You are on page 1of 15

Quantitative Finance

ISSN: 1469-7688 (Print) 1469-7696 (Online) Journal homepage: https://www.tandfonline.com/loi/rquf20

Assessing the effectiveness of local and global


quadratic hedging under GARCH models

Maciej Augustyniak, Frédéric Godin & Clarence Simard

To cite this article: Maciej Augustyniak, Frédéric Godin & Clarence Simard (2017) Assessing
the effectiveness of local and global quadratic hedging under GARCH models, Quantitative
Finance, 17:9, 1305-1318, DOI: 10.1080/14697688.2017.1279342

To link to this article: https://doi.org/10.1080/14697688.2017.1279342

Published online: 21 Mar 2017.

Submit your article to this journal

Article views: 326

View related articles

View Crossmark data

Citing articles: 7 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=rquf20
Quantitative Finance, 2017
Vol. 17, No. 9, 1305–1318, https://doi.org/10.1080/14697688.2017.1279342

© 2017 iStockphoto LP

Assessing the effectiveness of local and global


quadratic hedging under GARCH models
MACIEJ AUGUSTYNIAK∗ †‡ and FRÉDÉRIC GODIN§¶ and CLARENCE SIMARD
†Département de mathématiques et de statistique, Université de Montréal, PO Box 6128 Station Centre-ville, Montreal,
Quebec, H3C 3J7, Canada
‡Quantact Actuarial and Financial Mathematics Laboratory, Centre de recherches mathématiques, Université de Montréal, PO
Box 6128, Station Centre-ville, Montreal, Quebec, H3C 3J7, Canada
§Department of Mathematics & Statistics, Concordia University, 1455 De Maisonneuve Blvd. W., Montreal, Quebec, H3G
1M8, Canada
¶École d’actuariat, Université Laval, 2425 rue de l’Agriculture, Quebec City, Quebec, G1V 0A6, Canada
Département de mathématiques, Université du Québec à Montréal, PO Box 8888, Station Centre-ville, Montreal, Quebec,
H3C 3P8, Canada

(Received 14 June 2016; accepted 30 December 2016; published online 21 March 2017)

1. Introduction idea is simple to understand in a discrete time setting (see e.g.


Föllmer and Schweizer 1989) since the objective is to find po-
Groundbreaking work in financial theory due in great part to sitions in the risky and riskless assets such that the incremental
Fischer Black, Robert Merton and Myron Scholes showed that cost of rebalancing the hedging portfolio at the next time step
it is theoretically possible to completely hedge away the risk is as small as possible in a quadratic sense. An alternative to
of any derivative position under a certain set of assumptions this local hedging approach is global risk-minimization. Rather
by dynamically trading in risky and riskless assets. This ap- than controlling incremental hedging costs, global hedging
proach to hedging is known as delta hedging and is based on aims to find a trading strategy that minimizes the expected
the paradigm of market completeness and perfect replication. value of a penalty function applied to the terminal hedging
Unfortunately, due to the impossibility of trading in contin- error, which corresponds to the deficit between the payoff of
uous time, to sudden price jumps, to market frictions and to the derivative and the terminal value of the hedging portfolio
the limited availability (or illiquidity) of traded assets, perfect net of all capital injections. This problem was first solved in dis-
replication is generally not feasible in a realistic setting. crete time under a quadratic penalty function†† by Schweizer
Local risk-minimization was proposed by Schweizer (1988) (1995), who also pointed out that the local and global hedging
and Schweizer (1991) as a way to approach the hedging prob- approaches are equivalent under a quadratic criterion when one
lem more effectively in incomplete markets. The underlying
††Global risk-minimization under a quadratic penalty function is also
∗ Corresponding author. Email: augusty@dms.umontreal.ca known as variance-optimal hedging and mean–variance hedging.
© 2017 Informa UK Limited, trading as Taylor & Francis Group
1306 Feature

assumes that discounted asset prices are martingales. However, hedging is impacted by model mis-specification. In this con-
in the more realistic semimartingale setting, global hedging text, we assume that the hedger constructs the hedging portfo-
should in general give a smaller terminal hedging error. lio using a Gaussian GARCH model, but the data-generating
To date, many papers building upon the seminal work of model is either a non-Gaussian GARCH, a non-Gaussian expo-
Schweizer (1988), Schweizer (1991) and Schweizer (1995) nential GARCH (EGARCH) or a regime-switching GARCH
have taken a mathematical perspective and investigated gen- (RS-GARCH) model. Our study concludes with an empirical
eralizations of the underlying theory. For example, Lamberton experiment based on S&P 500 data where we evaluate the
et al. (1998) consider local-risk minimization in the presence effectiveness of GARCH-based hedging, and compare it to
of transaction costs, Motoczyński (2000) studies the existence popular delta hedging benchmarks. In particular, we examine
of an optimal solution to the global risk-minimization prob- hedging effectiveness during and after the stock market crash
lem in a multivariate setting in the presence or absence of of 2008–2009.
transactions costs, Černý and Kallsen (2007) extend the results Our analysis assumes that a short position in a standard
of Schweizer (1995) in a general semimartingale setting, and European call option is hedged with different moneyness levels
Biagini and Cretarola (2007) show how to apply local-risk min- and times to maturity ranging from three months to three years.
imization in the context of defaultable derivatives. However, Although we restrict our attention to call options, our results
there have been relatively few articles examining the empirical also directly apply to put options. This is due to the fact that
and practical relevance of local and global risk-minimization a put option payoff can be expressed as the sum of a call
(e.g. Badescu et al. 2014, Godin 2016, Hocquard et al. 2015, option payoff and an attainable claim, which can be perfectly
Rémillard et al. 2014). replicated with a static hedge. Overall, we find that global
The objective of this article is to investigate and contrast the quadratic hedging can significantly reduce the risk of hedging
effectiveness of local and global quadratic hedging strategies vanilla options with long-term maturities (one year or more).
under GARCH models, both experimentally and empirically. In particular, the performance of this strategy stands out at a
By quadratic hedging strategies, we mean local and global 3-year maturity when compared to local quadratic hedging. Its
risk-minimization strategies based on a quadratic criterion. use should therefore be advocated when hedging Long-Term
Three reasons justify our focus on GARCH models. First, these Equity AnticiPation Securities (LEAPS), which are actively
models have become ubiquitous in the econometrics literature traded call and put options on the Chicago Board Options
due to their strength in explaining volatility dynamics, and Exchange that expire two to three years from their issuance.
have also been shown to perform well as option pricing models Moreover, we find that for global quadratic hedging to achieve
(Christoffersen et al. 2010a). Second, GARCH models can be its full potential at long-term maturities, it must be imple-
estimated and manipulated with greater ease than continuous mented under the physical probability measure. In contrast,
time stochastic volatility jump-diffusion models. Third, it al- we observe that implementing local quadratic hedging with a
lows us to expand on the works of Badescu et al. (2014), Ortega mean-correcting risk-neutral measure instead of the physical
(2012) and Rémillard and Rubenthaler (2013) who were the probability measure does not decrease hedging effectiveness
first authors to address GARCH-based quadratic hedging. at all maturities.
In particular, Rémillard and Rubenthaler (2013) discuss The paper is structured as follows. Section 2 introduces the
global quadratic hedging under the physical probability mea- notation and explains how to implement local and
sure and illustrate the performance of this strategy under global quadratic hedging for a general class of GARCH mod-
simulated data from a GARCH model. On the other hand, els by way of dynamic programming. Section 3 presents the
Badescu et al. (2014) and Ortega (2012) discuss local quadratic results of our quadratic hedging analysis. Section 4
hedging and implement it under a risk-neutral measure using concludes.
Monte Carlo simulations. Badescu et al. (2014) also conduct
an extensive empirical experiment and conclude that GARCH-
based local quadratic hedging outperforms a popular Black–
Scholes delta hedging benchmark that makes use of implied 2. Quadratic hedging
volatilities.
Although these articles discuss local or global quadratic 2.1. Definition of the financial market
hedging in the context of GARCH models, the following prac- We consider a frictionless and arbitrage-free financial market
tical questions remain: (i) what is the value added that global where a risky asset (stock) and a risk-free asset (bond) are
quadratic hedging can provide over local quadratic hedging, traded. We assume that trading occurs in discrete time over
(ii) how does the choice of measure (physical or risk-neutral) a fixed financial horizon T ∈ N at equally-spaced times in
impact the effectiveness of quadratic hedging strategies and the set T := {t | t = 0, 1, . . . , T }. The stock price fol-
(iii) how is the effectiveness of quadratic hedging impacted lows a square-integrable stochastic process {St }t∈T defined
by model risk? The main contribution of our article lies in on a probability space (, FT , P) and adapted to a filtration,
providing some insight into these three questions. F := {Ft }t∈T , where Ft := σ (S0 , . . . , St ) represents an
Our analysis starts off with an experiment based on sim- information set containing observed market prices up to and
ulated data in which we assume that the hedging and data- including time t. The bond price follows a deterministic process
generating models coincide. This allows us to understand the {Bt }t∈T that is assumed to grow at the constant one-period
differences between local and global quadratic hedging in a risk-free rate r from an initial value of one: Bt = exp(r t).
theoretical setting. We then conduct a model risk experiment The measure P denotes the physical (real-world, objective)
to evaluate how the effectiveness of GARCH-based quadratic probability measure.
Feature 1307

2.2. Terminology This deficit is called the terminal hedging error and can be
equivalently expressed as
Definition 1 (Trading strategy) A trading strategy is a pair  
of stochastic processes θ := (θ (B) , θ (S) ) such that θ (B) := T = H − BT V0θ + G θT , (1)
(B) (S)
{θt }t∈T is an adapted process and θ (S) := {θt+1 }t∈T is a which shows that T does not depend on the positions in the
(B)
predictable process. The Ft -measurable random variables θt bond asset for all t ∈ T \ {0}.
(S)
and θt+1 represent, respectively, the number of bond asset Local and global quadratic hedging approaches offer two
shares and the number of shares of stock held over the time different ways to minimize the deviations of T from zero.
period [t, t + 1). On the one hand, global quadratic hedging aims to find a self-
financing hedging strategy θ and an initial capital V0θ that solve
Definition 2 (Value of hedging portfolio) Let Vtθ denote the
value of a hedging portfolio at time t (after rebalancing) based
arg min E 2T . (2)
on the trading strategy θ . For t ∈ T , we have
(V0θ ,θ )
(B) (S)
 
Vtθ := θt Bt + θt+1 St . Note that T = H − VTθ and θt(B) = θt−1
(B) (S) (S)
− θt+1 − θt
(S) (S)
We assume θT +1 = θT , that is, the stock position is not Bt−1 St , t = 1, . . . , T , for a self-financing strategy.
adjusted at time T . On the other hand, local quadratic hedging works with strate-
(B)
gies that are not necessarily self-financing by allowing θt to
Definition 3 (Gains process) Let G θt denote the discounted be freely determined at each time t ∈ T , and imposes VTθ =
hedging gain from time 0 to t, based on the trading strategy θ , T  θ θ

H . In this case, we obtain T = BT t=1 Ct − Ct−1 .
and set G θ0 := 0. For t ∈ T \ {0}, we have Locally risk-minimizing strategies are computed through a
  series of local optimizations having for objective to minimize

t
−1
G θt := θn(S) Bn−1 Sn − Bn−1 Sn−1 . the expected squared incremental hedging costs at each trading
n=1 period. More precisely, starting from t = T, T − 1, . . . , 1,
(B) (S)
the aim is to find the positions (θt−1 , θt ) at time t − 1 that
Definition 4 (Cost process) The discounted hedging cost from  θ θ
2
time 0 to t is defined by Ctθ := Bt−1 Vtθ − G θt , with C0θ = V0θ . minimize E[ Ct − Ct−1 | Ft−1 ], under the constraint that
θ (B) (S)
Vt is constructed from the positions (θt , θt+1 ) that satisfy
Definition 5 (Incremental cost process) The incremental dis- this optimality criterion at time t (for t = T , VTθ = H is im-
θ . Since
counted hedging cost at time t is defined as Ctθ − Ct−1 posed). It can be shown that although locally risk-minimizing
−1 θ (S) strategies are not generally self-financing, they are nevertheless
Ctθ − Ct−1
θ
= Bt−1 Vtθ − Bt−1 Vt−1 − θt
  mean-self-financing, so that the resulting cost process is a P-
−1
× Bt−1 St − Bt−1 St−1 martingale.
   
(B) (B) (S) (S)
= θt − θt−1 + θt+1 − θt Bt−1 St , Remark 1 The statement of the local risk-minimization prob-
lem presented here differs from the original definition given
it corresponds to the total outlay needed at time t (in discounted by Schweizer (1988). However, it is nevertheless equivalent to
terms) to rebalance the hedging portfolio. it as explained by Schweizer (2001, pp. 550–552).
Definition 6 (Self-financing hedging strategy) A trading strat- Remark 2 The trading strategies considered in the local and
egy θ is called self-financing if its cost process is constant, that global quadratic hedging problems must satisfy some square-
is Ctθ = V0θ for all t ∈ T . integrability
 conditions. Itis generally sufficient to assume that
(S)
θt Bt−1 St − Bt−1
−1
St−1 ∈ L2 (P) for t ∈ T \ {0}, and Vtθ ∈
Definition 7 (Mean-self-financing hedging strategy) Atrading
strategy θ is called mean-self-financing if its cost process is a L2 (P) for t ∈ T . See Lamberton et al. (1998) and Schweizer
P-martingale. (1995) for more details.

2.4. Local quadratic hedging algorithm


2.3. Hedging problem
The local risk-minimization problem can be solved with the
We study the problem of an agent who has entered into a following backward recursive scheme for t = T, T − 1, . . . , 1
short position on a European contingent claim with payoff at initiated with VTθ = H (see Lamberton et al. 1998, proposi-
time T given by an FT -measurable square-integrable random tion 2, for more details):†
variable H . The agent’s objective is to establish a trading
strategy such that the value of its portfolio at time T , net (1) Compute
of all capital injections, will be close to H . A natural way (S) Cov e−r Vtθ , t | Ft−1
θt = ,
to assess the attainment of this objective is to compute the Var [t | Ft−1 ]
deficit between the claim’s payoff and the terminal value of
the hedging portfolio net of incremental costs, that is, †Note that our definition of the value of the hedging portfolio at time t
  (see definition 2) corresponds to the one adopted by Lamberton et al.
T
 θ  (1998), and thus differs slightly from the one given in Schweizer
θ θ
T := H − VT − BT Ct − Ct−1 . (1988). However, both definitions lead to equivalent solutions as
t=1 shown by Lamberton et al.(1998, section 4.1).
1308 Feature
(S)
where t := St e−r − St−1 . This value of θt min- νt := E[(1 − t bt /at )νt+1 | Ft−1 ], (14)
imizes e−r E [Ht (1 − t bt /at )νt+1 | Ft−1 ]
  Ht−1 := . (15)
Var Bt−1 Ctθ − Ct−1 θ  Ft−1 νt

(S) 
= Var e−r Vtθ − θt t  Ft−1 . Although global quadratic hedging gives rise to a problem
that may seem more difficult to solve computationally than
(2) Determine Vt−1θ based on the mean-self-financing local risk-minimization, the additional numerical burden is
property of the hedging strategy: in fact relatively small in a discrete time setting (Černý and
 Kallsen 2009).
θ (S) 
Vt−1 = E e−r Vtθ − θt t  Ft−1 ,
Remark 3 When applied under a risk-neutral measure, the
(B) −1 θ (S)
and set θt−1 = Bt−1 (Vt−1 − θt St−1 ). backward recursive schemes in algorithms 1 and 2 are equiva-
lent, in the sense that they yield the same positions in the stock
Note that the terminal condition VTθ = H can always be asset and the same initial value V0θ . This follows from the fact
(B) (S)
satisfied by setting θT = BT−1 (H − θT ST ). that when Bt−1 St is a martingale, we have ∀t, āt = at , b̄t =
The backward recursive scheme presented above can be bt = 0, d̄t = dt , ᾱt = αt and H̄t = Ht . Although solving the
efficiently implemented with the dynamic program set up in underlying optimization problems under a risk-neutral measure
algorithm 1 (see Coleman et al. 2006, for an alternative dy- leads to a simpler solution from a computational standpoint,
namic program). there is no guarantee that the resulting risk-minimizing strategy
Algorithm 1 (Local quadratic hedging dynamic will be close to optimal for minimizing E 2T under P (what
program) Set H̄T = H . For t = T, T −1, . . . , 1, the solution really matters in the end).
to the local quadratic hedging problem is given by:
(S)
θt = ᾱt ,
(B) −1  
θt−1 = Bt−1 H̄t−1 − ᾱt St−1 , 2.6. Implementation under GARCH models
where Many popular GARCH models considered in the econometrics
and option pricing literature can be written as special cases of
āt := E 2t | Ft−1 , (3)
the following GARCH(1, 1) class of models.
b̄t := E [t | Ft−1 ] , (4)
Definition 8 (GARCH(1, 1) class of models) The GARCH
−r
c̄t := e E H̄t | Ft−1 , (5) class models the log-return yt := log(St ) − log(St−1 ), con-
d̄t := e−r E H̄t t | Ft−1 , (6) ditionally on Ft−1 , as
d̄t − b̄t c̄t yt = μt + σt t ,
ᾱt = , (7)
āt − b̄t2
where μt := E[yt | Ft−1 ] represents the conditional mean
H̄t−1 = c̄t − ᾱt b̄t . (8) of yt , σt2 := Var[yt | Ft−1 ] represents the conditional vari-
ance of yt , and t is an innovation independent of Ft−1 with
zero mean and unit variance. The GARCH(1, 1) class that
2.5. Global quadratic hedging algorithm we consider further assumes that this innovation has finite
exponential moments E[exp(σ t )] < ∞, ∀σ > 0, and that
The theoretical solution to the global quadratic hedging prob-
μt = g(σt ) and σt = h( t−1 , σt−1 ), where g(·) and h(·)
lem defined in equation (2) is given by Schweizer (1995).
are parametric functions of their arguments constrained as to
Bertsimas et al. (2001) and Černý (2004) show how to imple-
guarantee second-order stationarity of the process {yt }.†
ment this solution efficiently by way of dynamic programming.
Algorithm 2 summarizes the resulting dynamic program. Note Under this GARCH(1, 1) specification (St−1 , σt ) is a bivari-
that Rémillard and Rubenthaler (2013) used such an approach ate Markov process. This Markov property is key to reduce
in the context of GARCH models. the dimensionality of our dynamic programming algorithms,
because it implies that the processes (3)–(8) in algorithm 1 and
Algorithm 2 (Global quadratic hedging dynamic
the processes (10)–(15) in algorithm 2 can all be expressed as
program) The solution to the global quadratic hedging prob-
functions of only (St−1 , σt ). More precisely, with respect to
lem defined in equation (2) is fully determined by V0θ = H0 and
algorithm 2, there exist functions ãt , b̃t , d̃t , α̃t , ν̃t , H̃t−1 , such
the following backward recursive scheme initiated at t = T ,
that
HT = H and νT +1 = 1:
(S) θ
θt = αt − Vt−1 bt /at , (9) †In the context of quadratic hedging, it is generally assumed that
where St is square-integrable (e.g. Schweizer 1995). Accordingly, the
assumptions on the existence of exponential moments and on second-
at := E 2t νt+1 | Ft−1 , (10) order stationarity are included because they ensure that moments
of St are well-defined. However, we note that Gugushvili (2003)
bt := E [t νt+1 | Ft−1 ] , (11) and Melnikov and Nechaev (1999) were able to weaken the square-
−r integrability requirements commonly encountered in the quadratic
dt := e E [Ht t νt+1 | Ft−1 ] , (12)
hedging literature, and remove the nondegeneracy condition of
αt := dt /at , (13) Schweizer (1995).
Feature 1309

at = ãt (Z t−1 , σt ), bt = b̃t (Z t−1 , σt ), Table 1. Maximum likelihood parameter estimates of the GARCH
model.
dt = d̃t (Z t−1 , σt ), αt = α̃t (Z t−1 , σt ),
νt = ν̃t (Z t−1 , σt ), Ht−1 = H̃t−1 (Z t−1 , σt ), (16) μ ω α γ β

where Z t := log St . Our dynamic program computes these 0.0292 0.0179 0.0545 0.591 0.911
functions backwards in time for t = T, T − 1, . . . , 1 on a (0.0108) (0.0025) (0.0086) (0.103) (0.008)
discretization of the state space of the Markov chain (Z t−1 , σt ). Notes: These parameter estimates were computed on a time series of daily
At every time step, this state space is represented with a two- percentage log-returns on the S&P 500 price index for the period 31 December
dimensional grid built as the Cartesian product of two discrete- 1986–4 January 2010 (5863 observations). Asymptotic standard errors are
given in parentheses.
valued sets:
Gt := {Z (t−1,it ) | i t = 1, . . . , It } × {σ (t, jt ) | jt = 1, . . . , Jt }.
gives an annualized stationary volatility of 17.1% (based on a
The method for selecting the nodes in Gt is analogous to year with 260 business days).
the procedure described in François et al. (2014). Refer to
appendix 1 for more details on the computation of the functions
in (16). 3.1.2. Risk-neutral measure Q. We now introduce a
standard change of measure considered in the GARCH
3. Analysis of quadratic hedging option pricing literature (e.g. Badescu et al. 2014, Christof-
fersen et al. 2010a, Duan 1995, Ortega 2012) that allows us
Our analysis of quadratic hedging under GARCH models starts to define a specific risk-neutral measure, which we denote by
off in section 3.3, where we investigate the improvement in Q. Simply put, this measure transformation shifts the condi-
hedging effectiveness that results from using global rather tional mean, but keeps the conditional variance unchanged.
than local quadratic hedging under simulated data. Then, in Define the conditional mean shift at time t by ηt ∈ Ft−1 ,
section 3.4 we conduct an empirical experiment with S&P and assume that, conditionally on Ft−1 , ˜t := t + ηt is a
500 data to assess how much value can GARCH-based global Gaussian innovation with zero mean and unit variance under
quadratic hedging add over benchmark hedges. These two some measure Q. For measure Q to qualify as a risk-neutral
studies are preceded by two sections that, respectively, describe measure, the expected one-period effective return on the stock
the specific GARCH hedging model that is considered (section price must correspond to the one-period risk-free rate of return
3.1), and present a GARCH delta hedging benchmark (section r , which translates into the following condition (note that yt =
3.2). Our analyses assume that a short position in a European μ − σt ηt + σt ˜t ),
call option is hedged with different moneyness levels and times  
to maturity. More precisely, we consider an initial stock price exp(r ) = exp μ − σt ηt + σt2 /2 ,
of S0 = 100, and the payoff H = max(ST − K , 0), with
or equivalently,
strike price K = 90 (in-the-money, ITM), 100 (at-the-money,
ATM), or 110 (out-the-money, OTM) and maturity of T = 60 μ − r + σt2 /2
ηt = .
(3 months), 260 (1 year), or 780 (3 years) days. σt
This conditional mean shift characterizes the measure transfor-
3.1. GARCH model mation and ensures that the discounted stock price process is
a martingale under Q. The Q-dynamics of the GARCH model
Definition 9 (GARCH model) The model that we will refer to can thus be expressed as
as GARCH in our analysis is a special case of the GARCH(1, 1)
class of models introduced in definition 8. It includes a Gaus- yt = r − σt2 /2 + σt ˜t ,
sian innovation, a constant conditional mean μ (i.e., μt = μ, σt2 = ω + ασt−1
2
(|˜t−1 − ηt−1 | − γ (˜t−1 − ηt−1 ))2 + βσt−1
2
.
∀t), and models the conditional variance as
σt2 = ω + ασt−1
2
(| t−1 | − γ t−1 )
2
+ βσt−1
2
.
This volatility specification was introduced under a slightly 3.2. Delta hedging benchmark
different, but equivalent, parametrization by Glosten et al. Delta hedging is implemented in our analysis to serve as a
(1998) and is commonly known as the GJR-GARCH model. benchmark. The definition of the call option delta given by
Duan (1995) corresponds to‡:
3.1.1. Measure P. The P-dynamics of the GARCH model
were estimated on a time series of daily percentage log-returns provided that the denominator is positive.
on the S&P 500 price index for the period 31 December 1986– ‡Delta hedging under a GARCH model is somewhat of an ambiguous
4 January 2010 (5863 observations). Table 1 reports maximum concept because this model is incomplete and defined in discrete
likelihood parameter estimates and associated asymptotic stan- time. Different approaches to specify a continuous time limit could, in
dard errors. The daily stationary volatility† is 1.061%, which principle, lead to different definitions of delta hedging. Accordingly,
Duan (1995)’s definition of the delta was subject to criticism, notably
by Garcia and Renault (1998). Nevertheless, Badescu et al. (2014)
†The stationary (or unconditional) variance corresponds to show that this delta appears as a component of an approximate locally
ω risk-minimizing strategy by letting the time between two consecutive
E[σt2 ] = ,
1 − α(1 + γ 2 ) − β observations approach zero.
1310 Feature

(S) ∂ −r (T −t) Q hedging error T (see equation (1)); the realization of T on


θt+1 = e E [max(ST − K , 0) | Ft ]
∂ St   path i is denoted by T,i . In particular, we consider the root-

ST  mean-square error (RMSE), defined as
= e−r (T −t) EQ 1{ST ≥K }  Ft ,
St 

where 1 A is the indicator function of set A, K denotes the 1  N
RMSE =  2T,i ,
option’s strike price and the superscript Q on the expectation N
i=1
term indicates that the expectation is computed under the risk-
neutral measure Q. Although Duan (1995) evaluated this delta and the 95% Value at Risk (VaR), which corresponds to the
with Monte Carlo simulations, it is also possible to compute it 501th largest value of {T,i }i=1
10000 . Since global quadratic hedg-
(S)
by setting up a dynamic program as follows. Set θt+1 = ξt , and ing minimizes the RMSE by definition, it will of course be the
iterate backwards starting from t = T − 1 and ξT = 1{ST ≥K } winner with respect to this criterion. Our goal, however, is to
according to equation ξt = e−r EQ [e yt+1 ξt+1 | Ft ].† determine how large the gain actually is and evaluate whether
minimizing this measure also translates into an improvement
with respect to the downside risk of hedging, as measured by
3.3. Experiments with simulated data the 95% VaR. Our analysis of the hedging results presented in
table 3 is structured around the following three questions.
We start off our analysis by investigating the effectiveness of
GARCH-based quadratic hedging in a setting where there is no • By what margin is global quadratic hedging more effective
model risk. We simulated N = 10 000 daily stock price paths than local quadratic hedging?
with the Gaussian GARCH model under measure P and, on We observe that the relative improvement in hedging
each one of these paths, we applied local and global quadratic effectiveness resulting from global quadratic hedging under
hedging approaches under measures P and Q. The simulations P becomes larger when the option’s maturity increases. For
were initiated with a stock price of S0 = 100 and a daily sta- short-term options with a 3-month maturity, global hedging
tionary volatility of σ1 = 1.061%. The bond asset is assumed does not seem to add much value. For medium-term options
to grow at an annualized risk-free rate of 2%, which yields a with a 1-year maturity, the improvement in the RMSE is of
daily risk-free rate assumption of r = 0.00769%. the order of 5%, which is still modest. However, for a 3-year
To allow for a comparison on equal grounds, the hedging maturity, global hedging succeeds at reducing the RMSE by
strategies are all assumed to be initiated with the same capital 10–12%, and does even better with respect to the 95% VaR,
for a given option strike price and maturity. The initial capital which it trims down by 20–31%. This shows that a global
is chosen as the optimal capital associated with GARCH-based approach to quadratic hedging can also considerably reduce
global quadratic hedging under P, that is V0θ = H0 (see equa- the downside risk of hedging at long-term maturities. We note
tion (15)). Table 2 presents the values of this capital at different that Heath et al. (2001a) and Heath et al. (2001b) performed a
moneyness levels and times to maturity, as well as alternative comparison of local and global quadratic hedging in a contin-
choices for V0θ . We observe that the optimal capital obtained uous time stochastic volatility setting and also found that the
with global quadratic hedging under P is almost identical to the relative performance of the global approach improves when
capital computed under Q. Therefore, the choice between these the option’s maturity increases.
two approaches seems irrelevant for determining an initial
capital in our experiment.‡ Moreover, note that our choice of • How does the choice of measure (P or Q) impact the effec-
initial capital does not put any of the hedging strategies at tiveness of quadratic hedging strategies?
a disadvantage because the terminal discounted hedging gain Ortega (2012) claimed that the loss of effectiveness from im-
G θT is impacted by V0θ only in the context of global quadratic plementing local quadratic hedging under measure Q instead
hedging under P. of P should be small in the GARCH setting on a daily time
Hedging effectiveness can be assessed by examining the scale.§ Since he did not implement local quadratic hedging
distribution of the N =10 000 realizations of the terminal under P, he was unable to verify this claim experimentally. Our
results support his claim as we only observe negligible differ-
ences between statistical measures reported for local quadratic
†This equation is obtained as follows: hedging strategies under P and Q. However, in the context of
   global quadratic hedging, this claim is no more true, because
ST 
ξt = e−r (T −t) EQ 1{ST ≥K }  Ft we observe large differences between global quadratic hedging
St
    under P and global/local quadratic hedging under Q for long-
ST 
= e E e t+1 e−r (T −(t+1))
−r Q y 1{ST ≥K }  Ft term maturities. Therefore, the choice of measure (P or Q) does
St+1
     matter for global quadratic hedging.
ST  
= e−r EQ e yt+1 e−r (T −(t+1)) EQ 1{ST ≥K }  Ft+1  Ft
St+1

= e−r EQ e yt+1 ξt+1  Ft .
§Ortega (2012) showed that when the drift term in the GARCH model
is ‘sufficiently small’ the linear Taylor expansions of the locally risk-
‡For a comparison of option prices under different pricing measures, minimizing value process in the drift term under measures P and
including measures related to the local and global quadratic hedging Q coincide. Based on this result, he asserted that ‘carrying out the
problems in continuous time, see Henderson and Hobson (2003), risk-minimizing program with respect to the physical measure or the
Henderson (2005) and Henderson et al. (2005). equivalent martingale measure …yields virtually the same results’.
Feature 1311

Table 2. Initial capital V0θ computed for ATM (K = 100), OTM (K = 110) and ITM (K = 90) call options with T = 60, 260 and 780 days
to maturity.

ATM OTM ITM


Days to maturity 60 260 780 60 260 780 60 260 780

GARCH capital under P 3.44 7.83 15.04 0.36 3.45 10.37 10.98 14.42 20.89
GARCH capital under Q 3.44 7.81 15.01 0.36 3.44 10.34 10.98 14.41 20.87
Black–Scholes capital 3.51 7.79 14.58 0.59 3.85 10.28 10.74 13.95 20.14
Notes: GARCH capital under P: Optimal capital associated with GARCH-based global quadratic hedging under P, V0θ = H0 (see equation (15)). GARCH
capital under Q: Risk-neutral expectation of the discounted payoff under the GARCH model and measure Q, V0θ = e−r T EQ [max(ST − K , 0)]. Black–Scholes
capital: Black–Scholes call price based on an annualized risk-free rate assumption of 2% and an annualized volatility parameter of 17.1%, which corresponds
to the stationary annualized volatility of the GARCH model.

Table 3. RMSE and 95% VaR of the terminal hedging error for ATM (K = 100), OTM (K = 110) and ITM (K = 90) call options with
T = 60, 260 and 780 days to maturity.

ATM OTM ITM


Days to maturity 60 260 780 60 260 780 60 260 780

RMSE
Global under P 0.80 1.36 1.73 0.42 1.26 1.85 0.58 1.21 1.52
Local under P 1% 5% 11% 1% 5% 10% 0% 5% 12%
Global/Local under Q 1% 5% 11% 1% 5% 10% 0% 6% 12%
Delta hedging 17% 30% 41% 12% 31% 44% 15% 27% 37%

95% VaR
Global under P 1.43 2.35 2.71 0.75 2.21 2.92 0.91 1.91 2.16
Local under P 3% 5% 22% 1% 5% 20% 4% 12% 31%
Global/Local under Q 2% 7% 23% 1% 7% 22% 4% 13% 32%
Delta hedging 18% 22% 32% −4% 14% 25% 30% 36% 49%
Notes: These statistical measures are computed based on 10 000 realizations of T which depend on 10 000 trajectories generated from the GARCH model
under measure P. The statistical measures associated with global quadratic hedging are expressed in absolute terms (bold font), whereas those associated with
all the other hedging strategies are expressed relative to these values (% increase).

• How does the effectiveness of quadratic hedging strategies for every time t ∈ T . To assess the performance of each
compare to delta hedging? hedging strategy with respect to this type of criterion, we
computed the maximal hedging error on each simulated path,
We observe that delta hedging induces a much greater vari-
that is maxt∈T t , by assuming that Htmarket = e−r (T −t) EQ
ability in the terminal hedging error when compared to
[max(ST − K , 0) | Ft ].
quadratic hedging. For instance, the RMSE associated with
Moreover, we also calculated the total turnover in the stock
delta hedging is approximately 30% and 40% larger than the
position, defined as
RMSE associated with global quadratic hedging under P for
1 and 3-year maturities, respectively. The global strategy also T  
  (S) (S) 
adds considerable value over delta hedging with respect to (S)
θ := θt+1 − θt  . (17)
the downside risk of hedging. However, the same cannot be t=1
said of local quadratic hedging as we observe that this strategy
results in more moderate improvements, especially at 3-year Between two similarly effective hedging strategies, traders will
maturities and for OTM options. generally prefer the strategy with the lowest θ (S) , because
this strategy implies less adjustments to the stock position and
also implicitly leads to smaller transaction costs.
Results are presented in table 4. For a maturity of three
3.3.1. Pathwise criteria. Until now our assessment of hedg-
years, we observe that local quadratic hedging is slightly more
ing effectiveness was based on the terminal hedging error. Even
effective at avoiding large hedging shortfalls throughout the
though the ultimate goal of hedging is to reduce this error,
term of the contract (by approximately 5%). This result there-
day-to-day behaviour of the hedging strategy also matters to
fore suggests that there is a price to pay inside the hedging
traders in practice (Ahmad and Wilmott 2005). For instance,
trajectory when the strategy is established solely based on
for any time t ∈ T , the shortfall between the market value
a terminal hedging error criterion. However, this price does
of the hedged option, denoted by Htmarket , and the value of
not seem significant enough to discredit the benefits of global
the hedging portfolio net of incremental costs should also be
quadratic hedging shown in table 3. Furthermore, with respect
small. More precisely, traders want to avoid large values of the
to the total turnover in the stock position, we do not observe
hedging error,
any significant differences between local and global quadratic
t := Htmarket − Bt (V0θ + G θt ), hedging. Delta hedging, on the other hand, does leads to less
1312 Feature

Table 4. Mean of the maximal pathwise hedging error and of the total turnover in the stock position.

ATM OTM ITM


Days to maturity 60 260 780 60 260 780 60 260 780

Maximal pathwise hedging error (average over all paths)


Global under P 0.56 1.06 1.47 0.27 1.00 1.60 0.33 0.87 1.24
Local under P 0% −1% −5% −1% −2% −6% 1% −1% −5%
Global/Local under Q 0% −1% −4% 0% −1% −5% 1% 0% −4%
Delta hedging 11% 11% 3% −2% 2% −2% 22% 20% 8%

Total turnover in the stock position (average over all paths)


Global under P 2.40 4.84 7.73 1.18 4.98 8.79 1.06 3.39 6.11
Local under P 0% −1% −2% −1% −2% −2% 0% −1% −2%
Global/Local under Q 0% −1% −2% −2% −2% −2% 1% 0% −2%
Delta hedging −1% −7% −14% 12% 0% −10% −18% −16% −18%
Notes: These results are based on 10 000 trajectories generated from the GARCH model under measure P. ATM (K = 100), OTM (K = 110) and ITM
(K = 90) call options are considered with T = 60, 260 and 780 days to maturity. The maximal pathwise hedging error is defined as maxt∈T t , whereas the
total turnover in the stock position is defined in equation (17). The values associated with global quadratic hedging are expressed in absolute terms (bold font),
whereas those associated with all the other hedging strategies are expressed relative to these values (% increase).

adjustments in the stock position across all moneyness levels (9)), which entails that this strategy has the ability to correct
for a 3-year maturity. itself when model error generates unanticipated changes in the
value of the hedging portfolio.

3.3.2. Robustness to model risk. We now digress from the


idealized setting considered previously, and introduce model
3.3.3. Applicability of results to other payoffs. Although
risk. The rationale is simple; in the real world it is impossi-
our previous analyses were restricted to call option payoffs,
ble to pick the right hedging model. Therefore, it is of high
all of our results also apply to put option payoffs. This is due
practical importance to determine the extent to which the ef-
to put-call parity, which shows that a put option payoff can be
fectiveness of quadratic hedging can be impacted by model
expressed in terms of the payoffs of a call option and a forward
mis-specification. To investigate this impact, we repeat the
contract. As a result, the hedging error t for both options is
experiment of table 3, but we now assume that daily stock price
equal with probability one for all t ∈ T on any given stock
paths are simulated with one of the following three models
price path.
(hedging is still performed with the Gaussian GARCH model
Since our results show that global quadratic hedging adds
under measures P and Q): (1) GARCH model with a general-
value for long-dated vanilla options, such as LEAPS, this strat-
ized error distribution (GED) innovation, (2) EGARCH model
egy should also prove beneficial when hedging vanilla market-
with a GED innovation and (3) RS-GARCH model with two
linked certificates of deposit (CDs), which are structured
regimes.† The GED was considered by Nelson (1991) in the
products offering embedded call option payoffs with a typi-
context of the EGARCH model to allow for an innovation
cal term of three to five years. These types of products can
distribution with thicker tails than the normal. These three
generally be decomposed into a zero-coupon bond and vanilla
models are estimated on the same data set as our Gaussian
options (see Deng et al. 2013, section 4.4, for an overview).
GARCH model. Refer to appendix 2 for more details.
However, the value added of global quadratic hedging is not
Table 5 presents the results of our model risk experiment
necessarily identical for all types of payoffs. For example,
for an ATM call option with T = 60, 260 and 780 days
table 6 repeats the experiment of table 3 for a digital option
to maturity. Since results for OTM and ITM options lead to
(also known as a binary or cash-or-nothing option) with payoff
similar observations, they are not reported. We observe that the
H = 1001{ST ≥K } , where K = 90, 100 and 110. We observe
general conclusions reached in the idealized setting of table 3
that the benefits of global quadratic hedging are much more
continue to hold true, but the dominance of global quadratic
modest in this context. For instance, at a 3-year maturity the
hedging is even more significant in the presence of model risk.
improvement of global versus local quadratic hedging is of 2–
This result suggests that global quadratic is more robust to
4% with respect to the RMSE (compared to 10–12% for a call
model risk than local risk-minimization. This can be explained
option) and of 1–5% with respect to the 95% VaR (compared
on the basis that the stock position in the global strategy is a
to 20–31% for a call option). Moreover, although quadratic
function of the value of the hedging portfolio (see equation
hedging significantly improves over delta hedging with respect
to the RMSE, this advantage does not translate to the 95%
†Note that in this experiment we assume that the stock price process VaR. The difference in results may be due to the discontin-
is generated from a physical probability measure that is distinct from uous nature of the digital option payoff. This type of payoff
P (and Q). The hedging strategy is therefore constructed based on a
mis-specified Gaussian GARCH model. For an alternative way to is difficult to hedge dynamically because small moves in the
assess the robustness of quadratic hedging to model risk, see Di Nunno underlying around the strike price can have very large effects
et al. (2015). on the option’s value. Further research is therefore warranted to
Feature 1313

Table 5. RMSE and 95% VaR of the terminal hedging error for ATM (K = 100) call options with T = 60, 260 and 780 days to maturity.

GARCH-GED EGARCH-GED RS-GARCH


Days to maturity 60 260 780 60 260 780 60 260 780

RMSE
Global under P 0.95 1.57 1.82 0.84 1.28 1.25 0.76 1.28 1.61
Local under P 3% 13% 30% 2% 12% 34% 2% 7% 16%
Global/Local under Q 3% 13% 30% 2% 12% 34% 2% 7% 16%
Delta hedging 15% 33% 58% 17% 37% 76% 17% 32% 52%

95% VaR
Global under P 1.64 2.59 2.33 1.62 2.64 2.21 1.29 1.85 1.76
Local under P 3% 18% 61% 2% 8% 41% 2% 14% 30%
Global/Local under Q 4% 18% 63% 2% 10% 43% 3% 15% 33%
Delta hedging 11% 16% 42% 13% 16% 35% 20% 27% 38%
Notes: These statistical measures are computed based on 10 000 realizations of T which depend on 10 000 trajectories generated from different data-generating
models (see appendix 2 for model definitions and parameters). The statistical measures associated with global quadratic hedging are expressed in absolute
terms (bold font), whereas those associated with all the other hedging strategies are expressed relative to these values (% increase).

Table 6. RMSE and 95% VaR of the terminal hedging error for ATM (K = 100), OTM (K = 110) and ITM (K = 90) digital options with
T = 60, 260 and 780 days to maturity.

ATM OTM ITM


Days to maturity 60 260 780 60 260 780 60 260 780

RMSE
Global under P 15.6 11.3 8.8 12.4 12.7 9.4 9.7 9.2 7.7
Local under P 0% 1% 3% 0% 1% 2% 0% 2% 4%
Global/Local under Q 0% 1% 4% 0% 2% 6% 0% 2% 6%
Delta hedging 4% 9% 18% 6% 8% 18% 7% 14% 19%

95% VaR
Global under P 23.9 15.9 11.0 19.7 19.2 12.4 10.9 11.2 9.2
Local under P −2% −2% 1% 0% 1% 2% −1% 1% 5%
Global/Local under Q 0% 0% 3% 1% 3% 2% 1% 2% 6%
Delta hedging −10% −10% −4% −9% −8% −8% −21% −8% 8%
Notes: These statistical measures are computed based on 10 000 realizations of T which depend on 10 000 trajectories generated from the GARCH model
under measure P. The statistical measures associated with global quadratic hedging are expressed in absolute terms (bold font), whereas those associated with
all the other hedging strategies are expressed relative to these values (% increase).

investigate how the payoff structure impacts the effectiveness paths are not all independent. For simplicity, we continue to
of quadratic hedging. assume a constant daily risk-free rate of r = 0.00769% (2%
annualized). Figure 1 illustrates important characteristics of
the empirical data set that we consider.

3.4. Empirical test


In this section, we move away from evaluating hedging per- 3.4.1. Hedging strategies. On each stock price path, we
formance on simulated data and assess the effectiveness of compute the terminal hedging error for ATM (K = 100), OTM
GARCH-based quadratic hedging on empirically observed re- (K = 110) and ITM (K = 90) call options with the GARCH-
turn paths of the S&P 500 price index. We consider T = 60, 260 based hedging strategies considered in previous sections (re-
and 780-day call option contracts issued on every business day sults for local quadratic hedging under P are not shown because
between 4 January 2008 and 30 March 2012. The stock price they are essentially identical to those for global/local quadratic
path associated with each one of these contracts is assumed to hedging under Q). Note that since we implement these strate-
start at S0 = 100 and then to mimic the returns on the S&P gies with the GARCH model presented in table 1, which was
500 price index during the term of the contract. For example, estimated on S&P 500 data from 31 December 1986–4 January
the 780-day contract issued on 4 January 2008 is exposed 2010, option contracts issued between 1 April 2008 and 31
to S&P 500 return data from 4 February 2008–4 May 2011 March 2010 are not fully out-of-sample in our experiment.
inclusively whereas the 780-day contract issued on 30 March However, we prefer to include contracts issued before the stock
2012 is exposed to S&P 500 return data from 4 February 2012– market crash of 2008–2009 as this allows us to test hedging
8 May 2015 inclusively. In total, we obtain 1 010 stock price effectiveness in an environment where volatility explodes and
paths. It must be noted that since these paths are based on series where the underlying asset is exposed to significant market
of overlapping returns, the hedging errors computed on these swings. Moreover, the parameters of our GARCH model are
1314 Feature

140
120
Stock

100
80
60

2009 2010 2011 2012 2013 2014 2015

10
Returns

−5

−10
2009 2010 2011 2012 2013 2014 2015

80
GARCH

60

40

20

2009 2010 2011 2012 2013 2014 2015

Figure 1. Characteristics of S&P 500 data for the time period 1 April 2008–8 May 2015. Top: Stock price path assuming an index value of
100 on 1 April 2008. Middle: S&P 500 price returns (in %). Bottom: Daily volatility (in %) computed from the GARCH model with parameters
in table 1 (shown on an annualized basis).

Table 7. RMSE and 95% VaR of the terminal hedging error for ATM (K = 100), OTM (K = 110) and ITM (K = 90) call options with
T = 60, 260 and 780 days to maturity.

ATM OTM ITM


Days to maturity 60 260 780 60 260 780 60 260 780

RMSE
Global GARCH under P 1.75 2.13 1.49 1.02 1.68 1.39 1.79 2.23 1.41
Global/Local GARCH under Q 1.57 2.26 2.62 0.92 1.85 2.51 1.61 2.33 2.55
GARCH delta hedging 1.99 3.29 3.11 1.07 2.66 3.06 2.05 3.43 2.77
Local Heston 1.41 2.39 2.70 0.72 1.82 2.47 1.71 2.68 2.85
Black–Scholes delta hedging 1.41 3.41 2.43 1.22 2.90 2.67 2.12 3.48 2.43

95% VaR
Global GARCH under P 4.30 5.34 4.02 1.56 3.83 2.62 4.44 6.13 3.98
Global/Local GARCH under Q 4.03 5.49 6.33 1.58 4.28 4.75 4.04 5.75 6.42
GARCH delta hedging 5.39 8.73 8.32 1.71 6.74 7.03 4.80 9.36 7.08
Local Heston 3.40 5.79 7.26 1.14 4.07 5.13 4.78 6.85 7.92
Black–Scholes delta hedging 3.50 8.28 3.47 1.30 7.12 3.09 5.71 8.32 4.15
Notes: These statistical measures are based on 1 010 values of T computed from empirically observed return paths of the S&P 500 price index starting on
every business day between 1 April 2008 and 30 March 2012. The lowest statistical measure for each moneyness and maturity is indicated in bold font.

little affected when the final two years of the estimation period realized S&P 500 volatility on a rolling window of one year
are removed.† (based on daily returns). The second benchmark consists of the
To put the results of GARCH-based hedging strategies in locally risk-minimizing delta hedge under the stochastic volatil-
perspective, we also compute the terminal hedging error with ity model of Heston (1993), as presented by Poulsen et al.
two benchmarks that are unrelated to GARCH models. The (2009).† The parameters that we consider for this model are
first benchmark is a standard Black–Scholes delta hedge with
a volatility parameter that is dynamically calibrated to the †This delta hedge is computed under the minimal equivalent
martingale measure in Heston’s model, which is known explicitly, and
†The maximum likelihood parameter estimates of the GARCH model thus corresponds to the locally risk-minimizing delta hedge under the
based on the S&P 500 price index for the period 31 December 1986–1 physical probability measure of this model. Although this delta hedge
April 2008 are: μ = 0.0304, ω = 0.0184, α = 0.0560, γ = 0.560 has the advantage of being available in semi-closed-form, it requires
and β = 0.909. These estimates are very close to the ones provided us to infer the unobserved instantaneous variance in Heston’s model.
in table 1. To infer this variance, we use a particle filtering algorithm constructed
Feature 1315

Global GARCH Local GARCH Local Heston GARCH delta Black−Scholes delta 10

5
ATM

−5

2009 2010 2011 2012

Global GARCH Local GARCH Local Heston GARCH delta Black−Scholes delta 10

5
OTM

−5

2009 2010 2011 2012

Global GARCH Local GARCH Local Heston GARCH delta Black−Scholes delta 10

5
ITM

−5

2009 2010 2011 2012

Figure 2. Evolution of the terminal hedging error for S&P 500 call options with T = 780 days to maturity issued on every business day
between 1 April 2008 and 30 March 2012. The horizontal axis specifies the option’s issue date. Top: ATM call option (K = 100), Middle:
OTM call option (K = 110), Bottom: ITM call option (K = 90).

based on those of model SV in Kaeck and Alexander (2013, hedging strategies on equal grounds, we still need to assume a
table 6) (the S&P 500 data period used by these authors to common initial capital. As in previous sections, we select V0θ
estimate this model is the same as the one we consider to fit as the optimal capital associated with GARCH-based global
our GARCH model). quadratic hedging under P. Because this capital is a function of
the GARCH volatility at the time the option contract is issued,
V0θ varies in time in our empirical test.
3.4.2. Initial capital. One should note that the initial capital
used to compute the global quadratic hedging strategy does not
necessarily need to correspond to the actual payment received 3.4.3. Results. Table 7 presents the results of our empiri-
by the option writer. Suppose for example that the market is cal test. We observe that the performance of global quadratic
illiquid, and that call option contracts are sold at a significant hedging under P stands out from all other strategies at 1 and
premium. It would then not be sensible to want that a self- 3-year maturities. In fact, at these maturities, it is always the
financing strategy initiated from the option’s sale price be most effective hedging strategy with respect to the RMSE, and
as close as possible to the call option payoff at maturity. In whenever it is not the winner with respect to the 95% VaR,
fact, we would expect that the terminal value of this strat- it is a close second. Moreover, the improvement in hedging
egy be larger than this payoff on average. This suggests that effectiveness is far from being marginal as it is not uncom-
the global quadratic hedging criterion is most appropriate for mon to observe a 30–50% risk reduction relative to competing
self-financing hedging strategies that are started with a good strategies. However, the dominance of global quadratic hedg-
approximation of the amount of money required to replicate ing fades when it is implemented under the Q measure. For
the option’s payoff. Therefore, a distinction needs to be made instance, at a 3-year maturity statistical measures associated
between the actual cash amount received by the option writer with the strategy under Q are 60–80% higher than those under
and the value of V0θ that is used to compute the global quadratic P. Nevertheless GARCH-based quadratic hedging under Q still
hedging strategy. For this reason, we will not base V0θ on compares reasonably well to Heston’s locally risk-minimizing
option prices that were actually observed during the time pe- delta hedge. These empirical results underscore the importance
riod considered in our empirical test. However, to compare all of making the right choice of measure to take full advantage
of global quadratic hedging at long-term maturities.
similarly to the one proposed by Christoffersen et al. (2010b, section Because the statistical measures in table 7 are computed
3). The particle filtering algorithm is applied on a discretized form of based on hedging error outcomes on overlapping paths, we may
Heston’s model. It aims to approximate the conditional distribution
of the instantaneous variance given the observed data (known as the gain some additional insight by analysing the evolution in time
filtering distribution) sequentially in time with a discrete distribution of the terminal hedging error. Figure 2 illustrates the evolution
whose support points are called particles. of this error for call options with 780 days to maturity. We first
1316 Feature

observe that option contracts issued at the outset of the stock effectiveness were not large enough to discriminate between
market crash of 2008–2009 generally led to higher hedging local and global quadratic hedging at short-term maturities,
losses. This is because 3-year contracts issued between July the performance of the global approach really stood out at
and November 2008 were actually hit by two highly turbulent a 3-year maturity. This can be explained on the basis that
periods: (i) the financial crisis of 2008–2009 at inception of the global strategy can adapt and correct itself according to
the option contract, and (ii) a period of severe stock market the hedging error previously incurred, whereas the local strat-
volatility near maturity associated with the European sovereign egy is inert in this respect. Our results therefore suggest that
debt crisis (August to October 2011). Contracts issued before global quadratic hedging should be advocated when hedging
July 2008 were able to escape the second turbulent period and LEAPS and other long-term derivatives with embedded call (or
thus generally lead to lower hedging errors. After the height of put) option payoffs, such as vanilla market-linked CDs. More-
the financial crisis in early 2009, we observe that the hedging over, we found that to take full advantage of global quadratic
error is trending downwards. The downward kink in the third hedging at long-term maturities, it is imperative to implement
quarter of 2011 is due to the fact that contracts issued after that this strategy under the physical probability measure rather
time avoided the two turbulent periods mentioned previously, than a risk-neutral measure. However, we observed that the
and were thus not exposed to severe volatility. choice of measure is rather inconsequential for local quadratic
Moreover, the two strategies that performed the best at the hedging, and therefore no longer of first-order importance.
onset of the 2008–2009 financial crisis were global quadratic Finally, our results indicated that the effectiveness of global
hedging under P and Black–Scholes delta hedging. However, quadratic hedging is robust to model mis-specification.
Black–Scholes delta hedging was generally inferior to other In future research, it would be interesting to investigate the
strategies during mid-2009 to mid-2010 whereas global performance of quadratic hedging for payoff structures other
quadratic hedging rarely performed worse than its local coun- than calls or puts. For example, some non-vanilla market-
terpart and Heston’s locally risk-minimizing delta hedge. linked CDs include ratchet features or Asian-type payoffs (see
Despite this, Black–Scholes delta hedging still compares rea- Deng et al., 2013). Additionally, it would be worthwhile to
sonably well with other strategies overall for ATM and OTM compare the effectiveness of GARCH-based quadratic hedging
options. Furthermore, note that GARCH delta hedging does to global risk-minimization strategies constructed
surprisingly well post-crisis and rarely generates a hedging based on non-quadratic criteria, such as VaR. For instance,
loss. This can be explained on the basis that this hedge over- François et al. (2014) developed a flexible methodology for
shoots the position in the stock asset when compared to risk- implementing global risk-minimization under any desired
minimizing hedges.† Since a larger stock position inevitably penalty function within a general regime-switching framework.
leads to additional gains in a rising market, the better per- Given the ability of regime-switching models to account for
formance of the GARCH delta hedge after 2009 was to be time-variation in parameters, it would also be interesting to
expected in some sense. Of course, there is no free lunch here as evaluate whether the effectiveness of global risk-minimization
this good performance is overshadowed by a poor performance can further be improved based on a RS-GARCH modelling
associated with contracts issued at the onset of the stock market framework.
crash.
Disclosure statement
4. Conclusion
No potential conflict of interest was reported by the authors.
This article studied the effectiveness of local and global
quadratic hedging strategies under GARCH models, both ex- Funding
perimentally and empirically. Our analysis addressed three im-
portant practical issues: (i) the value added of global over local Maciej Augustyniak gratefully acknowledges financial support from
quadratic hedging, (ii) the importance of the choice of mea- the Natural Sciences and Engineering Research Council of Canada
[RGPIN-2015-05066]. Frédéric Godin gratefully acknowledges fi-
sure (physical or risk-neutral) when implementing quadratic nancial support from the Fonds de recherche du Québec – Nature et
hedging and (iii) the robustness of quadratic hedging to model technologies [2017-NC-197517].
mis-specification.
Overall, we found that global quadratic hedging can signifi- References
cantly reduce the risk of hedging vanilla options with long-term
Ahmad, R. and Wilmott, P., Which free lunch would you like today,
maturities (one year or more). Although differences in hedging Sir? Delta hedging, volatility arbitrage and optimal portfolios.
Wilmott Mag. 2005, (November), 64–79.
†Unlike risk-minimizing delta hedges, the GARCH delta hedge Badescu,A., Elliott, R.J. and Ortega, J.-P., Quadratic hedging schemes
ignores that the stock price and its volatility could be related for non-Gaussian GARCH models. J. Econ. Dyn. Control, 2014,
through a leverage effect. This well-documented stylized fact of 42, 13–32.
stock market returns entails that negative returns are generally Bertsimas, D., Kogan, L. and Lo, A.W., Hedging derivative securities
followed by an increase in volatility. Therefore, when returns are and incomplete markets:An -arbitrage approach. Oper. Res., 2001,
negative, the call option directly loses value through the stock price 49(3), 372–397.
decrease and indirectly gains value through an expected higher future Biagini, F. and Cretarola, A., Quadratic hedging methods for
volatility, resulting in a natural partial hedge. This feature is not taken defaultable claims. Appl. Math. Optim., 2007, 56(3), 425–443.
into account by the GARCH delta hedge, which explains why it Černý, A., Dynamic programming and mean-variance hedging in
overhedges the call option with a larger stock position than is needed. discrete time. Appl. Math. Finance, 2004, 11(1), 1–25.
See also Badescu et al. (2014, p. 21) for a mathematical justification Černý, A. and Kallsen, J., On the structure of general mean-variance
of this observation. hedging strategies. Ann. Prob., 2007, 35(4), 1479–1531.
Feature 1317

Černý, A. and Kallsen, J., Hedging by sequential regressions revisited. Motoczyński, M., Multidimensional variance-optimal hedging in
Math. Finance, 2009, 19(4), 591–617. discrete-time model–a general approach. Math. Finance, 2000,
Christoffersen, P., Dorion, C., Jacobs, K. and Wang, Y., Volatility 10(2), 243–257.
components, affine restrictions, and nonnormal innovations. J. Bus. Nelson, D.B., Conditional heteroskedasticity in asset returns: A new
Econ. Stat., 2010a, 28(4), 483–502. approach. Econometrica, 1991, 59(2), 347–370.
Christoffersen, P., Jacobs, K. and Mimouni, K., Volatility dynamics Ortega, J.-P., GARCH options via local risk minimization. Quant.
for the S&P500: Evidence from realized volatility, daily returns, Finance, 2012, 12(7), 1095–1110.
and option prices. Rev. Financ. Stud., 2010b, 23(8), 3141–3189. Poulsen, R., Schenk-Hoppé, K.R. and Ewald, C.-O., Risk
Coleman, T.F., Li, Y. and Patron, M.-C., Hedging guarantees in minimization in stochastic volatility models: Model risk and
variable annuities under both equity and interest rate risks. empirical performance. Quant. Finance, 2009, 9(6), 693–704.
Insurance: Math. Econ., 2006, 38(2), 215–228. Rémillard, B., Hocquard, A., Lamarre, H. and Papageorgiou, N.A.,
Deng, G., Dulaney, T., Husson, T., McCann, C. Structured certificates Option pricing and hedging for discrete time regime-switching
of deposit: Introduction and valuation. Working Paper, 2013. models. Working Paper, 2014. Available online at: SSRN (ID:
Available online at: SSRN (ID: 2303718). 1591146).
Di Nunno, G., Khedher, A. and Vanmaele, M., Robustness of quadratic Rémillard, B. and Rubenthaler, S., Optimal hedging in discrete time.
hedging strategies in finance via backward stochastic differential Quant. Finance, 2013, 13(6), 819–825.
equations with jumps. Appl. Math. Optim., 2015, 72(3), 353–389. Schweizer, M., Hedging of options in a general semimartingale model.
Duan, J.-C., The GARCH option pricing model. Math. Finance, 1995, PhD Thesis, ETH Zürich, Zürich, 1988.
5(1), 13–32. Schweizer, M., Option hedging for semimartingales. Stoch. Proc.
Föllmer, H. and Schweizer, M., Hedging by sequential regression: Appl., 1991, 37(2), 339–363.
An introduction to the mathematics of option trading. Astin Bull., Schweizer, M., Variance-optimal hedging in discrete time. Math.
1989, 18(2), 147–160. Oper. Res., 1995, 20(1), 1–32.
François, P., Gauthier, G. and Godin, F., Optimal hedging when the Schweizer, M., A guided tour through quadratic hedging approaches.
underlying asset follows a regime-switching Markov process. Eur. In: Option Pricing, Interest Rates and Risk Management,
J. Oper. Res., 2014, 237(1), 312–322. Handbooks in Mathematical Finance, edited by E. Jouini,
Garcia, R. and Renault, É., A note on hedging in ARCH and stochastic J. Cvitanic and M. Musiela, pp. 538–574, 2001 (Cambridge
volatility option pricing models. Math. Finance, 1998, 8(2), 153– University Press: Cambridge).
161.
Glosten, L.R., Jagannathan, R. and Runkle, D.E., On the relation
between the expected value and the volatility of the nominal excess Appendix 1. Sketch of the dynamic programming
return on stocks. J. Finance, 48(5), 1779–1801. algorithm
Godin, F., Minimizing CVaR in global dynamic hedging with
transaction costs. Quant. Finance, 2016, 16(3), 461–475. This section explains how to compute ãt . The computations for the
Gugushvili, S., Dynamic programming and mean-variance hedging other functions in (16) are analogous. First, note that under our
in discrete time. Georgian Math. J., 2003, 10(2), 237–246. GARCH(1,1) specification in definition 8, Z t = Z t−1 + g(σt ) + σt t
Haas, M., Skew-normal mixture and Markov-switching GARCH and σt+1 = h( t , σt ), which allows us to write
processes. Stud. Nonlinear Dyn. Econ., 2010, 14(4), Article 1, 1–
54. ãt (Z t−1 , σt ) = E 2t νt+1 | Ft−1
Haas, M., Mittnik, S. and Paolella, M.S., A new approach to Markov-
switching GARCH models. J. Financ. Economet., 2004, 2(4), 493– = E 2t ν̃t+1 (Z t , σt+1 ) | (Z t−1 , σt )
530.  
= E t t , Z t−1 , σt | (Z t−1 , σt ) ,
Heath, D., Platen, E. and Schweizer, M., A comparison of two
quadratic approaches to hedging in incomplete markets. Math. where
Finance, 2001a, 11(4), 385–413.
Heath, D., Platen, E. and Schweizer, M., Numerical comparison t (ε, z, σ ) := (e z+g(σ )+σ ε−r − e z )2 ν̃t+1 (z +g(σ )+σ ε, h(ε, σ )).
of local risk-minimisation and mean-variance hedging. Option Then, since t is independent from Ft−1 and thus also from
Pricing, Interest Rates and Risk Management, Handbooks in (Z t−1 , σt ), we obtain
Mathematical Finance, edited by E. Jouini, J. Cvitanic and
M. Musiela, pp. 509–537, 2001b (Cambridge University Press: ãt (Z (t−1,i t ) , σ (t, jt ) )
Cambridge).  
Henderson, V., Analytical comparisons of option prices in stochastic = E t t , Z (t−1,i t ) , σ (t, jt )
volatility models. Math. Finance, 2005, 15(1), 49–59.   
Henderson, V. and Hobson, D., Coupling and option price = t ε, Z (t−1,i t ) , σ (t, jt ) f t (ε)dε
comparisons in a jump-diffusion model. Stoch. Stoch. Rep., 2003, R
K  mk

75(3), 79–101.
Henderson, V., Hobson, D., Howison, S. and Kluge, T., A comparison = t (ε, Z (t−1,i t ) , σ (t, jt ) ) f t (ε)dε, (A1)
of option prices under different pricing measures in a stochastic k=1 m k−1
volatility model with correlation. Rev. Deriv. Res., 2005, 8(1), 5– where −∞ = m 0 < m 1 < . . . < m K −1 < m K = ∞ forms a
25. partition of the real line and f t (·) is the probability density function
Heston, S.L., A closed-form solution for options with stochastic of the innovation t . Provided that (m 1 , m K −1 ) covers the relevant
volatility with applications to bond and currency options. Rev. support of f t (·) and that |m k − m k−1 |, k = 2, . . . , K − 1, are
Financ. Stud., 1993, 6(2), 327–343. sufficiently small, we can approximate the integral in expression (A1)
Hocquard, A., Papageorgiou, N. and Remillard, B., The payoff to get our recursion step
distribution model: An application to dynamic portfolio insurance.
Quant. Finance, 2015, 15(2), 299–312. 
K  
Kaeck, A. and Alexander, C., Stochastic volatility jump-diffusions ãt (Z (t−1,i t ) , σ (t, jt ) ) ≈ t ε(k) , Z (t−1,i t ) , σ (t, jt )
for European equity index dynamics.. Eur. Financ. Manage, 2013, k=1
 
19(3), 470–496. × F t (m k ) − F t (m k−1 ) ,
Lamberton, D., Pham, H. and Schweizer, M., Local risk-minimization
under transaction costs. Math. Oper. Res., 1998, 23(3), 585–612. where F t is the cumulative distribution function of t and ε(k) ∈
Melnikov, A.V. and Nechaev, M.L., On the mean-variance hedging [m k , m k−1 ] for all k. Our implementation of this dynamic program is
problem. Theory Prob. Appl., 1999, 43(4), 588–603. based on choices of m k and ε(k) that are analogous to those considered
1318 Feature

by François et al. (2014); they are chosen as quantiles of the distri- When the innovation is assumed GED, E | t−1 | is given by equa-
bution of t . To improve accuracy of the approximation, a relatively tion (B2). The estimated parameters of the EGARCH-GED model are
larger number of quantiles are sampled in the tails of the distribution. given in table B2.
Finally, note that to compute t on the grid Gt , we must be able to
evaluate ν̃t+1 . Since the values of ν̃t+1 are only known on the grid Table B2. Maximum likelihood parameter estimates of the
Gt+1 , we used cubic spline interpolation to approximate this function EGARCH-GED model.
at points that do not lie directly on the grid nodes of Gt+1 .
μ ω α γ β υ
Appendix 2. Description of models used in table 5
0.0445 −0.0039 −0.090 0.116 0.987 1.34
B.1. GED
B.4. RS-GARCH
If X is a GED random variable with zero mean and unit variance, then
its density is given by (see Nelson 1991, equation (2.4)) The RS-GARCH model that we consider is based on the specifica-
    tion introduced by Haas et al. (2004) and extended by Haas (2010).
υ exp − 12 |x/λ|υ Let {X t } be a discrete time Markov chain with finite state space
f X (x) = , −∞ < x < ∞, 0 < υ ≤ ∞, {1, . . . , M} and transition matrix ( pi j )i,Mj=1 , where pi j = Pr(X t =
λ2(1+1/υ) (1/υ)
j | X t−1 = i), for i, j = 1, . . . , M. Assume further that this Markov
where (·) is the gamma function, and chain is hidden from market participants (i.e., unobserved) and that
1/2 the log-return yt depends on this Markov chain in the following way:
λ := 2(−2/υ) (1/υ)/ (3/υ) .
yt = μ X t + σt,X t t ,
The parameter υ controls the tail-thickness of the distribution.
When υ = 2, the GED collapses to the standard normal distribution. where t is a Gaussian innovation with zero mean and unit variance,
For υ < 2, the GED has thicker tails than the normal. The first absolute μ j , j = 1, . . . , M, is a regime-dependent conditional mean parame-
moment of X is given by (see Nelson 1991, equation (A1.8)) ter, and σt,2 j represents the conditional variance process in regime j,
j = 1, . . . , M, which evolves according to the GARCH structure,
E [|X |] = λ21/υ (2/υ)/ (1/υ). (B2)
σt,2 j = ω j + α j (|yt−1 − μ j | − γ j (yt−1 − μ j ))2 + β j σt−1,
2
j,
B.2. GARCH-GED where {ω j , α j , γ j , β j } are parameters of the conditional variance
The GARCH-GED model is the GARCH model in definition 9 with process in regime j, j = 1, . . . , M. The specific RS-GARCH model
a GED innovation. Its estimated parameters are given in table B1. that we estimated assumes M = 2 and the following parameters (see
table B3):    
p11 p12 0.930 0.070
Table B1. Maximum likelihood parameter estimates of the GARCH- =
p21 p22 0.977 0.023
GED model.

μ ω α γ β υ
Table B3. Maximum likelihood parameter estimates of the RS-
0.0483 0.0123 0.0499 0.556 0.922 1.33 GARCH model (M = 2).

j μj ωj αj γj βj
B.3. EGARCH-GED
1 0.066 0.006 0.039 0.536 0.933
The EGARCH model was proposed by Nelson (1991) as an alternative 2 −0.475 0.830 0.440 0.998 0.417
GARCH specification that ensures positivity of the conditional vari-
ance without parametric constraints. The idea consists in modelling
the logarithm of the conditional variance as
 
log(σt2 ) = ω + α t−1 + γ | t−1 | − E | t−1 | + β log(σt−12 ).

You might also like