 2008
Volatility Derivatives 66/69
Outline
Volatility DynamicsDiscrete-time ModelsContinuous-time Models
Option PricingChange of MeasureGARCH Model
Volatility DerivativesVariance SwapsTerm Structure ModelsOptimal InvestmentVIX
References
References 67/69
ReferencesI Andersen, Fusari, and Todorov (2015), “The Risk Premia Embedded in Index
Options,” Journal of Financial Economics, 117, 558–584.
I Barone-Adesi, Engle, and Mancini (2008), “A GARCH Option Pricing Modelwith Filtered Historical Simulation,” Review of Financial Studies, 21, 1223–1258.
I Badescu, Cui, and Ortega (2016), “Non-Affine GARCH Option Pricing Models,Variance Dependent Kernels, and Diffusion Limits,” working paper, SSRN.
I Bauwens, Laurent, and Rombouts (2006), “Multivariate GARCH Models: ASurvey,” Journal of Applied Econometrics, 21, 79–109.
I Bollerslev (1986), “Generalized Autoregressive Conditional Heteroskedasticity,”Journal of Econometrics, 31, 307–327.
I Bollerslev (1987), “A conditionally Heteroskedastic Time Series Model forSpeculative Prices and Rates of Return,” Review of Economics and Statistics,69, 542–547.
I Christoffersen, Jacobs, Ornthanalai, and Wang (2008), “Option Valuation withLong-Run and Short-Run Volatility Components,” Journal of FinancialEconomics, 90, 272–297.
I Duan (1995), “The GARCH Option Pricing Model,” Mathematical Finance, 5,13–32.
I Duffie, Pan, and Singleton (2000), “Transform Analysis and Asset Pricing forAffine-jump Diffusions,” Econometrica, 68, 1343–1376.
I Engle (1982), “Autoregressive Conditional Heteroscedasticity with Estimates ofthe Variance of United Kingdom Inflation,” Econometrica, 50, 987–1007.References 68/69
References (Continued)
I Engle and Gonzalez-Rivera (1991), “Semiparametric ARCH models,” Journal ofBusiness and Economic Statistics, 9, 345–359.
I Fang and Oosterlee (2008), “A Novel Pricing Method for European OptionsBased on Fourier-Cosine Series Expansions,” SIAM Journal on ScientificComputing, 31, 826–848.
I Filipovic, Gourier, and Mancini (2016), “Quadratic Variance Swap Models,”Journal of Financial Economics, 119, 44–68.
I Glosten, Jagannathan, and Runkle (1993), “On the Relation Between theExpected Value and the Volatility of the Nominal Excess Returns on Stocks,”Journal of Finance, 48, 1779–1801.
I Gruber, Tebaldi, and Trojani (2015), “The Price of the Smile and Variance RiskPremia,” working paper, SSRN.
I Heston (1993), “A closed-form solution for options with stochastic volatility,with applications to bond and currency options,” Review of Financial Studies, 6,327–343.
I Heston and Nandi (2000), “A Closed-form GARCH Option Valuation Model,”Review of Financial Studies, 13, 585–625.
I Majewski, Bormetti, and Corsi (2015), “Smile From The Past: A GeneralOption Pricing Framework with Multiple Volatility and Leverage Components,”Journal of Econometrics, 187, 521–531.
References 69/69