Ekonomi och samhälle

Economics and Society

Skrifter utgivna vid Svenska handelshögskolan

Publications of the Hanken School of Economics

Nr 220

Ihsan Ullah Badshah

Modeling and Forecasting Implied Volatility

Implications for Trading, Pricing, and Risk Management

<
Helsinki 2010
Modeling and Forecasting Implied Volatility: Implications for Trading, Pricing, and
Risk Management

Key words: Asymmetric Volatility, Modeling Implied Volatility, Options, Principal

Component Analysis, Quantile Regression, Spillovers, Value-at-Risk, Volatility
Index, Volatility Smirk/Skew, Volatility Surface

© Hanken School of Economics & Ihsan Ullah Badshah

Ihsan Ullah Badshah

Hanken School of Economics
Department of Finance and Statistics
P.O.Box 287, 65101 Vaasa, Finland

Distributor:

Library
Hanken School of Economics
P.O.Box 479
00101 Helsinki, Finland

Telephone: +358-40-3521 376, +358-40-3521 265

Fax: +358-40-3521 425
E-mail: publ@hanken.fi
http://www.hanken.fi

ISBN 978-952-232-106-0 (printed)

ISBN 978-952-232-107-7 (PDF)
ISSN 0424-7256

Edita Prima Ltd, Helsinki 2010

 i

Dedicated to my parents
ii
 iii

ACKNOWLEDGMENTS

This PhD thesis would not have been possible without the help, guidance, and support
of many individuals, all of whom deserve special thanks and appreciation.

First, I must express my deepest gratitude to my thesis supervisor—Johan Knif—for

providing me the opportunity to explore the area of volatility modeling and risk
management. He has been kind and helpful and has always provided valuable
suggestions and feedback on my thesis. He has extended his support to me on various
occasions. For example, his encouragement led me to teach advanced (MSc) level
courses at the department of finance, Hanken School of Economics, an experience that
opened my eyes to the world of teaching and bolstered my confidence. Johan, I am
grateful and indebted to you for all your help and support during my PhD.

I am also very thankful to Kenneth Hogholm, who has always been obliging, provided
timely help in various matters, including administrative matters, and offered useful
comments on my research.

I would also like to thank my thesis reviewers—Mika Vaihekoski and George

Skiadopoulos—for their thorough and valuable comments on my work. These
significantly improved the thesis, and I greatly appreciate and acknowledge all the
effort that went into them.

Eva Liljeblom deserves special mention for funding my PhD studies and trips abroad.
She has also been kind enough to comment on my articles and has always made
available her support.

I will always be indebted to my uncle Dr. Rahman Wali, who inspired and motivated
me to pursue a PhD. Thank you, uncle, for your support and guidance all these years.

My friends and colleagues have always been there for me, and I am grateful to them all.
Sheraz Ahmed and Kashif Saleem, both currently teaching at the LUT School of
Business, Finland, have supported and encouraged me throughout this journey and
even before. Muhammad Ali, Helsinki University Central Hospital, has always given me
invaluable advices. Sohail Farooq, currently studying in the University of Vaasa, has
been a good friend and discussion partner.

I owe special thank to Alireza Tourani-Rad, who has provided valuable suggestions, and
comments on my thesis articles and has always been supportive, so I will always be
indebted to you for all these. I would also like to thank Bart Frijns, Hossein Asgharian,
Seppo Pynnonen, Gregory Koutmos, David Simon, Jussi Nikkinen, Yakup Arisoy, and
Dudley Gilder for providing constructive comments on my thesis articles at various
international and local conferences. I must mention the director of the Graduate School
of Finance (GSF), Mikko Leppamaki, for his efforts in organizing research workshops
for the doctoral students to present their work at.

I would like to thank my colleagues at the department of finance: Benny Jern, Mujahid
Hussain, Anand Gulati, Hilal Butt, Kari Harju, Christian Johansson, Peter Nyberg,
Nadir Virk, Saint Kuttu, Nasib Nabulsi, and David Gonzalez. Their companionship and
constructive feedback at the internal presentations have been a great help. I also wish
to thank Khalid Bhatti, Mats Engsbo, and Chouki Sfandla for their time and the
stimulating discussions.
 iv

Finally, a very special thanks to my family—my parents, sisters, and brothers—to whom
I owe so much. You stood by me and gave me the confidence to take on such a
demanding and challenging project. Most important of all, I thank my dearest wife
Riffat. Your patience, support, and love never fail to amaze me. You give me the
strength to carry on, and it is your encouragement that helped me through the final
thesis submission.

The financial support from the Evald & Hilda Nissi Foundation, Center of Financial
Research (CEFIR), Hanken Foundation, Ella and Georg Ehrnrooth Foundation,
NASDAQ-OMX Foundation, and Marcus Wallenberg Foundation is greatly
acknowledged.

September 27, 2010

Ihsan Ullah Badshah

 v

CONTENTS

1 INTRODUCTION……………………………………………………………………….........................1

2 OPTION MARKETS’ EMPIRICAL REGULARITIES AND MODELING…………........8

3 METHODOLOGY……………………………………………………………………..........................15

4 SUMMARIES OF THE ESSAYS…………………………………………………….....................28

5 REFERENCES……………………………………………………………………….......................….34

THE ESSAYS

1. Badshah, I., (2010). “Modeling the Dynamics of Implied Volatility Surfaces”.

Manuscript, Hanken School of Economics, Finland........................................................41

2. Badshah, I., (2010). “Quantile Regression Analysis of Asymmetric Return-Volatility

Relation”. Manuscript, Hanken School of Economics, Finland......................................73

3. Badshah, I., (2010). “The Information Content of VDAX Volatility Index and
Backtesting Daily Value-at-Risk Models”. Manuscript, Hanken School of Economics,
Finland............................................................................................................................107

4. Badshah, I., (2010). “Modeling Risk Factors Driving the EUR, USD, and GBP
Swaption Volatilities”. Manuscript, Hanken School of Economics, Finland.................135
vi
 vii

PART I: BACKGROUND, METHODOLOGY AND FINDINGS

viii
 1

1 Introduction

Modeling and forecasting of volatility are important to both financial practitioners and
academics, especially in risk management activities, option pricing, option trading,
hedging of derivative positions, constructing volatility indices, portfolio construction
and diversification, and policy making, all of which require the accurate modeling and
forecasting of volatility. We have seen tremendous development in volatility forecasting
models since the introduction of the autoregressive conditional heteroskedasticity
(ARCH) model proposed by Engle (1982). Since then, many ARCH models have been
developed that attempt to forecast volatility using historical information. In 1986,
Bollerslev (1986) extended the ARCH model to a generalized ARCH model called
GARCH. Many extensions have also been made to the GARCH model (see Poon and
Granger, 2003, 2005); for example, Glosten, Jagannathan and Runkle (hereafter GJR)
(1993) have proposed the GJR model, which accounts for asymmetric volatility.

However, a growing body of literature supports the use of implied volatility (hereafter
IV) instead, calling IV the best forecast of future realized volatility (hereafter RV).
These studies empirically document that the information content of IV is richer than
and superior to that of historical volatility (hereafter HV) when forecasting the future
RV of the underlying asset (e.g., Day and Lewis, 1992; Christensen and Prabhala, 1998;
Fleming, 1998; Dumas, Fleming and Whaley, 1998; Blair et al., 2001; Ederington and
Guan, 2002; Poon and Granger, 2003; Mayhew and Stivers, 2003; Martens and Zein,
2004).1

Moreover, these previous studies assert that HV is backward-looking and develop

expectations about future volatility based on the past behavior of stock prices and other
relevant information. In contrast, IV is forward-looking, i.e., implied by the market
prices of options. Option prices reflect market participants’ consensus view (aggregated
beliefs updated through option trades) about the expected future volatility of an
underlying asset over the remaining life of an option.2 These volatility expectations can

1 Modeling of implied volatility is a less trite area of research in quantitative finance even today; however, it

has received far less attention from academic circles in comparison to volatility modeling using historical
information; for instance, there is an abundance of research articles on GARCH-type models.
2 Through option prices, investors update their beliefs about the underlying processes such as dividend

growth, macroeconomic variables, etc., thereby altering IVs and their shape [i.e., inducing smirk (skew), or
surface]. However, their dynamics are driven by latent risk factors (see, for economic explanations, studies
such as David and Veronesi, 2000; Guidolin and Timmermann, 2003; Garcia et al., 2003; and
Chalamandaris and Tsekrekos, 2009).
 2

be backed out by inverting option-pricing models such as Black and Scholes’ (hereafter
BS) (1973) model for equity options and Black’s model (1976) for interest rate options
(examples of these are swaptions and caps).

Since the 1987 market crash, one of the strongest empirical regularities in the U.S.
stock index option has appeared: IV recovered from stock index options appears to
differ between moneyness (the strikes dimension), called volatility smirk (skew), and
the term structure (the time-to-maturity dimension), called the volatility term
structure, if the two are examined at the same time. This implies that IVs present a rich
implied volatility surface (hereafter IVS) whose dynamics across strikes and over time
warrant further investigation. In fact, the assumptions behind the BS model do not
hold empirically, as asset prices are mostly influenced by risk factors such as stochastic
volatility or time-varying volatility, jumps, net-buying pressure of options, and
transaction costs (see, e.g., Carr et al., 2001; Heston, 1993; Bates, 1991; Leland, 1985;
David and Veronesi, 2000; Garcia et al., 2003; Bollen and Whaley, 2004).3 To account
for these deviations, practitioners employ different IVs for different strikes and
maturities. Consequently, the IV backed out using the BS model with the help of the
bisection method may reflect determinants of the option price that are not captured by
the BS model. Here, by repeating this procedure, we obtain IVs for every strike and
maturity; i.e., the volatility structure shows discrepancies between theoretical prices
and market prices. Therefore, an IVS that incorporates these features is essential in
practice, particularly for forecasting future volatility, pricing illiquid options, option
trading, pricing and hedging exotic derivatives, and risk management of options
portfolios.4

Additionally, it is of paramount importance for the practitioners to generate an IVS

from the IVs, which are implied by options representing the majority of market
sentiments (beliefs about bearish and bullish markets). Examples of this are accounting
for the put options in a volatility measure, particularly out-of-the-money (hereafter
OTM) put options that embed beliefs about market crashes, and, on the other hand,
OTM call options on market run-ups (see, e.g., Pan, 2002; Liu et al., 2005; Doran et al.,
2007; Bates, 2008; Camara and Heston, 2008). Because investors (particularly

3 Most of the BS model’s assumptions fail empirically; for instance, the BS model assumes that underlying

assets follow a geometric Brownian motion and constant volatility, implying that options with different
strikes and maturities on the same underlying asset should have the same IV.
4 See Skiadopoulos (2001) for the literature review on the implied volatility smiles and smile consistent

models.
 3

institutional investors) mostly hedge their portfolios’ downside risk with put options
and in market turmoil drive a greater buying demand for OTM puts than OTM calls,
higher IVs are generated (see, e.g., Bollen and Whaley, 2004). Therefore, it is essential
that the IVS be generated using the IVs implied by both put and call options (the most
informed options), thus leading the IVS to move with changes in options prices.5
Furthermore, Liu et al. (2005) also support the use of OTM put and call options and
thus argue that the rare-event premia play an important role in generating the volatility
smirk (skew) pattern observed for options across moneyness and that these rare-events
are embedded in the OTM options.6 Camara and Heston (2008) have proposed an
option model that accounts for both OTM put and call options. They derive the extreme
negative events from OTM puts and extreme positive events from OTM calls. The IV
measure that accounts for OTM options thus contains a broader set of information and
thereby should be more robust; as a result, the IV measure should be the perfect
candidate for generating IVS.

Furthermore, IVS can be viewed as highly correlated, as IVs present a high degree of
correlation among the underlying risk factors, thereby indicating a high dependence
therein. When few important sources of information are common to the risk factors, we
observe a high degree of correlation among the risk factors. By using principal
component analysis (hereafter PCA), most important risk factors can be extracted from
the correlated IVS, which can explain most of the dynamic therein (see, e.g.,
Skiadodopulos et al., 1999; Alexander, 2001; Cont and Da Fonseca, 2002). PCA helps in
reducing dimensionality and enhancing computational efficiency. The extracted
independent risk factors can be used for a variety of purposes, such as vega-hedging,
projecting IVS, value-at-risk (hereafter VaR) forecasting, and model calibration.
Later, Britten-Jones and Neuberger (2000) and Jiang and Tian (2005) derived a
model-free implied volatility (hereafter MFIV) under the pure diffusion assumption
and asset price processes with jumps.7 This MFIV measure has now been adapted by

5 For example, a negative or positive shock to the market induces adjustments in hedging and trading
strategies, consequently triggering changes in the prices of one type (i.e., put or call) of option. The IV then
changes in the direction of the market demand of a particular type of option and the underlying asset (see
Bollen and Whaley, 2004).
6 Similarly, Pan (2002) shows that the volatility smirk (skew) pattern is primarily due to investors’ fear of

large, undesirable jumps.

7 They show that the information content of MFIV is superior to that of the Black-Scholes implied volatility

(hereafter BSIV) because the MFIV measure accounts for all strikes when computing IV at a particular
point in time, whereas the BSIV measure is a point-based IV and does not account for all strikes in
computation; i.e., each strike has a separate IV. Moreover, BSIV is subject to both model and market
efficiency, while MFIV is only subject to market efficiency (see Poon and Granger, 2003).
 4

the major IV indices, which used to employ at-the-money (ATM) IV measures in their
methodologies.8 The MFIV index is forward-looking and captures market sentiments
(aggregated beliefs), that are often referred to as the “investors’ fear gauge” (e.g.,
Whaley, 2000).9 Similarly, The MFIV measure contains a broader set of information
and is thereby more robust; the MFIV index is, as a result, an excellent tool for some of
the challenging volatility modeling and forecasting issues.10The first of these is the
strong daily negative asymmetric return-volatility relationship, where the volatility
index reacts differently to negative and positive returns—the important stylized fact
empirically observed in the market. The second is daily implied-VaR forecasting, i.e.
incorporating the information content of MFIV into different specifications of daily
VaR models, and then comparing and backtesting their daily forecasts with other
horse-race VaR models afterwards. Referring to the former, there are two main existing
hypotheses in the literature that characterizes the negative asymmetric return-volatility
relationship: the leverage effect and feedback effect hypotheses. However, both the
leverage and feedback hypotheses have been unable to explain the observed strong
negative asymmetric return-volatility relationship at daily frequencies (see, e.g., French
et al., 1987; Breen et al., 1989; Schwert, 1989, 1990). Similarly, a recent study by
Hibbert et al. (2008) found a very strong contemporaneous negative asymmetric
return-volatility relationship using daily data, thereby empirically rejecting both the
leverage and volatility feedback hypotheses.11 Further empirical investigations are
indeed required here to characterize an asymmetric return-volatility relationship using
a daily MFIV index measure.12 Importantly, we believe that the estimation technique
currently being used for this relationship is the standard ordinary least squares (OLS),
which may not fully characterize this relationship because it ignores the volatility
8 The motives for adopting MFIV measures are the following. First, the MFIV index measure is
economically appealing and robust, as it accounts for OTM put and call options (i.e., volatility
smirk/skew). Second, the previous IV index measure (now called VXO) was upward-biased, induced by
trading-day conversion, which is now omitted from the new VIX measure. Finally, with the new robust
MFIV index measure, it is possible to replicate volatility derivatives (e.g., variance swaps), which was not
possible with the previous measure.
9 Major option exchanges, including the Chicago Board of Exchange (CBOE) and the Deutsche Börse, have

launched IV indices that robustly provide information on options using MFIV measures; examples of this
are the VIX index for the S&P 500 index, VXN for the NASDAQ 100 index, VDAX for the DAX 30 index
and VSTOXX for the Dow Jones (DJ) EURO STOXX 50 index.
10 A recent study by Konstantinidi et al. (2008) attempted to forecast the evolution of implied volatility

index.
11 Other studies by Simon (2003) and Giot (2005) have also found a very strong negative asymmetric

return-volatility relationship using daily data.

12 We also believe that the asymmetric return-volatility relationship should be more pronounced with the

new robust MFIV volatility index measure in contrast to the old BSIV volatility index measure. Therefore,
it is important to compare the asymmetric responses of both MFIV and BSIV measures to negative and
positive returns.
 5

responses at both tails of the IV changes’ distributions. Therefore, to obtain a complete

picture of this relation, the conditional quantile regression should be preferred over
OLS regression, especially to investigate the asymmetric responses of volatility at the
uppermost quantiles.13 Few well-known studies exist showing a significant negative and
asymmetric relations between stock index returns and changes in the IV index using
OLS (or mean) regressions (e.g., Fleming et al., 1995; Whaley, 2000; Giot, 2005;
Simon, 2003; Skiadopoulos, 2004; Low, 2004; Dennis et al., 2006).

However, referring to the latter, i.e., modeling and backtesting of implied-VaR, as we

believe that VaR modeling is all about extreme events, the MFIV measure should be
important as input for any VaR model because it is implied by OTM options, which are
informed on future events and contain information on negative and positive jumps
(see, e.g., Liu et al., 2005; Doran et al., 2007; Camara and Heston, 2008; Bates, 2008).
Therefore, it is possible to use the MFIV measure as an input to implied-VaR or
implied-VaR augmented by GJR model and then to compare the daily-VaR forecasts
with the forecasts of other horse-race VaR models such as the filtered historical
simulation (hereafter FHS) proposed by Barone-Adesi et al. (1998) and RiskMetrics.
Consequently, a gap in the literature needs to be filled by further investigation in this
direction as well.14

However, in addition to equity index option markets, there is an enormous over-the-

counter (OTC) interest rate option market (such as swaptions) that also needs
consideration. In particular, investigation of the IV dynamics of the larger swaption
markets (e.g., the EUR-, USD-, and GBP-denominated swaption markets) is indeed
indispensable. In fact, European swaption contracts are quoted to be consistent with
the Black (1976) model’s IVs. The IVs recovered from swaptions contain important
information as well; therefore, their dynamics need further modeling and exploration.
Example questions are how many risk factors are driving each of the EUR, USD, and
GBP swaption IVs, whether these risk factors are linked across markets, and whether
they respond to each other’s shocks. Furthermore, is it possible to reproduce (or
forecast) the whole swaption maturity IVS by calibrating a multifactor model to each of
these swaption markets separately?

13
Quantile regression estimates are robust to outliers, non-normal error distributions, etc.
14To our knowledge, no study yet compared and backtested forecasts of implied-VaR’s with those of FHS-
VaR.
 6

This PhD thesis is thus positioned within the area of modeling and forecasting implied
volatility discussed above. The first essay in this thesis is modeling IVS and identifying
risk factors that account for most of the randomness in the IVS. The approach extends
the Dumas, Fleming and Whaley (DFW) (1998) framework; for instance, using
moneyness in the implied forward price and OTM put-call options on the FTSE 100
index, a nonlinear parametric optimization technique is used to estimate different DFW
models and thereby produce rich, smooth IVSs. Here, the constant-volatility model fails
to explain the variations in the rich IVS. Next, PCA is applied to the IVS, and risk
factors are extracted that account for most of the dynamics in the shape of the IVS; it is
found that the first three PCs can explain about 69-88% of the variance in the IVS. Of
this, on average, 56% is explained by the level factor, 15% by the term-structure factor,
and the additional 7% by the jump-fear factor. The second essay proposes a quantile
regression model (QRM) for modeling the strong negative asymmetric return-volatility
relationship with newly adapted MFIV indices, especially to quantify the effects of the
positive and negative stock index returns at various quantiles of the IV changes’
distribution. This model is the generalization of standard mean regression models, such
as those of Simon (2003), Giot (2005) and Hibbert et al. (2008). The results show a
pronounced negative asymmetric return-volatility relationship that is monotonically
increasing when moving from the median quantile to the uppermost quantile (i.e.,
95%); therefore, OLS underestimates this relationship at upper quantiles. Additionally,
the asymmetric relationship is more pronounced with the MFIV volatility index
measure in comparison to the old BSIV volatility index measure. Nonetheless, the
volatility indices are ranked in terms of asymmetric volatility as follows: VIX, VSTOXX,
VDAX, and VXN. The third essay examines the information content of the new VDAX
volatility index to forecast daily VaR estimates and compares its VaR forecasts with the
forecasts of the FHS and RiskMetrics models. These daily VaR models are then
backtested from January 1992 through May 2009 using unconditional coverage,
independence, conditional coverage, and quadratic score tests. It is found that the
VDAX volatility index subsumes almost all information required for the volatility of
daily VaR forecasts for a portfolio of the DAX30 stock index; implied-VaR models
outperform VaR models of the FHS (GJR) and RiskMetrics.

Three essays discuss modeling and forecasting of IV for equity markets. However, the
fourth essay emphasizes modeling and exploring of IV dynamics for the important
swaption markets; the essay models risk factors driving the EUR, USD, and GBP
swaption IVs. Two main purposes are then achieved. First, the common underlying
 7

implied factors are identified that affect IV movements in the EUR, USD, and GBP
swaption markets. Then, the dynamic interactions between the implied factors are
examined using techniques such as Granger causality and the generalized impulse
response function. We find that the first three implied factors explain 94-97% of the
variation in each of the EUR, USD, and GBP swaption IVs. There are significant
linkages across factors, and bi-directional causality is at work between the factors
implied by EUR and USD swaption IVs. Furthermore, in innovation-accounting
investigations, the factors implied by EUR and USD swaption IVs respond to each
others’ shocks; however, surprisingly, GBP does not affect them. Second, the string
market model (SMM) is calibrated for each of the swaption markets using multivariable
nonlinear optimization. The calibration results show that the SMM can efficiently
reproduce (or forecast) the entire swaption volatility matrix (or surface) for each of the
EUR, USD, and GBP markets.

The rest of the introductory part to the thesis is organized as follows. Section 2
discusses option markets’ empirical regularities and modeling. Section 3 discusses the
methodology. Section 4 discusses the four essays, their contribution to the literature
and the implications of their results.
 8

2 Options Markets’ Empirical Regularities and Modeling

One of the strongest empirical regularities of all time encountered with U.S. stock index
options appeared after the U.S stock market crash in 1987; when IVs are plotted against
any moneyness measure (along the strike dimension), volatility smirk (or skew)
appears (see, e.g., the empirical evidence of Bates, 1991; Rubinstein, 1994; Ait-Sahalia
and Lo, 1998; Foresi and Wu, 2005, Zhang and Xiang, 2008; and Carverhill et al.,
2009).15 In fact, there are three commonly observed shapes of the plot of IVs against
moneyness: smile, skew and smirk. The volatility smile is a pattern; IV increases as the
strike price increases (or decreases). The volatility skew is a pattern; there is a high IV
for low strike prices and a low IV for high strike prices. Finally, the volatility smirk is a
pattern; IV increases more quickly at low strike prices than at high strike prices (see,
for further details, e.g., Jiang and Tian, 2005; and Foresi and Wu, 2005).16 Rubinstein
(1985) documented that the volatility smile pattern existed in U.S. stock index options
before the 1987 crash. However, it has changed its pattern, following the 1987 market
crash, to a typical volatility smirk (or skew) pattern. On the other hand, we have
another stylized fact regarding the IV: the IV shows a volatility term structure pattern
when plotted against time to maturity (time dimension). Therefore, when the IVs (of
different strike prices and times to maturity) on a particular day are plotted against
moneyness and time dimensions, the implied volatility surface is viewed. As we have
empirically observed a rich IVS, it has two features that have attracted the interest of
academics and practitioners. First, the IVs systematically change with the strike price
and maturity dimensions, which induce an instantaneously non-flat IVS. Second, the IV
changes dynamically as time passes by because with the arrival of new information in
the option market, option prices are later updated based on the new information, i.e.,
the influences beliefs or expectations of investors.17 The dynamics of the IVS are
empirically confirmed to be driven mostly by the underlying risk factors (see the
empirical evidence of Dumas et al., 1998; Skiadopoulos et al., 1999; Cont and da
Fonseca, 2002; Mixon, 2002; and Gocalves and Guidolin, 2006; and Chalamandaris
and Tsekrekos, 2009).

15 The volatility smirk (skew) can be interpreted in terms of a premium that the market is willing to pay for

insurance against any future crash.

16 The volatility smirk pattern is observed when OTM put options are much more expensive than the

corresponding OTM call options (see Foresi and Wu, 2005).

17 See, for further discussion, the studies of David and Veronesi, 2000; Guidolin and Timmermann, 2003;

Chalamandaris and Tsekrekos, 2009.

 9

An important contribution of the David and Veronesi (2000), and others such as Garcia
et al. (2003), and Guidolin and Timmermann (2003) is that they use a general
equilibrium approach, attributing these empirical regularities about IVS dynamics to
investors’ beliefs and updating from the option prices about the processes of
fundamentals that increase IVs, e.g., the time-varying volatility or stochastic volatility,
jumps, net-buying pressure of options, and transaction costs. Therefore, the shape of
the IVS incorporates the aggregated underlying beliefs (usually, most option traders are
institutional investors who posses professional skills) of the investors; therefore, IVS is
very informative.18 Consequently, modeling and generating IVS is of paramount
importance for forecasting future volatility, option trading, pricing illiquid options,
hedging and pricing exotic derivatives, and the risk management of an options
portfolio.

Figure 1 shows the average weekly profile of the implied volatility surface for March
2004 implied from FTSE 100 index options, i.e., the market IVs (pink dots) versus a
constant-BS model projection (green). As can be seen, a rich market IVS exists, but the
BS model projects a flat IVS and constant IVs at every segment of the IVS. This
empirical evidence clearly rejects the constant BS model and further supports the
existence of rich IVS, which is a stylized fact regarding IVs.

Figure 1. The market IVS versus BS model projection for March 2004 implied by
FTSE100 index options.

18 Chalamandaris and Tsekrekos (2009) discuss different approaches to modeling IVS dynamics.
 10

Figures 2 and 3 provide the proportion of factors and their loadings, respectively, that
drive the IVS for March 2004 implied by FTSE 100 index options. The first three
factors (or PCs) can explain, on average, approximately 82% of the variation in the IVS.

Average Implied Volatility Surface, March- 2004

100 100%

90 90%

80 80%

70 70%
Variance Explained (%)

60 60%

50 50%

40 40%

30 30%

20 20%

10 10%

0 0%
1 2 3 4 5 6 7 8 9
Principal Components

Figure 2. Principal components driving the market IVS for March 2004.

Three Factor Loadings of the Implied Surface, March-2004

0.8

0.6 PC1
PC3

0.4

0.2

-0.2

-0.4
PC2
-0.6

-0.8

5 10 15 20 25 30

Figure 3. Factor loadings of the first three PCs from the market IVS for March 2004.
 11

The first factor, which is common to all IVs, is systematic for the entire IVS in the same
direction. It is a volatility-level factor (or parallel shifts) that explains about 64% of the
variations in IVS. It is important for IVs implied from the short- and long-maturity
options because it alters the steepness of the volatility pattern of the near-term OTM
put-call options and the effect is greater on the OTM options than on the ATM options.
In particular, OTM puts are affected the most, featuring a mean-reverting stochastic
volatility after a shock that increases instantaneous volatility, which then persists for
some time (the correlation between stochastic volatility and jump).19 The second factor,
which can be interpreted as a term-structure factor (or tilt, which generates shifts in the
slope of the term structure of IVs), explains about 10% of the variations in IVS and
affects the short- and long-maturity IVs differently (i.e., the sign of the effects differs
between the short- and long-maturity IVs). The term-structure factor is also consistent
with the notion that the segments in the IVS that are implied from long-term options
do not completely respond to the segments of the IVS that are implied from short-term
options; therefore, this factor can be attributed to the long-term risks such as risks
related to the market and the macroeconomy. Moreover, the second factor affects IVs
across the moneyness dimension identically. The third factor can be interpreted as a
jump-fear factor (or curvature, which changes the steepness of IV smirk/skew) and
explains about 8% of the variations in IVs. The effects of the jump-fear factor on the
OTM put and call options significantly differ, thereby inducing steepness in IV smirk
(skew). For empirical evidence on the dynamics of the factors driving the IVS of index
options, see the studies of Skiadopoulos et al. (1999), Mixon (2002), Cont and da
Fonseca (2002) and Chalamandaris and Tsekrekos (2009).

Another stylized fact is the strong negative asymmetric return-volatility relationship,

which implies that the IV reacts to negative and positive returns differently. The studies
of Simon (2003), Giot (2005) and Hibbert et al. (2008) have documented a very strong
negative asymmetric return-volatility relationship using daily data. The results of these
studies and many others put in question the two main hypotheses in the existing
literature that characterize the asymmetric return-volatility relationship: the leverage
effect and feedback effect hypotheses (see also, e.g., French et al., 1987; Breen et al.,
1989; Schwert, 1989, 1990). The leverage hypothesis proposed by Black (1976) and
Christie (1982) attributes asymmetric volatility to the financial leverage of a firm; i.e.,

19 Factor 1 can be also interpreted as being consistent with the view that the change in IVs changes the

skewness of the corresponding implied risk-neutral density (see, for further discussion, the study of Mixon,
2002; Zhang and Xiang, 2008).
 12

when a firm’s debt increases, the firm’s value declines, triggering the value of its stock
to decline further because the equity of a firm is more exposed to the firm’s total risk,
thereby increasing the volatility of equity.20 On the other hand, the volatility feedback
hypothesis is proposed by French et al. (1987), Campbell and Hentschel (1992), Poterba
and Summers (1986) and Bekaert and Wu (2000), who attribute asymmetric volatility
to a volatility feedback effect. In contrast to the leverage effect hypothesis, the volatility
feedback hypothesis states that increases in volatility induce negative stock returns.
The economic explanation should be that time-varying risk premia induce volatility
feedback because they represent the linkage between fluctuations in volatility and
returns. Increases in volatility imply that the required expected future returns will
increase as well; as a result, current stock prices decline. Hibbert et al. (2008) and
Breen et al. (1989) empirically found that the volatility feedback hypothesis does not
always hold; i.e., we do not always find positive correlations between current volatility
and expected future returns. Nevertheless, the accounting and economic explanations
might be important for characterizing an asymmetric return-volatility relationship at
lower frequencies, such as quarterly or monthly, but not for daily or higher frequencies.
Further investigations are important at a daily frequency to characterize the strong
negative asymmetric return-volatility relationship using robust MFIV (smirk-adjusted)
volatility indices. Importantly, an estimation technique such as quantile regression
should be preferred over OLS to capture the responses of the entire distribution of IV
changes to the negative and positive stock returns. Because the OLS will characterize
this relationship just at the mean of the response variable not at the other parts of the
response variable’s distribution, however, which could be nicely characterized by using
quantile regression.

Figure 4 provides a daily time-series plot (levels) of the stock market indices (S&P 500,
NASDAQ 100, DAX 30, and DJ Euro STOXX 50) versus the corresponding volatility
indices (VIX, VXN, VDAX, and VSTOXX) from February 2, 2001 to May 29, 2009. It is
evident that volatility indices are moving in the opposite direction to the stock market
indices. This phenomenon is stronger in crisis periods, particularly the credit-crunch
and liquidity crises in the year 2008, where historically high volatility levels of the
volatility indices are witnessed; for instance, VIX level twice surpassed 80%, and the
corresponding stock market crashed afterwards. Therefore, the inverse relationship is

20 However, Schwert (1990) argues that it is too strong for leverage hypothesis to completely explain

asymmetric volatility.
 13

the empirical evidence of a daily strongly negative return-volatility relationship, which

needs further examination.

Figure 4. Stock indices versus MFIV indices from February 2, 2001 to May 29, 2009.
 14

Nonetheless, these empirically observed stylized facts in the option markets also bring
challenges for modeling: the first is determining a method to model IVS and explore the
dynamics of the factors driving IVS. The second is to incorporate the rich information
of the MFIV measure with the quantile regression framework to model and investigate
the strong daily negative asymmetric return-volatility relationship, i.e., how the entire
IV changes’ distribution would react to the negative and positive returns, particularly to
examine the responses of the upper-most quantiles of the IV changes’ distribution.
Moreover, there is the question of whether the relationship would differ between the
BSIV and MFIV measures in a quantile regression framework. The third is to
incorporate the information content from the MFIV measure (smirk-adjusted) into our
daily VaR models for VaR forecasts and, furthermore, to compare and backtest the
daily implied-VaR models with other benchmark VaR models. The fourth is to
determine how the factors implied from swaption maturity IVS are dynamically linked
to each other and across markets and, furthermore, to calibrate a multifactor model to
the swaption market and thereby reproduce the entire swaption maturity IVS.

The following section briefly discusses our methodology for the four essays, such as
how we approach modeling and forecasting of implied volatility in the presence of
above regularities.
 15

3 Methodology
3.1 Calculating Black-Scholes Implied Volatility (BSIV)

Merton (1973) extended the Black-Scholes (1973) model for European call and put
options on dividend paying, St stock or stock index with strike price K , time-to-

maturity  T  t , call with payoff Max(S t  K,0) , and put Max (K  S t ,0) . The BS
formulae for the values of these call and put options are

C S t ,K,K, S t e  q  d 1   Ke  r  d 2 , (1)

P S t ,K,K, Ke  r   d 2   S t e  q   d 1 , (2)

d1

ln S t K    r  q   2 2 , (3)
 

d2

ln S t K    r  q   2 2 . (4)
 
where q is the dividend rate on the underlying asset, r is the risk-free interest rate, and

 ˜ is the standard-normal cumulative distribution function. Now consider a market

where the BS assumptions do not hold, as previously empirically found by many
researchers (see the empirical evidence of Bates, 1991; Rubinstein, 1994; Ait-Sahalia
~ ~
and Lo, 1998); the quoted market prices of put and call options are C t and P t . Then,
BS
there is a unique implied volatility  t that equates the theoretical option price with

the market option price given all other observed parameters:

~
C BS (S t ,K,K,tBS ) C t , (5)
~
PBS (S t ,K,K,tBS ) Pt . (6)

Nonetheless, with the help of the Bisection Method, all IVs are obtained for each strike
price and time to maturity, i.e., for the OTM put and call as well as the ATM options.
The bisection method is efficient and fast, and the advantage is that IVs can be
estimated without the knowledge of Vega. In the first essay, IVs are backed out from
OTM put and call options as well as from ATM options for modeling of the dynamics of
implied volatility surfaces for the FTSE 100 stock index.
 16

3.2 Model-Free Implied Volatility (MFIV) Index

Britten-Jones and Neuberger (2000) derived a model-free risk-neutral implied

volatility under the pure diffusion assumptions

ªT § dS
0Q « ³ ¨¨ t
·
¸¸
2
º
» 2³
f
 
C T,Ke rT  max S 0  K,0 
dK, (7)
«¬ 0 © S t ¹ »¼ 0 K2

The risk-neutral integrated implied volatility between the current date and a future
date is completely specified by a group of options expiring on the future date. The price
and volatility processes are completely model-free. In fact, in the option markets, the
cross-sections of the strike price are available in discrete periods; therefore, the above
continuous equation cannot be implemented in this situation. To overcome this issue,
Jiang and Tian (2005) have derived a discrete form of the MFIV that holds under a
variety of assumptions, such as diffusion, stochastic volatility and jumps:

ª n 1 § S  S
E0Q «¦ ¨¨ ti 1 ti ·
¸¸
2
º
»
§
¨u  ¸ ¦
 
1 · m C T,K i e rT  max S 0  K i ,0 
, (8)
«¬ i 0 © S ti ¹ »¼ © u ¹i  m Ki

where
T
ti ih, for i 0, 1, 2,....., n, h , Ki S0 u i , for i 0, r 1, r 2,....., r m,
n
1
u 1  k  m . The CBOE introduced a new VIX volatility index in September 2003
based on options on the S&P 500 stock index. The VIX is based on a similar MFIV
measure. The VIX is then determined from the bid-ask prices of the options underlying
the S&P 500 index; it provides an estimate of expected future realized stock market
volatility for the 22 subsequent trading days (over 30 calendar days). However, the old
VIX index, based on options on the S&P 100 index, was introduced in 1993 and has
now moved to the new ticker symbol VXO. In contrast to the VXO, which is based on
near-the-money BSIV options on the S&P 100 index, the new VIX uses market prices of
options on the S&P 500 index.21 This new model-free VIX methodology accounts for
both OTM put and call options (i.e., volatility smirk/skew). The new methodology is
thus more appealing and robust. The CBOE’s introduction of the new VIX was
motivated by both theoretical and practical deliberations. First, the new VIX is

21 The options on the S&P 500 index, in comparison with the options on the S&P 100 index, contain a

much broader set of implied information; the new VIX is thus a more informative measure than the old
VIX (now VXO).
 17

economically more appealing, as it is based on a portfolio of options, whereas the old

VIX was based on the ATM option prices. Second, the new VIX makes it easy to
replicate variance swap payoffs while using static positions in a range of options and
dynamic positions in futures trading. Third, the new VIX has removed the induced
upward bias of the old VIX in the trading day conversion (see, e.g., Carr and Wu,
2006). Similarly, in September 2003, the CBOE introduced the VXN using the same
MFIV methodology as that of VIX. The CBOE has calculated price histories for VIX and
VXN back to the years 1986 and 2001, respectively. 22 The model-free formula for the
new VIX volatility index calculation is

2
2 K i rT 1 § Ft ·
 2 t,T  ¦ e Ot K i ,T   ¨¨  1 ¸¸ , (9)
T  t i Ki2
T  t © K0 ¹
where T is time to expiry date of all S&P 500 index options (best bid and best ask of all
options), Ft is the forward price derived from the prices of the S&P 500 index options,

K i is the strike price of the ith OTM option, Ot K,T  is the mid-price (of the bid-ask

spread) of the OTM option at strike K i , K 0 is the first strike price below the forward

K i 1  K i
price Ft , K i is the interval between the strike prices, and r is the risk-
2
free rate at time t. Finally, the above equation is used to calculate  2 t,T  at two of the
nearest maturities of the available options. Then, via linear interpolation between the
two nearest maturities, the VIX estimate is obtained, leading to the estimation of the
annualized 30-day VIX

§ § N T  N T 30 · § N  N T1 · · N 365
VIX t 100 ¨ T1  12 ˜¨ 2 ¸  T2  12 ¨ 30 ¸¸ . (10)
¨ ¨ NT  NT ¸ ¨ N T  N T ¸ ¸ N 30
© © 2 1 ¹ © 2 1 ¹¹

where N T1 and N T2 denote the number of days to maturities T1 and T2 of the two S&P

500 index options, respectively, and N 365 is the time for a standard year.23

22 Later, the Deutsche Börse and Goldman Sachs jointly developed a similar model-free methodology for

the new VDAX and VSTOXX indices. The VDAX is based on options on the DAX 30 stock index, whereas
VSTOXX is based on options on the Dow Jones (DJ) Euro STOXX 50 stock index, which consists of the
Eurozone’s 50 largest blue-chip stocks. The price histories for both VDAX and VSTOXX were calculated
back to the years 1992 and 1999, respectively.
23See, for further detail on the VIX construction, the CBOE VIX white paper:
www.cboe.com/micro/VIX/vixwhite.pdf.
 18

3.3 Modeling Implied Volatility Smirk (Skew) and Surface

To model and obtain smooth implied volatility surfaces for FTSE100 stock index
options, we use a parsimonious parametric approach similar to the approach proposed
by Dumas et al. (1998). Their model is known as the practitioner BS model (see, e.g.,
Christoffersen and Jacobs, 2004). However, our approach somewhat differs from the
original work of Dumas et al. (1998). First, we incorporate most informed and liquid
options, i.e., OTM put and call options. Because put options represent the natural
hedging instrument for the underlying stock index portfolio, institutional investors
always hedge their portfolios using OTM puts, particularly when their forecast shows
any market declines in future (see, e.g., Pan, 2002; Bollen and Whaley, 2004; Foresi
and Wu, 2005; Bates, 2008; Camara and Heston, 2008).24 Second, we obtain IVS with
a practitioner-friendly moneyness measure that takes into account forward prices as
well as time dependence because the IVS should be able to be updated continuously on
the changes in both IVs and stock prices. We use a moneyness in implied forward price
that is similar to that of Gross and Waltner (1995). Moneyness in implied forward price,
r
m, is defined as the logarithm of the implied forward price Se t over the strike price
K , normalized by square root of time to maturity  , i.e.,

log(S t e r  t /K)
m , (11)

Four different structural forms of model specifications are estimated, leading to smooth
IVSs, each of which differs in the parametric structure that serves to characterize the
IVs:

Model 1: I m,  0   , (12)

Model 2: I m,  0  1 m   2 m 2  , (13)

Model 3: I m,  0  1 m   2 m 2   3    4  m  , (14)

Model 4: I m,   0   1 m   2 m 2   3    4  m   5  2  . (15)

24 However, if investors who forecast any market spikes in the future go long in the OTM call positions with

limited downside risks which remain to the premiums of the options only, instead taking positions in the
underlying stock index.
 19

Model 1, the flat IV function, is a constant in the BS model that gives volatility equal to
ATMIV. This does not account for volatility skew or smirk and term structure IV effects.
Model 2 attempts to capture volatility skew and volatility smirk across moneyness;
here, the slope, 1 , represents volatility skew, and the curvature, 2 , represents
volatility smirk effects.25 Model 3 captures an additional variation attributable to time
(e.g., time to maturity slope),  3 , and a combined effect of the skew and time

dimensions, 4 . Finally, Model 4 adds a parameter that captures the time-to-maturity

curvature effect,  5 . Further clarifications for the rest of notations are: I m,  , which

represents the dependence of IV on the moneyness level as well as time to maturity,

and 0 , which a parameter that is constant for all models.

In the first essay, the four structural models are estimated, and smooth surfaces are
obtained using the nonlinear optimization technique. In fact, the choice of the
nonlinear optimization technique is obvious because of the following reasons: first, it is
more common in the market practices, particularly in multifactor models such as the
well-known LIBOR Market Models. Second, optimization is more flexible. Third,
parameters need to be continuously calibrated to the market information. Fourth, this
technique can capture the nonlinear effects nicely. Nevertheless, we optimize the
surface with a nonlinear least square function (lsqnonlin) in MATLAB.26 However, to
test for the goodness of fit of these models, we employ two important loss functions
proposed by Christoffersen and Jacobs (2004): relative root mean squared errors
(%RMSE) and implied volatility root mean squared errors (IVRMSE). The former
minimizes the relative difference between the market (i.e., BSIVs) and practitioner BS
model IVs; it assigns greater weight to the OTM options, as these values are small. The
latter function minimizes the error between the market IVs and model IVs, and it
assigns equal weights to IVs.

In the first essay of the thesis, aside from parametric structural specifications for IVS,
we proceed to explore the movements in the IVS without parametric specifications.
Principal component analysis (PCA) is used to reduce the whole surface dimension to
just a few independent underlying risk factors that drive the IVS; the rationale is to
extract independent factors that explain most of the dynamics of the IVS. With PCA,

25In Model 2, when the slope is zero, it becomes a pure smile.

26This is a built-in function. It follows an algorithm that minimizes the sum of squared errors between the
actual value and the prediction for a given vector of a parameter.
 20

independent risk factors can be indentified without any prior assumptions. Many
researchers have previously attempted to extract risk factors that explain most of the
variations in the IV smirk (skew) or surface: for FTSE 100 index IV, Alexander (2001)
and Cont and Da Fonseca (2002); for S&P 500 index IV, Cont and Da Fonseca (2002),
and Skiadopoulos et al. (1999); and for DAX index IV, Fengler et al. (2003). It was
confirmed that about 70-90% of the total variation in the IVS or volatility smirk (skew)
can be attributed to just three risk factors: parallel shift, tilt, and curvature. Our
approach is, therefore, similar to that of Skiadopoulos et al. (1999); however, our data
differ. We used FTSE 100 stock index IVs, whereas they used S&P 500 stock index IVs.

We thus compute the first-difference IVs across different moneyness levels and times to
maturity.27 PCA is applied to the pool data of IVs to analyze the dynamics of the IVS.
The data are then assembled into different groups on the basis of trading days left to
maturity, i.e., 8-30, 30-60, 60-90, 90-150, 150-220, and 220 days and above. After
grouping, PCA is applied to each group for a detailed investigation of the dynamics of
the IVS.

PCA is a method of matrix decomposition into eigenvectors and eigenvalue matrices. It

is applied to a pool of IVs across moneyness levels and term structures, effectively
T
decomposing the covariance matrix as  , where the diagonal elements of  are
the eigenvalues and the columns  are the associated eigenvectors. The choice of
column labeling in  allows the ordering of PCs such that e 1 belongs to the largest

eigenvalue 1 , e2 belongs to second largest eigenvalue  2 , and so forth. In a highly

correlated IVS, the first eigenvalue would be much larger than the others.
Consequently, the first PC can explain much of the variation in the IVS. If the first three
PCs explain most of the variations in the IVS, then these PCs can replace the original IV
variables without loss of much information (see, e.g., Skiadopoulos et al., 1999;
London, 2004).

3.4 Modeling the Asymmetric Return-Volatility Relationship

In the second essay, we propose a quantile regression model for the negative
asymmetric relationship between returns on the stock market index and IV changes in

27This is because the data input to PCA must be stationary, as IVs are often nonstationary; see, for more
discussion on the data issues, Skiadopoulos et al. (1999).
 21

the volatility index. Our model is the generalization of the standard mean-regression
models (MRM) of Simon (2003), Giot (2005) and Hibbert et al. (2008), who have
empirically confirmed the asymmetric return-volatility relationship. However, we
extend their MRM by modeling the asymmetric return-volatility relationship using the
conditional QRM to examine how negative and positive stock index returns vary across
different quantiles of IV changes, i.e., how much this asymmetric relationship changes
across different quantiles of IV changes. Their MRM is considered a standard model in
our analysis specified in the second essay. Their MRM assumes that the effects of both
types of returns are static across different IV changes (i.e., response variables);
therefore, an MRM would miss important information across quantiles of the IV
changes’ distribution that we could otherwise model using our QRM, particularly when
determining how the median or perhaps the 95th or 5th percentiles of the response
variable (IV changes) are affected by negative and positive stock return variables
(regressor variables).28

Nonetheless, first we regress the daily volatility index changes (denoted VI it , where

i=VIX, VXN, VDAX, VSTOX) on the daily percentage continuously compounded

returns of the stock market index (denoted R it , where i=S&P 500, NASDAQ, DAX, DJ

Euro STOXX 50), Rit was used for positive returns, and Rit was used for negative

returns). For the positive returns, Rit Rit if Rit ! 0 , and Rit 0 otherwise. On the

other hand, for the negative returns, Rit Rit if Rit  0 , and Rit 0 otherwise.

Koenker and Bassett (1978) were the first to introduce quantile regressions, which
could effectively model the quantiles of the distribution.29 QRM is a generalization of
the MRM and is therefore a robust regression, especially in situations where errors are
non-normally distributed, i.e., are skewed and leptokurtic. Nonetheless, the QRM is
used to examine the asymmetric return-volatility relationship; for instance, the QRM,
the specification of the qth QRM is

3 3 3
VI it  q   ¦  iLq  VI it  L  ¦
iLq  Rit L  ¦ iLq  Rit L  u t , (16)
L 1 L 0 L 0

28 See for further discussion Meligkotsidou et al. (2009).

29 Koenker (2005) provides mathematical details on the quantile regression as well its different extensions.
 22

where  q  is the intercept,  iLq  represents the coefficients for the lagged IV changes in
q 
a volatility index i , L 1 to 3 ,
iL represents the coefficients for positive returns and

iLq  represents the coefficients for negative returns of a stock market index i, where

L 0 to 3 for both type of returns and the errors ut are assumed to be independent

from an error distribution  q (u t ) with the qth quantile equal to zero. The above QRM

implies that the qth conditional quantile of the dependent variable VI i given

VI it 1 ,Vit  2 ,Vit 3 ,Rit ,Rit1 ,Rit 2 ,Rit3 ,Rit ,Rit1 ,Rit 2 ,Rit3 and


denoting Qq VI i VI it 1 ,..... it  3 ,
,Rit ,..,Rit3 ,Rit ,..,Rit3 , is equal to

3 3 3
 q   ¦  iLq  VI it  L  ¦
iLq  Rit L  ¦ iLq  Rit L . The main feature of this quantile
L 1 L 0 L 0

q 
regression framework is that the effects of the variables captured by  iL(q) ,
iL ,and

iLq  vary for each qth quantile within the range q  (0,1). Furthermore, the framework

allows for heteroskedasticity in error u t , and the coefficients are different for different

quantiles. Consequently, a quantile regression provides a broader set of information

about the asymmetric return-volatility relationship here (i.e., the effects of negative and
positive returns not only on the mean of the volatility changes but it captures the effects
on all parts of the distribution of the volatility changes) than an OLS regression would,
particularly when the error distribution is not symmetric.30 Nevertheless, the QRM is
estimated using the quantile regression method proposed by Koenker and Bassett
(1978), which minimizes the asymmetric sum of absolute residuals and robustly models
the conditional quantiles of the response variable, i.e., changes in the volatility index in
our case:31

ª
min« ¦  ˆ  q VIit  ˆ  ˆiL VIit L 
ˆiL RitL  ˆiL RitL
¬«t::it tˆ  iL VIit  L 
ˆiL Rit  L  iL Rit  L
ˆ
(17)
º
 ¦ (1  q) VIit  ˆ  ˆ iL VIit  L 
ˆiL Rit L  ˆiL Rit L » ˜
t::it ˆ  ˆiL VIit  L 
ˆiL Rit L  ˆiL Rit L »¼

30 Because the differences between the mean and the median produce asymmetric distributions, see, for a

more detailed explanation, Meligkotsidou et al. (2009).

31 For a discussion of quantile models and their estimation techniques, see Koenker (2005).
 23

3.5 Modeling and Backtesting Daily Value-at-Risk Models

Value at risk (VaR) is defined as the maximum expected loss in the value of a portfolio.
It has a certain probability over a certain holding period (for details on VaR, see, e.g.,
Duffie and Pan, 1997; Jorion, 2000; Dowd, 2005; Christoffersen, 2008). VaR forecasts
are fundamental to financial risk management and risk regulation. However, the
importance and recognition of VaR as a risk management tool came from the Market
Risk Amendment (1996) to the Basel Capital Accord of 1998 and also because of the
popularity of RiskMetrics introduced by J.P. Morgan (see Jorion, 2000; Dowd, 2005).
Afterwards, VaR became widely accepted by banks and was also imposed by regulators.
The aim of these two groups was to supervise and manage market risk, as market-
exposure risk was arising due to unfavorable movements in equity, interest rates, and
exchange rates, among other indicators.32 Because VaR has become a standard measure
for market risk due to the immense trading activities and market positions taken by
large banks, most financial institutions and trading houses currently use VaR models to
assess their daily portfolio losses from significant trading activities. This practice leads
to backtesting of VaR models by observing when the portfolio returns exceed the VaR
forecasts, i.e., when the VaR forecasts match their expectations. As a result, accurate
VaR forecasts are crucial for market risk management. Given that accurate VaR
forecasts are heavily dependent on the accurate forecasting of the volatility of a
portfolio, this is an important parameter for any VaR model. For instance, the level of
the VaR over a one-day holding period is defined as the solution

P(rt P  VaRtP t 1 ) , (18)

where  is 1 minus the VaR confidence level (i.e., 99%) and rt P is the return of a

portfolio over a one-day holding period. Having conditional volatility specification ht

such that rt P ht  t , where the residuals are distributed as  t ~N 0,ht  , then a one-

period VaR at time t is

VaR t  1   ht , (19)

where  denotes the standard-normal cumulative distribution function. In fact, we

need to accommodate heavy tails in the VaR estimation; therefore, VaR needs to be

32
The Basel committee for banking supervision allows banks to use VaR as a benchmark to determine how
much additional capital is needed to cover market risk.
 24

estimated using a Student’s t density function.33 We assume that  t are distributed as

follows:

 t ht    2  ~ ,
12
(20)

where   is the standardized Student’s t-distribution with  degrees of freedom.

To estimate the %daily VaR using Student’s t density function, we follow the method
described by Dowd (2005):34

 1  §¨   2  ·¸
1
2
VaRt, ht , (21)
© ¹
where the shape  is to be estimated. In the third essay, we used four different types of
volatility forecasts [implied volatility (VDAX), implied volatility (VDAX) with GJR, FHS
(GJR), and RiskMetrics] as parameters for the daily VaR models.

For the implied-VaR model, we considered the new VDAX index (MFIV index), which
is a robust and more informed volatility measure. The rationale for considering the new
VDAX is that we believe that VaR modeling is about extreme events; implying volatility
from the extreme outcomes in the options market is of paramount importance to VaR
forecasting. These extreme events are embedded in the IV derived from OTM options
(see, for instance, Liu et al., 2005; Camara and Heston, 2008; and Bates, 2008), i.e.,
the IV smirk-adjusted volatility measure. Likewise, the importance of the new VDAX
measure increases because IV smirk accounts for the net buying pressure of the put
options as well (see Bollen and Whaley, 2004). Volatility smirk (skew) is an obvious
phenomenon previously documented by many other researchers and is important to
capture in any volatility measure (e.g., Skiadopoulos et al., 1999; Alexander, 2001; Cont
and Da Fonseca, 2002; Low, 2004; Goncalves and Guidolin, 2006; Badshah, 2008).
Furthermore, information from trading strategies and other shocks are well absorbed
into the new VDAX index, as it accounts for a cross-section of options. Finally, as the
majority of option traders is very informed and possesses professional skills, the new
VDAX represents the beliefs of the informed traders (see, e.g., Low, 2004 and
Chakravarty et al., 2004). As a result, the new VDAX is a good candidate for a volatility
parameter in the daily VaR model to quantify a daily VaR forecast for the DAX 30 stock

33 Most previous studies have concluded that distribution functions accounting for fat tails are

fundamental to VaR modeling. See, for instance, Huisman et al. (1998), Alexander and Sheedy (2008) and
Giot (2005) for VaR forecasts obtained through different density functions.
34 Also, because the returns on the DAX 30 are non-normally distributed, the VaR forecasts are estimated

using Student’s t distribution function.

 25

index portfolio.

However, for the daily VaR model, a daily-variance parameter is needed in place of the
standard deviation because VaR uses variance as an input, as the VDAX is expressed in
annualized standard-deviation units. As a result, transformation is essential; for
instance, at time t, we insert VDAX t 1 into the daily VaR model as35

2
ht  imp,1,t 1 , (22)

while
2
 imp,1,t 1  VDAX t  1 2
252
. 23)

However, Granger and Poon (2003, 2005) point out that BSIV is biased and is always
higher than the actual volatility. They suggest using HV for calibration as

ht    imp,1,t
2
1 , (24)

where  and need to be estimated. However, we assert that VDAX is calculated

using the MFIV measure, whose IV value should not be subject to model risk, and that
this bias should thus not be of great concern; therefore, any of the above two variance
measures can be used equally. Furthermore, we have a combined specification for
variance using GJR-GARCH(1,1) extended with the lagged IV as

ht  0   1  t21   2  t2 1 d t 1  h t  1   imp,1,t

2
1 . (25)

This equation has a dummy variable used to capture asymmetry. For instance, the
dummy variable d t 1 is equal to 1 when  t 1  0 and is equal to 0 otherwise. Therefore,

estimation of an % daily implied (VDAX) VaR forecast or the combined VaR forecast
using implied volatility (VDAX) plus GJR can be done with the following specification:

 1  §¨   2  ·¸
1
2
VaRt, ht . (26)
© ¹

where  is the standardized Student’s t-distribution function, the shape  parameter

needs to be estimated, and D is the quantile (which in our case is 99%, 97.5% or 95%).
Another two volatility forecasts (FHS-GJR and RiskMetrics) as parameters for the
daily-VaR models are also used for the comparison with implied-VaR models. In the

35 A similar transformation scale is used by Blair et al. (2001) for the VIX index.
 26

third essay, we thus backtest the daily-VaR models [implied volatility (VDAX), implied
volatility (VDAX) with GJR, FHS (GJR), and RiskMetrics] from January 1992 through
May 2009 using unconditional coverage, independence, and conditional coverage tests.
A quadratic score was also estimated for each of the models.

3.6 Modeling Risk Factors driving the Swaption Volatilities

In the fourth essay, a PCA technique is used to reduce the swaptions maturity surface to
only a few risk factors and then further model their implied dynamics across markets in
a vector-autoregressive (VAR) framework (i.e., the EUR, USD, and GBP swaption
markets). PCA can be applied to the Black (1976) consistent swaption IVs; as in the
swaption markets, swaption IVs are quoted in a manner consistent to that used by the
Black model. Numerous studies have been conducted on the dynamics of IVS implied
from equity index options.36 However, the study by Bruggemann et al. (2008) most
closely meets this objective; in this study, the stochastic properties of the factors of IVs
(i.e., those obtained from options on the DAX 30 index) were analyzed and modeled in
a VAR framework and significant interactions were found between the factor loadings.

Therefore, in a similar vein, the VAR model of Sims (1980) is used to investigate the
dynamic interactions between the first two implied factors (i.e., PC1s, PC2s extracted
from each of the EUR, USD, and GBP Swaption IVs).37 This model successfully treats
each endogenous variable in a system as a function of the lagged values of all
endogenous variables in the dynamic simultaneous equation system. Therefore, the
VAR model for the factor loading dynamics can be represented mathematically

L
Ft   ¦  i Ft  i  u t , (27)
i 1

where Ft (z t1EUR ,z t2EUR ,z t1USD ,z t2USD ,z t1GBP ,z t2GBP )c is an m u 1 vector of endogenous variables

representing the factors in the EUR, USD, and GBP swaption IVs, ^i , i 1, 2, 3, ˜ ˜ ˜ , L`

is an m u m matrix of coefficients, and ut is an m u 1 vector of innovations that can be

36 For example, Cont and da Fonseca (2002), Mixon (2002), Skiadopoulos et al. (1999), Fengler et al.

(2003), Badshah (2008) and many others.

37 The first two PCs are selected because they are able to explain on average 94% of the variation in each of

the market IVs. We do not consider PC3s in our VAR framework. It helps to maintain a reasonable number
of parameters, increase efficiency, and achieve parsimony.
 27

contemporaneously correlated and uncorrelated with its own lagged values and with
other variables. Accordingly, the VAR model is estimated by OLS estimation.

However, to investigate the direction between PCs, we apply the Granger (1969)
causality test, which establishes a causal relationship between the PCs and hence
confirms the lead-lag relationship. The Granger causality test establishes that PC i is

Granger-caused by PC j if the information in the past and present values of PC j helps

to improve the forecasts of PC i . However, Granger causality in a VAR system using the

F-test provides information only about which variables impact the future values of each
of the variables in the VAR system. In this way, F-test values do not provide the sign of
the relationship, speed or persistence. The VAR’s impulse response functions (IRF)
could provide information about this kind of dynamic relationship (see, e.g., Brooks,
2002). Therefore, a generalized version of Pesaran and Shin (1998) is preferred, as it
does not require the orthogonalization of shocks and is invariant to the reordering of
variables in the VAR system. Therefore, an IRF measures the responses of the variables
in the dynamic VAR system (in our case, the first two factors of each the EUR, USD,
and GBP swaption IVs) when a shock is given to each factor: a one-standard-error
shock is applied to the error of a factor, and the effect on the dynamical VAR system
over a specified period of time is recorded.

Furthermore, in the fourth essay, we model and reproduce the swaption maturity IVS
(or volatility matrix) on a particular day using orthogonal PCs in an SMM framework
for each of the EUR, USD, and GBP swaption markets. Santa-Clara and Sornette
(2001), and Longstaff et al. (2001) were the first to introduce the SMM model, also
known as the high- or infinite-dimensional model. Later, Kerkhof and Pelsser (2002)
showed that the SMM and the LMM are observationally equivalent. However, for the
SMM calibration, we select a separated approach proposed by Gatarek et al. (2007) that
is a calibration technique for multifactor interest rate models used widely in the
financial industry. Separated calibration could be implemented in a straightforward
way.38 For instance, the SMM model could be calibrated to the whole swaption
maturing IVS. In fact, calibration of the model to a swaption matrix is preferred over
the caps because it is essential in a situation where the prices of derivatives instruments
are dependent on covariance/correlation structures in addition to volatilities such as

38 Technical details of the separated calibration are provided in Appendix A of the fourth essay, or see

Gatarek et al. (2007).

 28

the values of exotic instruments including as the Bermudan type of swaption, which are
more dependent on swaption prices than on caps.39

4 Summaries of the Essays

This PhD thesis, Modeling and Forecasting Implied Volatility: Implications for
Trading, Pricing, and Risk Management, consists of four essays. All of the four essays
are single-authored. This section provides a brief summary of the essays, their
contribution to the literature and their implications.

4.1 Essay 1: Modeling the Dynamics of Implied Volatility Surfaces

This essay models implied volatility surfaces and identifying risk factors that account
for most of the randomness in the volatility surfaces for the FTSE 100 index options. In
general, the essay attempts to model and uncover important stylized facts in the stock
index options appearing after the 1987 crash. The approach is an extension that follows
the framework by Dumas, Fleming and Whaley (1998); for instance it uses moneyness
in the implied forward price and OTM put-call options on the FTSE 100 stock index,
and a nonlinear parametric optimization technique is then used to estimate different
DFW (1998) models.

We estimated four parametric models of DFW (1998); the constant-volatility Model 1

fails to capture variations in the IVSs, yielding only flat IVSs. However, Model 2, which
captures the volatility skew and volatility smirk effects, does a good job and captures
most of the volatility skew and smirk effects. When we estimated Models 3 and 4, which
account for both the volatility term structure and volatility smirk/skew effects, they
performed better than the one-dimensional smirk/skew model. However, both models
fit the market IVS well, generating the best IVSs for the observed data. In the second
part of the study, which uses PCA (a nonparametric approach), we find that the first
three factors can explain about 69-88% of the variances in the IVS. Of this fraction, an
average of 56% is explained by the level factor, 15% is explained by the term-structure
factor, and an additional 7% is explained by the jump-fear factor. Finally, when
applying PCA to the maturity groups, we find that level factors (or parallel shifts) are
evident and thus important for shorter- and longer-maturity IVs. However, the term-

39 See Gatarek et al. (2007) for a detailed discussion.

 29

structure factor is important for IVs with medium maturities, and the jump-fear factor
is important for the IVs with the shortest (8-90 days) maturities.

IVS is important in practice; for instance, in the FTSE 100 index option market, 100
options on average are traded daily with a variety of strike prices and time to
maturities. If the practitioner needs to quote more options for different strikes or
maturities, IVS can be used as a reference; the practitioner could easily quote new
options by taking estimates from IVS. Moreover, the shape of the IVS is important in
option trading; one of its most important characteristics is that it can be used to
identify mispricing in the market. Second, the IVS can be used to manage exotic
derivative positions, which could be hedged with plain-vanilla options such as
European options. Third, a robust volatility index could easily be constructed from the
IVS and consequently used to price volatility derivatives. Fourth, risk can be managed
for securities that underlie the FTSE 100 stock index by using independent risk factors
from the IVS.

4.2 Essay 2: Quantile Regression Analysis of Asymmetric Return-

Volatility Relation

Essay 2 proposes a quantile regression model (QRM) for modeling of the negative
asymmetric return-volatility relationship with newly adapted IV indices, especially to
quantify the effects of the positive and negative stock index returns at various quantiles
of the IV changes’ distribution; in particular, this essay is an attempt to model and
explore another important stylized fact: the strong negative asymmetric return-
volatility relationship at high frequencies. The quantile regression model that we
propose is the generalization of the standard mean regression models (MRM) of Simon
(2003), Giot (2005) and Hibbert et al. (2008). Nonetheless, we investigated the
asymmetric return-volatility phenomenon in the newly adapted robust volatility indices
(i.e., the VIX, VXN, VDAX, and VSTOXX) using quantile regression. In particular, we
quantified the effects of positive and negative stock index returns at different quantiles
of IV changes’ distributions, asking about the degree to which the asymmetric
responses at the uppermost quantiles are comparable to the responses of median (or
mean) regressions. Additionally, as Bollen and Whaley (2004) have documented, the
net buying pressure for stock index put options from institutional investors seeking to
hedge their portfolios induces increases in IVs. Likewise, new IV indices incorporate
both OTM put and call options and are thus highly informed and robust measures.
 30

Accordingly, they should present more pronounced asymmetric return-volatility

relationships in comparison with their older counterparts.

There is noticeable evidence that the volatility indices VIX, VXN, VDAX, VDAXO and
VSTOXX from February 2001 through May 2009 responded in a pronounced
asymmetric fashion to the negative and positive returns of their corresponding stock
indices; the asymmetry monotonically increases when moving from the median
quantile to the uppermost quantile (i.e., 95%). Therefore, OLS underestimates this
relationship at upper quantiles. These IV indices thus sharply rise during market
declines (fear) and fall during market rallies (exuberance). The VIX presents the
highest asymmetry, followed by the VSTOXX, VDAX and VXN volatility indices.
Second, our argument that asymmetry with a volatility smirk (skew)-adjusted volatility
index measure (MFIV) should be pronounced is confirmed by comparing the
asymmetric responses of VDAX (MFIV) and VDAXO (BSIV); the MFIV index responds
in a pronounced fashion when compared to the BSIV index. Third, we also confirmed
that a significant amount of asymmetry occurs contemporaneously rather than with a
lag, thus rejecting the leverage hypothesis, and that a similar conclusion can be drawn
for the feedback hypothesis.

Our results have a number of implications. First, as we found that newly adapted MFIV
indices are strongly negatively correlated with their corresponding stock indices and
that the MFIV indices are important instruments for hedging stock portfolios.
Derivatives exchanges provide liquid markets for the futures and options underlying
MFIV indices. Therefore, a position in futures or options on a MFIV index can
completely hedge a stock portfolio position without consideration of complicated stock
index option-trading strategies; for instance, a call option on the MFIV index can be
seen as put option on the underlying stock index. Second, when the stock index drops,
the MFIV index rises sharply. Therefore, MFIV indices are useful for assessing not only
potential risks but also speculative transactions by risk-seeking investors. Third,
because the MFIV indices are based on the robust MFIV concept and provide better
tradability, it is easier for issuers of derivatives to engineer structured products based
on the MFIV indices. Fourth, trading strategies with regard to range could generate
profits; an example of this could be a volatility-long position in decreasing volatility
markets paired with a volatility-short position in increasing volatility markets.
 31

4.3 Essay 3: The Information Content of the VDAX Volatility Index and
Backtesting Daily Value-at-Risk Models

The third essay examines the information content of the new VDAX volatility index to
forecast daily VaR estimates for a DAX 30 index portfolio and compare its VaR
forecasts with the forecasts of the other horse-race models such as Filtered Historical
Simulation (FHS) and RiskMetrics. Many researchers have recently documented that
the information content of a volatility measure implied from the OTM put and call
options contains broader market-wide sentiments (beliefs that are updated upon option
trading) about future volatility; OTM options embed negative and positive volatility
jumps, stochastic volatility, and demand-supply pressure (see, e.g., Pan, 2002; Liu et
al., 2005; Bates, 2008; Camara and Heston, 2008; Doran et al., 2007, Bollen and
Whaley, 2004). We know that VaR modeling is all about the extreme events; therefore,
the new VDAX (MFIV index) index, which incorporates most of the information
required for VaR forecasting, is a perfect candidate to consider for VaR forecasting.

The information content of the new VDAX was incorporated into daily VaR forecasts
and compared with the VaR forecasts from the FHS (GJR) and the RiskMetrics models
at various confidence levels (i.e., 99%, 97.5% and 95%), using unconditional coverage,
independence, and conditional coverage tests for backtesting of each of the VaR
models. Furthermore, a quadratic score was estimated for each VaR model for the
period from January 1, 1992 through May 29, 2009. The backtesting results showed
that the new VDAX index contains significant information about actual volatility in VaR
models. The null hypotheses of independence and conditional coverage backtests were
never rejected for implied volatility, implied volatility plus GJR, or FHS (GJR) VaR
models. The number of VaR exceptions was not significantly different from the set
coverage rates. However, the null hypotheses of the RiskMetrics model were rejected
for lower confidence levels. It was also found that implied volatility and implied
volatility-GJR VaR models presented the fewest VaR exceptions and clusters of
exceptions, in contrast to the FHS (GJR) and RiskMetrics models. On the other hand,
the quadratic score for each model suggests the following ranking of VaR models:
implied volatility (VDAX), combined (implied volatility plus GJR), FHS, and
RiskMetrics. Our findings have implications for traders who hold long positions, risk
managers (internal), and regulators (external).
 32

4.4 Essay 4: Modeling Risk Factors Driving the EUR, USD, and GBP
Swaption Volatilities

The fourth essay models risk factors driving EUR, USD, and GBP swaptions’ implied
volatilities. In particular, the essay attempts to answer the following: first, how many
common underlying implied-risk factors are driving the IV maturity surfaces of the
EUR, USD, and GBP swaption markets? This essay then attempts to examine the
dynamic interactions between the implied risk factors across markets by using
techniques such as Granger causality and the generalized impulse response function.
Finally, this study aims to calibrate a multifactor model such as a string market model
(SMM) to swaption IVS by using multivariable nonlinear optimization to reproduce the
swaption surfaces for the EUR, USD, and GBP swaption markets.

First, we applied PCA to each of the EUR, USD, and GBP swaption IVs to discover the
important risk factors. We found that three risk factors explain about 94 -97% of the
variance in each of the EUR, USD, GBP swaption IVs. Second, the significant implied
factors present high correlations across swaption markets; consequently, there are
strong linkages across the three markets. Bi-directional causality is at work between the
implied factors from each of the EUR and USD swaption markets. The factors from
EUR and USD swaption markets Granger-cause the factors from the GBP swaption
market, but not vice-versa. Furthermore, in innovation-accounting investigations,
shocks to both factors implied by the EUR and USD IVs are found to be influential for
the factor implied from the GBP IVs. However, a shock to the GBP factors does not
affect the factors observed in the other two markets. Finally, there are many similar
characteristics between EUR and GBP markets, in contrast to the USD swaption
market. The whole swaption matrix for the EUR, USD and GBP markets is reproduced
using the SMM model. The fewest differences are observed between the theoretical and
market volatilities for EUR, GBP, and USD, respectively. Here too, similar
characteristics are found between EUR and GBP markets.

The identification of risk factors is important in practice. These factors can be used for
hedging of the portfolio position, particularly Vega-hedging, generating smooth implied
volatility surfaces, managing risk, and calibrating models. First, as in the interest rate
market, we find securities with a variety of maturities; therefore, it is almost impossible
for a trader to Vega-hedge portfolios against each and every individual risk. Therefore,
in the easiest way to account for most of the risks, a trader uses the major risk factors
 33

instead of hedging against every underlying risk. Second, we find a limited number of
OTC-traded swaptions in the swaption markets. When a trader needs to quote options
with different maturities, he is generally restricted. This could be solved by a swaption
volatility surface, thereby allowing us to use the risk factors to generate a swaption
volatility surface. A trader can easily quote new options by taking estimates from the
volatility surface. Third, by employing these factors, a trader can manage the overall
downside risk for his portfolio, i.e., the value-at-risk for a portfolio consisting of
interest rate derivatives. Finally, market models could be easily calibrated to only few
important factors, i.e., LIBOR market models.

The SMM is an important pricing and hedging tool. First, this model enriches the LMM
by calibrating to the whole swaption matrix, whereas the LMM is calibrated to only a
few risk factors. Therefore, the SMM calibration accounts for all independent risk
factors in a parsimonious fashion. Second, once a whole swaption matrix is reproduced,
the pricing and hedging of exotic derivatives such as Bermudan swaptions could be
done using these plain-vanilla products. Finally, in conjunction with SMM,
multivariable nonlinear optimization could be used for increased accuracy and
efficiency in the pricing and hedging of exotic interest rate derivatives.
 34

References

Ait-Sahalia, Y., and Lo, A., 1998. Nonparametric estimation of state-price densities
implicit in financial asset prices. Journal of Finance 53, 499-547.
Alexander, C., 2001. Principal component analysis of implied volatility smiles and
skews. ISMA Discussion Paper, University of Reading,UK.
Alexander, C., Sheedy, E.,2008. Developing a stress testing framework based on market
risk Models. Journal of Banking & Finance 32, 2220-2236.
Badshah, I., 2008, Modeling the dynamics of implied volatility surfaces. European
Financial Management Association (EFMA)-2008, Athens, Greece, Conference
Proceedings.
Barone-Adesi, G., Bourgoin, F., Giannopoulos, K., 1998. Don’t look back. Risk 11, 100-
103.
Bates, D., 1991. The crash of 87 : was it expected ? The evidence from options markets.
Journal of Finance 46, 1009-1044.
Bates, D., 2008. The market for crash risk. Journal of Economic Dynamics & Control
32, 2291- 2321.
Bekaert, G., and Wu, G., 2000. Asymmetric volatility and risk in equity markets.
Review of Financial Studies 13, 1-42.
Black, F., and Scholes, M., 1973. The pricing of options and corporate liability, Journal
of PoliticalEconomy 81, 636-654.
Black, F., 1976. The pricing of commodity contracts. Journal of Financial Economics 3,
167-179.
Black, F., 1976. Studies of stock market volatility changes. Proceedings of the American
statistical association, Business and Economic Statistics Section, 177-181.
Blair, B., Poon, S., and Taylor, S., 2001. Forecasting S&P 100 volatility: The
incremental information content of implied volatilities and high frequency
index returns. Journal of Econometrics 105, 5-27.
Bollen, N., and Whaley, R., 2004. Does net buying pressure affect the shape of the
implied volatility functions? Journal of Finance 59, 711-753.
Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. Journal
of Econometrics 31, 307-327.
Breen, D., Glosten, L., and Jagannathan, R., 1989. Economic significance of predictable
variations in stock index returns. Journal of Finance 44, 1177-1189.
Britten-Jones, M., and Neuberger, A., 2000. Option prices, implied processes, and
stochastic Volatility. Journal of Finance 55, 839-866.
Brooks, C., 2002. Introductory econometrics for finance. First edition, Cambridge
University Press, Cambridge, United Kingdom.
Bruggemann, R., Härdle, W.,Mungo, J.,Trenkler, C., 2008. VAR modeling for dynamic
loadings driving volatility strings. Journal of Financial Econometrics 6, 361-381.
Camara, A., Heston, S., 2008. Closed form option pricing formulas with extreme
events. Journal of Futures Markets, 28, 213-230.
Campbell, J., and Hentschel, L., 1992. No news is good news: an asymmetric model of
changing volatility in stock returns, Journal of Financial Economics 31, 281-
318.
Carverhill, A., Cheuk, T., and Dyrting, S., 2009. The smirk in the S&P 500 futures
options prices: A linearized factor analysis. Review of Derivatives Research 12,
109-139.
Carr, P., Helyette, G., Madan, D., and Yor, M., 2002. The fine structure of asset returns,
an empirical investigation. Journal of Business 75, 305-332.
Carr, P., and Wu, L., 2006. A tale of two indices. Journal of Derivatives 13, 13-29.
Chakravarty, S., Gulen, H., and Mayhew, S., 2004. Informed trading in stock and
option markets. Journal of Finance 59, 1235-1258.
 35

Chalamandaris, G., and Tsekrekos, A., 2009. Predicting dynamics in implied volatility
surfaces from OTC currency options. Forthcoming in the Journal of Banking
and Finance.
Christoffersen, P., 2009. Value-at-Risk models. Handbook of Financial Time Series,
First edition, Springer Verlag, Berlin, Germany.
Christensen, B.J., and Prabhala, N.R., 1998. The relation between implied and realized
volatility. Journal of Financial Economics 50, 125-150.
Christoffersen, P., and Jacobs, K., 2004. The importance of the loss function in option
valuation. Journal of Financial Economics 72, 291-318.
Christie, A., 1982. The stochastic behaviour common stock variances: Value, leverage,
and interest rate effects. Journal of Financial Economics 10, 407-432.
Cont, R., and Da Fonseca, J., 2002. Dynamics of implied volatility surfaces.
Quantitative Finance 2, 45-60.
David, A., and Veronesi, P., 2000. Option pricing with uncertain fundamentals: Theory
and evidence on the dynamics of implied volatilities. CRSP Working Paper
No.485.
Day, T., and Lewis, C., 1992. Stock market volatility and the information content of
stock index options. Journal of Econometrics, 52, 267-287.
Dennis, P., Mayhew, S., and Stivers, C., 2006. Stock returns, implied volatility
innovations, asymmetric volatility phenomenon. Journal of Financial and
Quantitative Analysis 41, 381-406.
Doran, J., Peterson, D., and Tarrant, B., 2007. Is there information in the volatility
skew? Journal of Futures Markets 27, 921-959.
Dowd, K., 2005. Measuring market risk. Second edition, Wiley Finance, Sussex,
England.
Duffie, D., Pan, J., 1997. An overview of Value at Risk. Journal of Derivatives 4, 7-49.
Dumas, B., Fleming, J., and Whaley, R.E., 1998. Implied volatility functions: Empirical
tests. Journal of Finance 53, 2059-2106.
Ederington, L., and Guan, W., 2002. Is implied volatility an informationally efficient
and effective predictor of future volatility? Journal of Risk, 4, 29-46.
Engle, R., 1982. Autoregressive conditional heteroskedasticity with estimates of the
variance of UK inflation. Econometrica 50, 987-1008.
Fengler, M., Härdle, W., and Villa, C., 2003. The dynamics of implied volatility: A
common principal components analysis approach. Review of Derivative
Research 6, 179-202.
Fleming, J., Ostdiek, B., and Whaley, R., 1995. Predicting stock market volatility: A new
Measure. Journal of Futures Markets 15, 265-302.
Fleming, J., 1998. The quality of market volatility forecasts implied by S&P100 index
option prices. Journal of Empirical Finance 5, 317-345.
Foresi, S., and Wu, L., 2005. Crash-o-phobia: a domestic fear or a worldwide concern.
Journal of Derivatives 13, 8-21.
French, K., Schwert, W., and Stambaugh, R., 1987. Expected stock returns and
volatility. Journal of Financial Economics 19, 3-29.
Garcia, R., Luger, R., Renault, E., 2003. Empirical assessment of an intertemporal
option pricing model with latent variables. Journal of Econometrics 116, 49-83.
Gatarek, D., Bachert, P., and Maksymiuk, R., 2007. The Libor market model in practice.
First edition, Wiley Finance, Sussex, England.
Glosten, L., Jagannathan, R., and Runkle, D., 1993. On the relation between expected
excess return on stocks. Journal of Finance 48, 1779-1801.
Granger, J., 1969. Investigating causal relationships by econometric models and cross-
spectral Analysis. Econometrica 37, 424-438.
Gross, L., and Waltner, N., 1995. S&P500 options: Put volatility smile and risk
aversion, Salomon Brothers, mimeo.
 36

Ghysels, E., Santa-Clara, P., and Valkanov, R., 2005. There is a risk-return tradeoff
after all. Journal of Financial Economics, 76, 509-548.
Giot, P., 2005. Implied volatility indices and daily value at risk models. Journal of
Derivatives 12, 54-64.
Giot, P., 2005. Relationships between implied volatility indices and stock index returns.
Journal of Portfolio Management 31, 92-100.
Goncalves, S., and Guidolin, M., 2006. Predictable dynamics in the S&P 500 index
options implied volatility surfaces. The Journal of Business 79, 1591-1635.
Guidolin, M., and Timmermann, A., 2003. Option prices under Bayesian learning:
Implied volatility dynamics and predictive densities. Journal of Economic
Dynamics and Control 27, 717-769.
Heston, S., 1993, A closed form solution for options with stochastic volatility with
applications to bond and currency options. Review of Financial Studies 6, 327-
343.
Hibbert, A., Daigler, R., and Dupoyet, B., 2008. A behavioural explanation for the
negative asymmetric return-volatility relation. Journal of Banking and Finance
32, 2254-2266.
Huisman, R., Koedijk, K., and Pownall, R., 1998. VaR-x: Fat tails in financial risk
management. Journal of Risk 1, 47-61.
Jiang, G., and Tian, Y., 2005. The model-free implied volatility and its information
contents. Review of Financial Studies 18, 1305-1342.
Jorion, P., 2000. Value at Risk: the benchmark for controlling market risk. Second
edition, McGraw-Hill, New York, United States.
Kerkhof, J., and Pelsser, A., 2002. Observational equivalence of discrete string models
and market models. Journal of Derivatives 10, 55-61.
Koenker, R., 2005. Quantile regressions. First edition, Cambridge University Press,
New York, United States.
Koenker, R., Bassett, G., 1978. Regression quantiles. Econometrica 46, 33-50.
Konstantinidi, E., Skiadopoulos, G., and Tzagkaraki, E., 2008. Can the evolution of
implied volatility be forecasted? Evidence from European and US implied
volatility indices. Journal of Banking and Finance 32, 2401-2411.
Leland, H., 1985. Option pricing and replications with transaction costs. Journal of
Finance 40, 1283-1301.
Liu, J., Pan, J., and Wang, T., 2005. An equilibrium model of rare-event premia and its
implication for option smirks. Review of Financial Studies 18, 131-164.
London, J., 2004. Modeling derivatives in C++. First edition, Wiley Finance, New
Jersey, United States.
Longstaff, F., Santa-Clara, P., and Schwartz, E., 2001. The relative valuation of caps and
swaptions: Theory and empirical evidence. Journal of Finance 56, 2067-2109.
Low, C., 2004. The fear and exuberance from implied volatility of S&P 100 index
options. Journal of Business 77, 527-546.
Martens, M., and Zein, J., 2004. Predicting financial volatility: High-frequency time-
series forecasts vis-à-vis implied volatility. Journal of Futures Markets, 24,
1005-1028.
Mayhew, S., and Stivers, C., 2003. Stock return dynamics, option volume, and the
information content of implied volatility. Journal of Futures Markets, 23, 615-
646.
Meligkotsidou, L., Vrontos, I., and Vrontos, S., 2009. Quantile regression analysis of
hedge fund strategies. Journal of Empirical Finance 16, 264-279.
Merton, R., 1973. Theory of rational option pricing. Bell Journal of Economics and
Management Sciences 4, 141-183.
Mixon, S., 2002. Factors explaining movements in the implied volatility surface.
Journal of Futures Markets 22, 915-937.
 37

Pan, J., 2002. The jump-risk premia implicit in options: evidence from an integrated
time series study. Journal of Financial Economics 63, 3-50.
Persaran, M., and Shin, Y., 1998. Impulse response analysis in linear multivariate
models. Economics Letters 58, 17-29
Poon, S.H., and Granger, C.W.J., 2003. Forecasting volatility in financial markets: A
review. Journal of Economic Literature 41, 478-539.
Poon, S., and Granger, C., 2005. Practical issues in forecasting volatility. Financial
Analyst Journal, 61, 45-56.
Poterba, J., and Summers, L., 1986. The persistence of volatility and stock market
fluctuations. American Economic Review 76, 1142-1151.
Rubinstein, M., 1994. Implied binomial trees. Journal of Finance 49, 771-818.
Santa-Clara, P., and Sornette, D., 2001. The dynamics of the forward interest rate curve
with stochastic string shocks. Review of Financial Studies 14, 149-185.
Schwert, G.W., 1989. Why does stock market volatility change over time? Journal of
Finance 44, 1115-1153.
Schwert, G.W., 1990. Stock volatility and the crash of “87.” The Review of Financial
Studies 3, 77-102.
Simon, D., 2003. The Nasdaq volatility index during and after the bubble. Journal of
Derivatives 11, 9-24.
Sims, C.A., 1980. Macroeconomics and reality. Econometrica, 48, 1-48.
Skiadopoulos, G., 2001. Volatility smile consistent option models: A survey.
International Journal of Theoretical and Applied Finance 4, 403-437.
Skiadopoulos, G., 2004. The Greek implied volatility index: construction and
properties. Applied Financial Economics 14, 1187-1196.
Skiadopoulos, G., Hodges, S., and Clewlow, L., 1999. The dynamics of the S&P500
implied volatility surface. Review of Derivatives Research 3, 263-282.
Whaley, R., 2000. The investor fear gauge. Journal of Portfolio Management 26, 12-17.
Zhang, J., and Xiang, Y., 2008. The implied volatility smirk. Quantitative Finance 8,
263-284.
38
 39

PART II: THE ESSAYS

40
 41

Modeling The Dynamics of Implied Volatility Surfaces

Ihsan Ullah Badshah

Hanken School of Economics, Department of Finance and Statistics, P.O. Box 287,
FIN-65101 Vaasa, Finland. Phone: +358-6-3533 721, Fax: +358-6-3533 703,
Email: ibadshah@hanken.fi

September 20, 2010

Abstract

The purpose of this study is to model implied volatility surfaces and identify risk factors
that account for most of the randomness in these volatility surfaces. The approach is
similar to that of the study by Dumas, Fleming and Whaley (DFW) (1998). We use
moneyness in the implied forward price and out-of-the-money put-call options on the
FTSE 100 stock index. After adjustments, a nonlinear parametric optimization
technique is used to estimate different DFW models to characterize and produce
smooth implied volatility surfaces. Next, principal component analysis is applied to the
implied volatility surfaces to extract principal components that account for most of the
dynamics in the shape of the surfaces. We then estimate and obtain smooth implied
volatility surfaces with the parametric models that account for both smirk (skew) and
time to maturity. We find that the constant-volatility model fails to explain the
variations in the surfaces. However, the first three principal components (or factors)
can explain about 69-88% of the variance in the implied volatility surfaces, in which, on
average, 56% is explained by the level factor, 15% is explained by the term-structure
factor, and the remaining 7% is explained by the jump-fear factor. Applications of our
study include options trading, hedging of derivatives positions, risk management of
options, and policymaking.

Keywords: implied volatility, implied volatility surface, options, principal component

analysis, smirk

JEL classification: C13, C53, G12, G13

The Author would like to thank Johan Knif, Mika Vaihekoski, George Skiadopoulos, Gregory Koutmos,
Kenneth Högholm, David Simon, and Yakup Arisoy for providing useful comments. The author would also
like to thank to the participants of the EFMA conference (2008), Athens, Greece; the Global Finance
Conference (2007), Melbourne, Australia; and the Graduate School of Finance (GSF) Research workshops
for their useful comments. The author acknowledges Evald & Hilda Nissi and HANKEN foundations for
providing financial support.
 42

1 Introduction

Modeling and exploring the dynamics of the volatility of financial assets has been an
active area of research in quantitative finance, for both academics and practitioners.
Volatility is fundamental for risk management, option pricing, option trading, hedging
derivative positions, constructing and diversifying portfolios and policy making. Most
previous research has focused on historical volatility (HV), but implied volatility (IV)
has recently received attention due to some breakthrough studies, such as those of
Christensen and Prabhala (1998), Fleming (1998), Skiadopoulos et al. (1999), Dumas et
al. (1998), Blair et al. (2001), Ederington and Guan (2002), Poon and Granger (2003)
and Gocalves and Guidolin (2006). They have shown that implied information is
superior to historical information when forecasting volatility and thereby suggest that a
precise view of the market's judgment and expectation of volatility is vital for
forecasting volatility. HV is backward looking and incorporates expectations about
future volatility based on the past behavior of stock prices and other relevant
information. In contrast, IV is forward looking; that is, it is implied by the market
option prices. Option prices are the common consensus of market participants
regarding the expected future volatility of an underlying asset over the remaining life of
an option; usually, the majority of option markets’ participants are institutional
investors who can make professional judgments on the future direction of volatility.
The volatility expectation of market participants can be recovered by inverting the
option-pricing model. However, since the 1987 stock market crash, one of the strongest
empirical regularities in the stock index options has been that IV recovered from stock
index options has appeared to differ between moneyness (strike dimension), the
volatility smirk (skew), and term structure (time-to-maturity dimension), the volatility
term structure if the two are examined at the same time, for example, on the same day.
This indicates that implied volatilities (IVs) present a rich implied volatility surface
(IVS), whose dynamics across strikes and over time warrant further investigation.

The Black and Scholes (BS) (1973) model assumes that underlying assets follow
geometric Brownian motion and constant volatility, implying that options on the same
underlying asset with different strike prices and times to maturity should have the
same IV. In fact, the assumption of constant volatility does not hold empirically
because asset prices are mostly influenced by risk factors such as jumps, stochastic
volatility or time-varying volatility, the demand/supply pressure of options, and
transaction costs (see, e.g., Carr et al., 2001; Heston, 1993; Leland 1985; David and
 43

Veronesi, 2000; Garcia et al., 2003; and Bollen and Whaley, 2004). To account for
these deviations, practitioners use different IVs for different strikes and maturities.
Therefore, the IV of an option reflects the determinants of the option value that are not
captured by the constant volatility BS model – this volatility structure illustrates
discrepancies between theoretical prices and the market. Therefore, an IVS that
includes all of these features is essential in practice, particularly for forecasting future
volatility, trading options, pricing illiquid options, pricing and hedging exotic
derivatives, and managing the risk of the options portfolio.

Furthermore, it is important for practitioners to generate IVS from the IVs implied by
options that contain the majority of market sentiments (beliefs on bearish and bullish
markets). Therefore, implying volatility from the out-of-the-money (OTM) put and call
options in addition to at-the-money (ATM) options is of paramount importance to any
volatility measure or IVS. OTM put options incorporate beliefs on the market crashes,
and OTM call options incorporate beliefs on market spikes (see, e.g., Pan, 2002; Liu et
al., 2005; Doran et al., 2007; Bates, 2008; Camara and Heston, 2008). Because we
know that investors (particularly institutional investors) always hedge their stock index
portfolios with put positions, during market declines, there is greater demand for OTM
puts (with less supply) than OTM calls (with less demand), leading to higher IVs (see,
e.g., Bollen and Whaley, 2004). Therefore, it is essential that IVS be generated (or
forecasted) using IVs from both OTM puts and calls; these IVs contain a broader set of
information that includes information on future negative and positive jumps.

However, the IVS can be viewed as highly correlated because IVs present a high degree
of correlation among risk factors, indicating high dependence. When few important
sources of information are common to the risk factors, we find a high degree of
correlation among risk factors. Principal component analysis (PCA) is a tool that
extracts the most important independent risk factors from the correlated IVS and can
explain most of the dynamics. Therefore, PCA reduces the dimensionality and enhances
computational efficiency. After extracting common risk factors from the IVs, these can
be used for a variety of purposes, such as projecting IVS, Vega hedging, Value-at-Risk
modeling, and model calibration.

There are few studies that estimate and obtain IVS and further study the dynamics of
these IVSs by applying PCA. However, many studies have examined either the volatility
smirk/skew/smile (strike dimension) or the volatility term structure. One well-known
 44

study on IVS was conducted by Cont and Da Fonseca (2002), who examined the
dynamics of IVSs of S&P 500 and FTSE 100 index options. The price of an index option
on a given date is represented by a corresponding IVS, which has volatility smirk (skew)
and term structure features. They used different methods to obtain smooth IVSs.
However, the main findings were, first, that an IVS has a non-flat shape in both
dimensions (i.e., strike and time). Second, the shape of the IVS changes with time.
Third, IVs present positive autocorrelation, indicating the property of mean reversion.
Fourth, variations in IVS can be explained by two or three principal components (PCs).
Finally, the movements in the underlying assets are not correlated with movements in
the IVs. In another study on IVS, le Roux (2006) developed an econometric model to
estimate IVSs and then examined the dynamics of the estimated IVS using PCA. He
used VIX index data and options on the S&P500 index and found that 75.2% of the
variations of the IVS can be explained by the first PC and that another 15.6% can be
explained by the second PC. Moreover, he found that processes for these factors appear
to be independent of the process of the VIX index. Alentorn (2004) adapted a
parametric form of the DFW (1998) approach and estimated IVS for options on the
FTSE 100 futures. He used three DFW models. He further suggests how to implement
this methodology in real time. This study differs from the above two in that it only
estimates IVS, whereas the other two additionally study the dynamics of the IVS.

Many studies have been conducted on the dynamics of either smirk/skew/smile or

surfaces using PCA, such as that by Skiadopoulos, Hodges, and Clewlow (SHC) (1999),
who investigated the factors and shape of shocks that move IV smiles and IVS of
S&P500 index options. They formed maturity buckets within the IVS, the average IVs
of options whose maturities fall into them, and applied PCA to each bucket separately.
They identified two factors that explain about 60% of the variance. They suggest that to
effectively price hedge future options, one needs only three risk factors: one for the
underlying asset and the other two for IV. In another study, Alexander (2001)
developed a new PCA model of fixed-strike volatility deviations from at-the-money
(ATM) volatility. She focused on its application to the volatility skew and used FTSE
100 index options with a maximum of three-month maturities. She investigated
whether the IVS would move continuously if the index moved and, furthermore,
whether second and higher PCs with non-zero conditional correlation with the index
changes cause non-parallel movements in the IVS as the index moves. She found that
most of the variation in the IV of FTSE 100 index options can be explained by just three
 45

risk factors: parallel shifts, tilts, and curvature changes. However, Alexander (2001)
investigated only skew, whereas SCH (1999) investigated both skew and surface.

The objective of this essay is to estimate and obtain smooth IVSs using parametric
models and to study the dynamics of these IVSs. Our study is related to that of DFW
(1998) in terms of functional forms of estimated models for different IVSs, whereas the
second part of our study is related to that of SHC (1999) in terms of applying PCA to
IVSs. Our study, however, differs in some ways from the two cited studies: first, we
estimate IVS using OTM puts for low strikes and OTM calls for high strikes in addition
to ATM options, which are not accounted for in DFW (1998). Second, we study IVSs
using moneyness in the forward price and time to maturity, whereas DFW (1998)
incorporated the strike instead. Third, we employ PCA to extract those risk factors that
explain most of the dynamics in the IVSs to determine what these factors look like, how
many factors can explain these IVSs, and their interpretation. Finally, we study the
dynamics of the IVSs of FTSE 100 index options (i.e., over the most recent four years),
whereas SHC (1999) studied futures options on the S&P500 index.

Our main findings are that the DFW constant volatility model fails to capture the
variations in the IVSs and, consequently, generates only flat IVSs. Model 2, which
captures the volatility skew and volatility smirk effect, does a good job and captures
most of the volatility skew and smirk effects. When we estimated Models 3 and 4, which
account for both the volatility term structure and volatility smirk/skew effects; these
models performed better than the one-dimensional smirk/skew model. However, both
models fit the market IVS well and generate the most appropriate IVSs for the observed
data. In the second part of the study, which uses PCA (nonparametric approach), we
find that the first three factors can explain about 69-88% of the variances in the IVS:
within this, on average, 56% is explained by the level factor, 15% is explained by the
term-structure factor, and an additional is explained 7% by the jump-fear factor.
Finally, when applying PCA to the maturity groups, we find that level factor (or parallel
shifts) is evident and thus important for shorter- and longer-maturity IVs. However,
the term-structure factor is important for the IVs with medium maturities, and the
jump-fear factor is important for the IVs with the shortest (8-90 days) maturities.

This essay is organized as follows. Section 2 describes the methodology. Section 3

describes our data. Section 4 presents results, and Section 5 concludes our study.
 46

2 Methodology
2.1 Model for Implied Volatilities

Merton (1973) extended the Black-Scholes (1973) model for European call and put
options on dividend-paying S t stock or stock index with a strike price of K , a time to

maturity of  T  t , a call with a payoff of Max(S t  K,0) , and a put with a payoff of

Max (K  S t ,0) . The BS formulae for the values of these call and put options are

C S t ,K,r,K, S t e  q  d 1   Ke  r  d 2 , (1)

P S t ,K,r,K, Ke  r   d 2   S t e  q   d 1 , (2)

d1

lnS t K    r  q   2 2 , (3)
 

d2

lnS t K    r  q   2 2 . (4)
 
where q is the dividend rate on the underlying asset, r is the risk-free interest rate, and
( ˜ ) stands for the standard normal cumulative density. Consider an option market in
which the BS assumptions do not hold empirically (see the empirical evidence of Bates,
1991; Rubinstein, 1994; Ait-Sahalia and Lo, 1998); then, the quoted market prices of
~ ~
BS
put and call options are C t and P t . Thus, there is a unique implied volatility,  t ,

which equates the theoretical option price with the market option price given all other
observed parameters:

~
C BS (S t ,K,r,K,tBS ) Ct , (5)
~
PBS (S t ,K,r,K,tBS ) Pt . (6)

The BSIV can be found uniquely because of the monotonicity of the BS formula in the
volatility parameter,

wBS
!0. (7)
w
 47

For a fixed value of (K,T) ,  tBS (K,T) is a stochastic process, and for a fixed value of t,

its value depends on the characteristics of the option such as maturity T and strike K.40
Therefore, the function is

 tBS :(K,T) o  tBS (K,T). (8)

This is the implied volatility surface at date t. However, with moneyness of the option,
the IVS is a function of moneyness and time-to-maturity, that is,

I t m,   tBS (m(S(t),t  ) . (9)

Nonetheless, with the help of the bisection method, IVs are obtained for each strike
price and time to maturity, that is, for the OTM put and call as well as for ATM options.
The bisection method is efficient and fast with the advantage that IVs can be estimated
without knowing Vega.

2.2 Modeling Implied Volatility Surfaces

To model and obtain smooth implied volatility surfaces, we prefer a parametric

approach similar to the approach proposed by Dumas et al. (1998). This approach is
also known as the Practitioner Black and Scholes model (PBS) (see, e.g., Christoffersen
and Jacobs, 2004; Goncalves and Guidolin, 2006). Its parsimony and application
motivated us to study it thoroughly and extend it further, particularly to fit PBS to the
FTSE100 index options’ information. However, we approach the problem somewhat
differently than the original work of DFW (1998): first, we use the most informed and
liquid options besides ATM, which are the OTM puts and calls. Because the put option
is the natural hedging instrument for the underlying stock index portfolio, institutional
investors always hedge their portfolios’ downside risk with OTM put options; this trend
is obvious when they forecast any market uncertainty (see, e.g., Pan, 2002; Bollen and
Whaley, 2004; Foresi and Wu, 2005; Bates, 2008; Camara and Heston, 2008).41
Second, we obtain IVS with more practitioner-friendly moneyness, which takes into
account forward prices and time dependence, because, for the informed IVS, daily
calibration is essential to the changes in stock price and IV. The IV as a function of

40 See, for a detailed discussion, the study of Cont and da Fonseca (2002).
41However, when they forecast any market spikes in the future they go long in the OTM calls with limited
downside risks, which remain to the premium of the options only, instead of taking positions in the
underlying stock index.
 48

strike does not adequately capture the market movements, whereas IV as a function of
the moneyness parameter does. Third, we use FTSE 100 index options data, whereas
DFW (1998) used S&P 500 index options data. Fourth, we use a parametric
optimization technique to estimate an IVS that is common in practice.

Fung and Hsieh (1991) and Jackwerth and Rubenstein (1996) were the first to propose
a simple moneyness measure, that is, to take the ratio of the strike price and stock
price. However, rather than absolute moneyness, practitioners express moneyness
based on the log of the implied forward price versus the strike. For a strike, K , and an
r 
implied forward price, Se t , where rt is the interest rate until the maturity of the

option, and  T  t , is the time to maturity. However, this moneyness measure is

time independent. Time dependence is important when updating IVS. There are many
measures in the market that account for different information in the market (on the
implied forward price moneyness measures, see, e.g., Foresi and Wu, 2005; Goncalves
and Guidolin, 2006; Tompkins, 2001; Alentorn, 2004). However, we use a similar
moneyness in implied forward price as proposed by Gross and Waltner (1995).
Moneyness in implied forward price, m, is defined as the logarithm of the implied
r 
forward price, Se t , over the strike price, K, normalized by the square root of the time
to maturity,  T  t , that is,

log(S t e r  t /K)
m . (10)

Four different structural forms of model specifications are estimated, leading to smooth
IVSs, each of which differs in the parametric structure that serves to characterize the
IVs:

Model 1: I m,  0   , (11)

Model 2: I m,  0  1 m   2 m 2  , (12)

Model 3: I m,  0  1 m   2 m 2   3    4  m  , (13)

Model 4: I m,   0   1 m   2 m 2   3    4  m   5  2  . (14)

In Model 1, the flat IV function is a constant, that is, the BS model that gives volatility
equal to ATMIV does not account for volatility skew, smirk or term-structure effects.
 49

Model 2 attempts to capture volatility skew and volatility smirk across moneyness;
here, the slope, 1 , represents volatility skew, and the curvature, 2 , represents
volatility smirk effects.42 Model 3 captures additional variation attributable to time
(e.g., time to maturity slope),  3 , and a combined effect of the skew and time

dimension,  4 . Finally, Model 4 adds a parameter that captures the time-to-maturity

curvature effect,  5 . The remaining notations are as follows: I m,  represents the

dependence of IV on the moneyness level and time to maturity, and 0 is a parameter

that is constant for all models.

The four structural models were estimated and smooth IVSs were obtained using the
nonlinear optimization technique. In fact, the choice of the nonlinear optimization
technique is obvious for the following reasons: first, we chose nonlinear optimization
because this is more common in market practice, particularly for multifactor models
such as the well-known LIBOR market models. Second, optimization is more flexible.
Third, the parameters need to be continuously calibrated to the market information.
Fourth, this can capture the nonlinear effects nicely. Therefore, we optimize the IVS
with a nonlinear least square function (lsqnonlin) in MATLAB.43

However, to test for the goodness of fit of the models and for comparison, we use two
important loss functions proposed by Christoffersen and Jacobs (2004): relative root
mean squared errors (%RMSE) and IV root mean squared errors (IVRMSE). The
%RMSE loss function is as follows:

¦ §¨© e  OptionPric e ·¸¹

N 2
%RMSE   1
N
i . (15)
i 1 i

The %RMSE loss function minimizes the relative difference between the market IVs
(i.e., BSIVs) and PBM model IVs. Therefore, %RMSE assigns a greater weight to the
deep OTM options because these values are small. For instance, we can see that the
denominator of the %RMSE loss function includes the market price of the option.
Therefore, when the market price of the option is small, the difference between the IVs

42In Model 2, when the slope is zero, it becomes a pure smile.

43This is a built-in function. It follows an algorithm that minimizes the sum of squared errors between the
actual value and the prediction for a given vector of a parameter.
 50

is amplified. However, the IVRMSE loss function assigns equal weights to IVs, as
follows:
N
IVRMSE  ¦    
1 2
N i i . (16)
i 1

Therefore, the IVRMSE loss function minimizes the error between the market IVs and
PBS model IVs.44

2.3 Principal Component Analysis of Implied Volatility Surfaces

In addition to the parametric approach for modeling the IVS, we proceed to examine
the dynamics of the IVS using the principal component analysis (PCA) approach.45 PCA
is used to reduce the IVS dimensions to only a few uncorrelated underlying risk factors
that drive the IVS, the rationale is to extract independent factors that explain most of
the dynamics of IVS. With PCA, independent factors can be identified without any prior
assumptions. Many researchers have attempted to extract those risk factors that
explain most of the variation in the volatility smirk (skew) or surface; Alexander (2001)
and Cont and Da Fonseca (2002) (FTSE 100 index IV); Cont and Da Fonseca (2002)
and Skiadopoulos et al.(1999) (S&P 500 index IV); and Fengler et al. (2003) (DAX
index IV). Their findings confirmed that about 70-90% of the total variations in the IVS
or skew can be attributed to just three factors: parallel shift, tilt, and curvature, which
are changes that are captured by the first three PCs. Therefore, our approach is similar
to that of Skiadopoulos et al. (1999) but our data differ. We used FTSE 100 index IV
data, whereas they used S&P500 index IV data.

We compute the first-difference IVs across different moneyness levels and times to
maturity because the data input to PCA must be stationary, and IVs are often
nonstationary (for more discussion on the data issues, see, e.g., Skiadopoulos et al.,
1999). PCA was applied to pool data of IVs to analyze the IVS dynamics. Then, the data
were assembled into different groups based on the number of trading days left to
maturity, that is, 8-30, 30-60, 60-90, 90-150, 150-220, and 220 days and above. After

44 See Rouah and Vainberg (2007) for a nice detail discussion on the above loss functions.
45 However, recently Fengler et al. (2007) proposed a modified method to the PCA, that is a
semiparametric approach of modeling the dynamics of implied volatility surface, their approach might be
advantageous in modeling the dynamics, however, we believe that our approach is simple, parsimonious
and widely acknowledged in practice as well previously employed by many studies such as Skiadopoulos et
al. (1999) etc.
 51

making these groupings, PCA was applied to each group to investigate the IVS
dynamics in greater detail.

PCA is a matrix decomposition method that decomposes a matrix into eigenvectors and
eigenvalue matrices. It was applied to a pool of IVs across moneyness levels and term
T
structures to decompose the covariance matrix into  , where the diagonal
elements of  are the eigenvalues and the columns  are the associated eigenvectors.
The choice of column labeling in  allows PC ordering such that e1 belongs to the

largest eigenvalue 1 , e2 belongs to the second-largest eigenvalue, 2 , and so forth. In

a highly correlated IVS, the first eigenvalue would be much larger than the others.
Consequently, the first PC can explain much of the variation in the IVS. If the first three
PCs explain most of the variations in the IVS, then these PCs can replace the original IV
variables without losing much information (see, e.g., Skiadopoulos et al., 1999; London,
2004). That is, the original IV input data may be written as a linear combination of the
PCs, which reduces the dimensions of the system.

The first three PCs or factors can be interpreted in this study as follows: the first factor
(or PC1), which is common to the IVs and moves the entire IVS systematically in the
same direction, can be interpreted as the volatility level factor (or parallel shifts). It is
important for IVs implied from the shortest and longest maturity options. The second
factor (or PC2) can be interpreted as the term-structure factor (or tilt, which generates
shifts in the slope of the term structure of IVs), which affects the shorter- and longer-
maturity IVs differently (i.e., the sign of the effects differ for the shorter and longer
maturities IVs). The third factor (or PC3) can be interpreted as the jump-fear factor (or
curvature, which changes the steepness of IV smirk/skew); the effects of the jump-fear
factor on the OTM puts and calls significantly differ, which induces steepness in the
volatility smirk (skew). For empirical evidence on the dynamics of the factors driving
the IVS of index options, see, e.g., the studies by Skiadopoulos et al. (1999), Mixon
(2002), and Cont and da Fonseca (2002).
 52

3 Data

We used daily data of European options on the FTSE 100 stock index. The data were
extracted from the Euronext database for the period from January 01, 2004 to
December 31, 2007. The end-of-the-day prices were considered and were updated daily
in the Euronext database. The database includes the following daily information for
each call and put traded option: the type (call/put), date, delivery date, strike price,
volume, and open interest; opening, high, low, and closing option prices; and
underlying price. First, we used option data; the maturity ranged from eight days up to
one year and options varied across different strike prices. Second, data on the discount
curve, which had different maturities to proxy for a risk-free rate downloaded from
Thomson DataStream. Finally, dividend data; the dividend yield for the FTSE 100 stock
index was backed out from FTSE 100 Futures prices, using the cost-of-carry approach
(See e.g., Hull, 2002), the futures data were extracted from the same Euronext
database. We considered data for ATM options and OTM put and call options: OTM
puts were used for low strikes, and OTM calls were used for high strikes. However, the
data on in-the-money (ITM) puts and calls were not considered because they are
sensitive to the non-synchronicity problem. Moreover, options with fewer than eight
days to maturity were also excluded because they are very sensitive to small errors in
the option price (e.g., Skiadopoulos et al., 1999).

Figure 1 shows the scatter graph of the FTSE 100 index (level) against the ATMIV
(level) of short maturity (i.e., one month to maturity) – the FTSE 100 stock index level
is on the vertical axis, and the ATMIV level is on the horizontal axis. We observe a
noticeable negative relationship between the two historical series. This evidence is
consistent with previous research, which has found similar evidence that the stock
returns are negatively correlated with changes in IVs.
 53

FTSE100 Index versus ATM Implied Volatility

6,800

6,400
FTSE100 Index(Level)

6,000

5,600

5,200

4,800

4,400

4,000
8 10 12 14 16 18 20 22 24 26 28

ATM Implied Volatility(Level)

Figure 1. A scatter plot of FTSE100 Index (level) versus ATMIV (Level) from January
1, 2004, to December 31, 2007.

Table 1 presents the descriptive statistics for the ATMIVs and the FTSE 100 index from
January 1, 2004 to December 31, 2007. Panel I of Table 1 shows the descriptive
statistics for the level. The mean ATM volatility was 12.8% annually, with a maximum
of 27.2% and a minimum of 8%, whereas for the FTSE 100 index, the mean level was
5,501 points, with a maximum level of 6,732 points and a minimum level of 4,287
points. However, Panel II of Table 1 reports the descriptive statistics on the daily
percentage changes () in the one-month ATMIVs and the daily percentage
continuously compounded returns of the FTSE 100 stock index. We can observe from
the skewness that ATMIVs are positively skewed, whereas the FTSE 100 index returns
are negatively skewed; however, the kurtosis for both ATMIVs and FTSE 100 returns
is high (i.e., greater than the normal three). The Jarque-Bera tests for overall normality
reject that both ATMIVs and FTSE 100 returns are normally distributed. Therefore,
both ATMIVs and FTSE100 index returns are non-normally distributed. Some
researchers have documented that non-normality in the underlying should trigger non-
constant IVs. Panel III of Table 1 shows a correlation between the FTSE 100 index and
 54

one-month ATMIVs. The correlation between the FTSE 100 index returns and
ATMIVs is about -0.83, which is significantly negative and therefore reconciles with
the results of Figure 1.

Table 1. The descriptive statistics for the daily FTSE 100 index levels (and returns) and daily
levels (and changes) in the ATMIVs of the FTSE 100 index from January 1, 2004 to December
31, 2007.
Panel I: Daily ATM implied Volatility (level) and Daily FTSE 100 Index (level)
ATM Implied Volatility FTSE100
Mean 12.865 5501.783
Median 11.995 5595.450
Minimum 8.000 4287.040
Maximum 27.260 6732.400
Panel II: Daily percentage changes of ATM implied Volatility and percentage returns of FTSE
ATM Implied Volatility FTSE100
Mean 0.007812 0.038537
Median -0.02000 0.049429
Minimum -3.98000 -4.098649
Maximum 4.61000 3.504223
Std. Dev. 0.795208 0.786640
Skewness 0.878056 -0.373138
Kurtosis 8.834837 5.880725
J-Bera
1612.029 384.4766
Probability
0.000000 0.00000
Panel III: Correlation Analysis between ATM implied Volatility and FTSE100 index
ATM Implied Volatility FTSE100
ATM Implied Volatility 1.0000 -0.833891**
---------- (-48.7239)
FTSE 100(Returns) -0.833891** 1.0000
(-48.7239) ---------
T-statistics are in the parenthesis. ** Denote rejection of the null hypothesis at the 1% significance levels.

Table 2 presents descriptive statistics of the IVs recovered from the OTM put and call
options’ weekly data from 2004-2007 for the months of March and October. The mean
IV (level) is about 18% annually, whereas the mean percentage change () of the IVs is -
0.46. The maximum IV (level) is 41%, with a minimum of 7.5%. However, the
 55

maximum IV is 2.265, with a minimum of -6.058. Similarly, both IV (level) and IV
are non-normally distributed as well, which is confirmed by the Jarque-Bera statistics.

Table 2. The descriptive statistics for the crossection of weekly implied volatilities recovered
from the out-of-the-money put and call options from 2004-2007 for the months of March and
October. The descriptive statistics for both levels and changes in IVs are reported.
OTM-PutCall Implied Volatility(Level) OTM-PutCall Implied Volatility

Mean 18.45401 -0.465341

Median 17.47506 -0.448484
Minimum 7.562080 -6.058008
Maximum 41.33367 2.265540
St.Dev. 6.856906 0.407839
Skewness 0.607074 -0.549265
Kurtosis 2.705163 10.64076
J-Bera 497.9209** 19006.07**
T-statistics are in the parenthesis. ** Denote rejection of the null hypothesis at the 1% significance levels.

In Figure 2 the IVs recovered from OTM puts and calls are plotted against moneyness
levels from 2004 to 2007 for the months of March and October. An almost one-to-one
relationship between the IVs and moneyness levels exists--most of the data lie at a 45-
degree angle. This evidence supports the monotonic relationship between IV and the
option price; that is as volatility increases, the option price increases.
The data presents some important patterns consistent with previous empirical findings.
First, the underlying stock index returns are negatively correlated with the changes in
IVs. Second, IVs backed out from option prices with different strike prices and
maturities on the same underlying stock index differ considerably; therefore, the
assumption of BS model of constant volatility is rejected. Third, there is a one-to-one
relationship between volatility and moneyness that implies that option prices have a
monotonic relationship with volatility. This relationship has important implications for
option traders.
 56

OTM-PutCall implied volatility versus Moneyness

44

40

36
OTM-PutCall implied volatility (%)

32

28

24

20

16

12

4
-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

Moneyness

Figure 2. A scatter plot of IVs (level) implied by OTM Put and Call options versus
Moneyness (level).
 57

4 Empirical Results
4.1 Results of the Four Estimated Models of Volatility Surfaces

Table 3 reports the results of the estimated implied volatility surfaces in the months of
March and October for the years 2004, 2005, 2006, and 2007. These surfaces were
estimated with four parametric models using a nonlinear optimization technique. Panel
I of Table 3 shows the results for Model 1, the constant volatility model. The goodness
of fit criteria IVRMSE and %RMSE range from a minimum of 0.046 and 0.046 to a
maximum of 0.078 and 0.069, respectively, indicating very high levels. Figure 3 is a
visual display of the average IVS (March, 2004) (pink dots), which corresponds to
Model 1 projection (green). As shown in Figure 3, a rich market IVS exists, and the IVS
projected by the Model 1 (or BS Model) is flat; therefore, IV is constant at every point of
the IVS. We use this constant volatility model later as a reference model for comparison
purposes.

Figure 3. Estimated implied volatility surface with Model 1 for March 2004.
 58

Table 3. Results for estimated implied volatility surfaces (2004-2007)

Panel-I: Model 1 I m,t  0  

Month 0 IVRMSE %RMSE

Mar-2004 0.2082 0.0775 0.0640

Mar-2005 0.1468 0.0562 0.0639
Mar-2006 0.1529 0.0458 0.0501
Mar-2007 0.1849 0.0577 0.0465
Oct-2004 0.1752 0.0698 0.0666
Oct-2005 0.1633 0.0627 0.0689
Oct-2006 0.1631 0.0493 0.0461
Oct-2007 0.2262 0.0635 0.0468
Panel-II: Model 2 I m,   0  1m   2m 2
 
Month 0 1 2 IVRMSE %RMSE

Mar-2004 0.1674 0.2156 0.0077 0.0250 0.0174

Mar-2005 0.1098 0.1671 0.0794 0.0123 0.0110
Mar-2006 0.1221 0.1466 0.0669 0.0110 0.0080
Mar-2007 0.1515 0.2184 -0.0454 0.0180 0.0144
Oct-2004 0.1309 0.1850 0.0679 0.0164 0.0159
Oct-2005 0.1256 0.1810 0.0614 0.0141 0.0119
Oct-2006 0.1285 0.1888 0.0006 0.0105 0.0073
Oct-2007 0.1877 0.2089 -0.0053 0.0156 0.0104
Panel-III: Model 3 I m,   0  1m   2m 2
 3  4m  

Month 0 1 2 3 4 IVRMSE %RMSE

Mar-2004 0.1570 0.1756 0.0362 0.0186 0.0827 0.0221 0.0152

Mar-2005 0.1024 0.1535 0.0787 0.0163 0.0384 0.0100 0.0112
Mar-2006 0.1127 0.1187 0.0937 0.0185 0.0441 0.0068 0.0063
Mar-2007 0.1611 0.2150 -0.0513 -0.0195 0.0086 0.0170 0.0132
Oct-2004 0.1223 0.1600 0.0837 0.0182 0.0501 0.0139 0.0148
Oct-2005 0.1218 0.1491 0.0854 0.0073 0.0858 0.0126 0.0106
Oct-2006 0.1200 0.1716 0.0200 0.0198 0.0245 0.0076 0.0067
Oct-2007 0.1885 0.1938 0.0016 -0.0035 0.0355 0.0153 0.0097
Panel-IV: Model 4 I m,   0  1m   2 m 2
 3  4m  5 2
 

Month 0 1 2 3 4 5 IVRMSE %RMSE

Mar-2004 0.1438 0.1818 0.0319 0.0938 0.0739 -0.066 0.0214 0.0155

Mar-2005 0.0997 0.1552 0.0772 0.0313 0.0345 -0.0130 0.0100 0.0115
Mar-2006 0.1137 0.1185 0.0941 0.0133 0.0442 0.0046 0.0068 0.0063
Mar-2007 0.1745 0.2100 -0.0476 -0.0948 0.0218 0.0652 0.0161 0.0122
Oct-2004 0.1271 0.1588 0.0838 -0.0092 0.0525 0.0258 0.0138 0.0144
Oct-2005 0.1192 0.1503 0.0853 0.0230 0.0823 -0.0152 0.0126 0.0106
Oct-2006 0.1194 0.1717 0.0198 0.0230 0.0243 -0.0031 0.0076 0.0068
Oct-2007 0.1845 0.1947 0.0013 0.0218 0.0342 -0.024 0.0152 0.0099
 59

Panel II of Table 3 displays the results for Model 2, which attempts to capture volatility
skew and volatility smirk across moneyness, the quadratic skew model, which includes
additional parameters for moneyness slope (captures volatility skew) and curvature
(captures volatility smirk).46 As is evident from the goodness-of-fit criteria, IVRMSE
and %RMSE decreased for all six months, ranging from a minimum of 0.010 and 0.007
to a maximum of 0.025 and 0.017, respectively, and improved by 0.036 to 0.053 and
0.039 to 0.050 points, respectively, compared with Model 1. Figure 4 shows the
corresponding 3D graph for the March 2004 IVS. As shown, the pink dots present the
observed term structure of the smirk/skew (also called strings), that is, skewed in the
moneyness level, and show the decline in the term structure. Figure 4 demonstrates
two important findings. First, in addition to the volatility level effect (ATMIV captured
by the Model 1), there are strong volatility skew and smirk effects (captured by the
slope and curvature parameters); therefore, model 2 fits well into the observed rich
IVS. Second, the smirk (skew) IVS of Model 2 is contrary to the BS model’s assumption
of constant volatility and therefore suggests the existence of stylized facts regarding IV.
Third, we observe high volatilities for options with shorter maturities than options with
longer maturities. Fourth, the level and volatility of moneyness have a monotonic
relationship, which is consistent with the findings of previous research.

Figure 4. Estimated implied volatility surface with Model 2 for March 2004.

46 The volatility smirk can be also interpreted a fear factor in the market. This volatility smirk pattern

arises when OTM put options are more expensive than OTM call options (see Foresi and Wu, 2000; Zhang
and Xiang, 2008; Carverhill et al., 2009).
 60

Panel III of Table 3 reports results for IVSs estimated with Model 3, which is Model 2
with additional parameters for the volatility term-structure effect (time dimension) and
combined effect of the volatility skew and volatility term structure. As can be seen from
the goodness of fit criteria, IVRMSE and %RMSE decreased for all six months with
ranges from a minimum of 0.007 and 0.006 to a maximum of 0.022 and 0.015,
respectively, and improved by 0.003 to 0.0032 and 0.001 to 0.002 points, respectively,
compared with Model 2 (smirk/skew model). Figure 5 displays a 3D graph of the IVs
for March 2004. The IVS generated by Model 3 is more closely fitted to the observed
IVS, with the last corner of the IVS stretched upward in comparison with the IVS in
Figure 4. Figure 5, therefore, suggests that there is volatility term structure effect in the
IVs, which is further confirmed by the results in Panel III of Table 3.

Figure 5. Estimated implied volatility surface with Model 3 for March 2004.

Panel IV of Table 3 presents the results for Model 4, which is Model 3 with an
additional parameter to capture the curvature effect of the volatility term-structure
effect. As can be seen from the goodness-of-fit criterion levels, IVRMSE and %RMSE
decreased for all six months with values ranging from a minimum of 0.007 and 0.006
to a maximum of 0.021 and 0.015, respectively, and improved from 0.00 to 0.001 and
0.00 to 0.001 points, respectively, compared with Model 3. Overall, compared with the
goodness of fit of the Model 3, very small improvements can be seen. Figure 6 presents
 61

the corresponding 3D IVS generated by Model 4. We observe a slight change in shape,

and the last corner of the IVS is now stretched somewhat further.

We conclude from the results presented in Table 3 that Model 1 cannot explain
variations in the IVS because it generates constant volatilities across both dimensions.
However, Model 4 explains most of the variations in the IVS. Therefore, Model 4 fits
the best among the models to the rich market IVS. However, the difference in
comparison with the fit with the Model 3 is marginal. Therefore, we conclude that the
forecasts of Model 3 and Model 4 for the IVSs are the best. Furthermore, we suggest
using the %RMSE goodness-of-fit criterion, particularly when IVs are implied by OTM
options such as in our study, because if we compare their levels, we find that %RMSE
throughout presents minimum levels in contrast to IVRMSE. Therefore, %RMSE is a
more suitable criterion because it assigns greater weight to OTM options.

Figure 6. Estimated implied volatility surface with Model 4 for March 2004.
 62

4.2 Results for Principal Component Analysis of Implied Volatility

Surfaces

Table 4 presents the results of PCA on the average IVSs for March and October for the
years 2004, 2005, 2006, and 2007. Columns 2, 3 and 4 show the results for the three
factors (or PCs) extracted from the IVS. Figure 7 presents the proportions of the factors
that explain the dynamics in the IVS for the March 2004, and the corresponding three
factor loadings for the same IVS are plotted in Figure 8.47 The first factor, which is
common to the entire IVS and moves the entire IVS systematically in the same
direction, can be interpreted as the volatility level factor (or parallel shift), which, on
average, explains about 56% of the variations in the implied volatility surfaces for
March and October in 2004, 2005, 2006, and 2007. It is important for IVs implied
from the shortest and longest maturity options because it alters the steepness of the
volatility structure of the near-term OTM put and call options, The effect is stronger on
the OTM options than ATM options; OTM puts are affected the most and feature mean-
reverting stochastic volatility after a shock that trigger instantaneous volatility. The
effect then persists for some time (the correlation between stochastic volatility and
jump)48. The second factor can be interpreted as a term-structure factor (or tilt, which
generates shifts in the slope of the term structure of IVs), which, on average, explains
about 15% of the variations in IVSs for all eight months and affects the shorter-and
longer-maturity IVs differently: the sign of the effects differs between the shorter- and
longer-maturity IVs. The term-structure effect factor is also consistent with the notion
that those segments in the IVS that are implied from long-term options do not
correspond entirely with those segments of the IVS that are implied from short-term
options; this factor can thus be attributed to long-term macroeconomic risks.
Moreover, factor 2 affects IVs across the moneyness dimension identically. The third
factor can be interpreted as the jump-fear factor (or curvature, which changes the
steepness of IV smirk/skew), which, on average, explains about 7% of the variation in
IVS for all eight months. The effects of jump-fear factor on the OTM put and call
options significantly differ and induce steepness in the IV smirk (skew). For empirical
evidence on the dynamics of the factors driving the IVS of index options, see the studies
by Skiadopoulos et al. (1999), Mixon (2002), and Cont and Da Fonseca (2002).

47Appendix includes Figures for the PCA of the other IVSs.

48The level factor can also be interpreted with the view that the change in IVs changes the skewness of the
corresponding implied risk-neutral density (for further discussion, see the studies by Mixon (2002) and
Zhang and Xiang (2008))
 63

Table 4. Principal components in the implied volatility surfaces

Month PC1 PC2 PC3 Total explained variance(%)

Mar-2004 64.0% 10.4% 7.7% b

82.1%

Mar-2005 53.6% 12.0% 10.0% 75.6%

Mar-2006 43.6% 31.3% 4.8% 79.7%

Mar-2007 61.8% 21.9% 4.7% 88.4%

Oct-2004 55.5% 13.0% 8.5% 77.0%

Oct-2005 59.3% 10.9% 7.0% 77.2%

Oct-2006 46.6% 12.5% 10.6% 69.7%

Oct-2007 67.3% 8.8% 6.5% 82.6%

Mean 56.4% 15.1% 7.4% 79.0%

Average Implied Volatility Surface, March- 2004

100 100%

90 90%

80 80%

70 70%
Variance Explained (%)

60 60%

50 50%

40 40%

30 30%

20 20%

10 10%

0 0%
1 2 3 4 5 6 7 8 9
Principal Components

Figure 7. Principal components in the IVS for March 2004

 64

Three Factor Loadings of the Implied Surface, March-2004

0.8

0.6 PC1
PC3

0.4

0.2

-0.2

-0.4
PC2
-0.6

-0.8

5 10 15 20 25 30

Figure 8. Three factor loadings of the IVS for March 2004.

For further investigation of the dynamics of IVs, we go beyond IVS and study each
group of maturities in the IVS, that is, 8-30, 30-60, 60-90, 90-150, 150-220 and 220
days and above. Table 5 documents the results for the different maturity groups. It is
evident that, as before, the first three factors can capture variations in each group.
Therefore, we find two distinct patterns in all six groups.

The level factor or parallel-shifts factor can explain, on average, 61% of the variance in
the first three maturity groups (from 8-90 days). However, for the remaining three
groups, the results are in complete contrast, where most of the dynamics can be
attributed to the level factor in the IVs; therefore, about 91% of the variations can be
explained by the level factor. In addition, the term structure factor, or tilt factor, can
explain about 19% of the variations in the IVs for the first three maturity groups,
whereas, for the last three groups, this factor, on average, explains 6% of the variance.
Finally, the jump-fear factor, or the curvature factor, is important for the first three
maturity groups (shorter maturities) that capture the jump-fear in the market (the
pattern implied by deep OTM options), which explain 10% of the variation in the IVs
but only 2% of the variation in the last three maturity groups.
 65

Table 5. Principal components in the different maturity groups

Range Month PC1 PC2 PC3 Total Month PC1 PC2 PC3 Total
8-30 Mar-04 68.03% 14.35% 7.84% 90.2% Oct-04 60.02% 25.98% 8.79% 94.7%

Mar-05 68.87% 14.97% 8.53% 92.3% Oct-05 66.63% 15.33% 9.97% 91.9%

Mar-06 72.79% 14.62% 6.00% 93.4% Oct-06 64.71% 19.01% 8.13% 91.8%

Mar-07 50.70% 35.73% 9.60% 96.0% Oct-07 65.81% 13.97% 9.59% 89.3%

Mean 64.7% 19.2% 8.5% 92.4%

30-60 Mar-04 69.72% 16.86% 8.31% 94.8% Oct-04 55.28% 22.17% 15.24 92.6%

Mar-05 63.28% 14.16v 10.99 74.2% Oct-05 52.15% 22.91% 11.00 86.0%

Mar-06 54.26% 15.23% 10.72 80.2% Oct-06 50.16% 20.39% 15.28 85.8%

Mar-07 77.50% 15.44% 4.00% 96.9% Oct-07 63.60% 15.07% 6.38% 85.0%

Mean 60.4% 17.7% 10.2% 86.9%

60-90 Mar-04 48.65% 21.24% 17.95 87.8% Oct-04 69.06% 11.99% 7.50% 88.5%

Mar-05 57.23% 20.09% 14.01 91.3% Oct-05 69.79% 14.25% 7.02% 91.0%

Mar-06 55.10% 22.22% 9.76% 87.0% Oct-06 52.63% 22.47% 14.33 89.4%

Mar-07 47.36% 32.08% 7.72% 87.1% Oct-07 77.81% 12.93% 5.31% 96.0%

Mean 59.7% 19.6% 10.4% 89.7%

90-150 Mar-04 94.74% 2.72% 1.44% 98.9% Oct-04 77.65% 15.71% 5.44% 98.8%

Mar-05 92.55% 5.62% 1.13% 99.3% Oct-05 94.87% 3.44% 1.10% 99.4%

Mar-06 90.98% 7.40% 1.03% 99.4% Oct-06 63.30% 18.64% 14.16% 96.1%

Mar-07 97.28% 1.67% 1.00% 99.5% Oct-07 88.37% 8.78% 1.23% 98.3%

Mean 87.4% 8.0% 3.3% 98.7%

150-220 Mar-04 95.90% 1.43% 1.27% 98.6% Oct-04 94.13% 2.73% 2.41% 99.2%

Mar-05 67.29% 29.78% 1.58% 98.6% Oct-05 97.91% 0.96% 0.75% 99.6%

Mar-06 90.58% 6.01% 1.93% 98.5% Oct-06 92.04% 4.08% 2.90% 99.0%

Mar-07 88.01% 10.43% 1.17% 99.6% Oct-07 96.52% 1.50% 0.92% 98.9%

Mean 90.3% 7.1% 1.6% 99.0%

220- Mar-04 93.17% 4.23% 1.65% 99.0% Oct-04 95.73% 3.15% 0.90% 99.7%

Mar-05 97.82% 1.32% 0.60% 99.7% Oct-05 96.58% 1.73% 1.10% 99.4%

Mar-06 92.52% 6.69% 0.63% 99.8% Oct-06 93.10% 3.40% 1.91% 98.4%

Mar-07 98.09% 1.28% 0.55% 99.9% Oct-07 96.51% 2.11% 1.00% 99.6%

Mean 95.4% 3.0% 1.0% 99.4%

 66

To summarize Table 5, we conclude that shocks that are common to the entire IVS that
can be captured by the first factor (or PC ) are very important for the shortest
maturities (8-30 days) and for the longest (90 days-above) maturities; therefore, this
factor is of sole importance for these IVs. However, the term structure factor is
important for the IVs with medium (particularly between 60-90 days) maturities, and
the jump-fear factor is important for IVs with shorter (8-90 days) maturities.

Figure 9 is a plot of the forecasted values (blue circles) and actual values (red line) for
the March 2004 IVS; the IVS is forecasted (predicted) using a different number of
retained factors (PCs) from the market (or actual) IVS. There are four subplots in
Figure 9: the first subplot shows predicted IVS values using only the first latent factor,
the second subplot forecasts IVS values with the first two factors, the third subplot
forecasts IVS values with the first three factors, and the last uses the first 20 factors to
forecast the IVS values. It is evident that the first three factors, the level factor, the term
structure factor, and the jump-fear factor, are more than adequate for forecasting IVS,
and these factors can explain most of the dynamics of the market IVS. The results in
Figure 9 match our previous analysis.

One PC and Imp Surface Mar-04 Two PCs and Imp Surface Mar-04
40 40
Implied Volatility(%)

Implied Volatility(%)

30
30
20
20
10

0 10
0 20 40 60 0 20 40 60
Term structure Term structure
Three PCs and Imp Surface Mar-04 20 PCs and Imp Surface Mar-04
40 40
Implied Volatility(%)

Implied Volatility(%)

30 30

20 20

10 10
0 20 40 60 0 20 40 60
Term structure Term structure

Figure 9. Forecasted IVS (blue circles) versus Market observed IVS (red line) using
PCs.
 67

5 Conclusion

We estimated and obtained implied volatility surfaces for FTSE 100 index options with
parametric structural forms of models proposed by Dumas et al. (1998). We modified
these models by considering moneyness in implied forward prices and OTM put and
call options. Further, we investigated the dynamics of the IVSs using principal
component analysis.

We estimated four parametric models of DFW (1998): the constant-volatility Model 1

fails to capture variations in the IVSs and yields only flat IVSs. However, Model 2,
which captures the volatility skew and volatility smirk effects, does a good job and
captures most of the volatility skew and smirk effects. Models 3 and 4, which account
for both the volatility term structure and the volatility smirk/skew effects, performed
better than the one-dimensional smirk/skew model. However, both models fit the
market IVS well and generated the best IVSs for the observed data. In the second part
of the study, which uses PCA (nonparametric approach), we find that the first three
factors can explain about 69-88% of the variance in the IVS: on average, 56% is
explained by the level factor, 15% is explained by the term structure factor, and an
additional 7% is explained by the jump-fear factor. Finally, when applying PCA to the
maturity groups, we find that the level factor (or parallel shifts) is evident and thus
important for shorter- and longer-maturity IVs. However, the term-structure factor is
important for IVs with medium maturities, and the jump-fear factor is important for
IVs with the shortest maturities (8-90 days).

IVS is important in practice; for instance, on average, in the FTSE 100 index option
market, 100 options are traded daily with a variety of strike prices and times to
maturity. If practitioners need to quote more options for different strikes or maturities,
then IVS can be used as a reference. The practitioner could easily quote new options by
taking estimates from the IVS. Moreover, the shape of the IVS is important in option
trading: one of its most important characteristics is that it can be used to identify
mispricing in the market. Second, the IVS can be used to manage exotic derivative
positions, which could be hedged with plain-vanilla options such as European options.
Third, a robust volatility index could easily be constructed from the IVS and
consequently used to price volatility derivatives. Fourth, risk can be managed for those
securities that underlie the FTSE 100 index by using independent risk factors from the
IVS.
 68

References
Ait-Sahalia, Y., and Lo, A., 1998. Nonparametric estimation of state-price densities
implicit in financial asset prices. Journal of Finance 53, 499-547.
Alentorn, A., 2004. Modelling the implied volatility surface: An empirical study for
FTSE options, Working Paper. University of Essex, UK.
Alexander, C., 2001. Principal component analysis of implied volatility smiles and
skews. ISMA Discussion Paper, University of Reading, UK.
Bates,D., 1991. The crash of 87 : was it expected ? The evidence from options markets.
Journal of Finance 46, 1009-1044.
Bates, D., 2008. The market for crash risk. Journal of Economic Dynamics & Control
32, 2291- 2321.
Black, F., and Scholes, M., 1973. The pricing of options and corporate liability. Journal
of Political Economy 81, 636-654.
Blair, B., Poon, S., and Taylor, S., 2001. Forecasting S&P 100 volatility: The
incremental information content of implied volatilities and high frequency
index returns. Journal of Econometrics 105, 5-27.
Bollen, N., and Whaley, R., 2004. Does net buying pressure affect the shape of the
implied volatility functions? Journal of Finance 59, 711-753.
Camara, A., Heston, S., 2008. Closed form option pricing formulas with extreme
events. Journal of Futures Markets, 28, 213-230.
Carr, P., Helyette, G., Madan, D., and Yor, M., 2002. The fine structure of asset returns,
an empirical investigation. Journal of Business 75, 305-332.
Carverhill, A., Cheuk, T., and Dyrting, S., 2009. The smirk in the S&P 500 futures
options prices: A linearized factor analysis. Review of Derivatives Research 12,
109-139.
Cassese, G., and Guidolin, M., 2006. Modelling the implied surface: Does market
efficiency matter? An application to MIB30 index options. International Review
of Financial Analysis 15, 145-178.
Christensen, B., and Prabhala, N., 1998. The relation between implied and realized
volatility. Journal of Financial Economics 50, 125-150.
Christoffersen, P., and Jacobs, K., 2004. The importance of the loss function in option
valuation, Journal of Financial Economics 72, 291-318.
Cont, R., and Da Fonseca, J., 2002. Dynamics of implied volatility surfaces.
Quantitative Finance 2, 45-60.
David, A., Veronesi, P., 2000. Option pricing with uncertain fundamentals: Theory and
evidence on the dynamics of implied volatilities. CRSP Working Paper No.485.
Doran, J., Peterson, D., and Tarrant, B., 2007. Is there information in the volatility
skew? Journal of Futures Markets 27, 921-959.
Dumas, B., Fleming, J., and Whaley, R.E., 1998. Implied volatility functions: Empirical
tests. Journal of Finance 53, 2059-2106.
Ederington, L., and Guan, W., 2002. Is implied volatility an informationally efficient
and effective predictor of future volatility? Journal of Risk, 4, 29-46.
Fengler, M., Härdle, W., and Mammen, E., 2007. A semiparametric factor model for
implied volatility surface dynamics. Journal of Financial Econometrics 5, 189-
218.
Fengler, M., Härdle, W., and Villa, C., 2003. The dynamics of implied volatility: A
common principal components analysis approach. Review of Derivative
Research 6, 179-202.
Fleming, J., 1998. The quality of market volatility forecasts implied by S&P100 index
option prices. Journal of Empirical Finance 5, 317-345.
Foresi, S., and Wu, L., 2005. Crash-o-phobia: a domestic fear or a worldwide concern.
Journal of Derivatives 13, 8-21.
 69

Fung, W., and Hsieh, D., 1991. Empirical analysis of implied volatility: Stocks, bonds
and currencies. paper presented at the 4th Annual Conference of Financial
Options Research Center,University of Warwick, UK.
Garcia, R., Luger, R., Renault, E., 2003. Empirical assessment of an intertemporal
option pricing model with latent variables. Journal of Econometrics 116, 49-83.
Goncalves, S., and Guidolin, M., 2006. Predictable dynamics in the S&P 500 index
options implied volatility surfaces. Journal of Business 79, 1591-1635.
Gross, L., and Waltner, N., 1995. S&P500 options: Put volatility smile and risk
aversion. Salomon Brothers, mimeo.
Heston, S., 1993. A closed form solution for options with stochastic volatility with
applications to bond and currency options. Review of Financial Studies 6, 327-
343.
Hull, J., 2002. Options, futures, and other derivatives. Fifth edition, Prentice Hall, New
Jersey, United States.
Jackwerth, J., and Rubenstein , M., 1996. Recovering probability distributions from
options price. Journal of Finance 51, 1611-1631.
Leland, H., 1985. Option pricing and replications with transaction costs. Journal of
Finance 40, 1283-1301.
Le Roux, M., 2006. A long-term model of the dynamics of the S&P500 implied
volatility surfaces. Working Paper, ING Institutional Markets.
Liu, J., Pan, J., and Wang, T., 2005. An equilibrium model of rare-event premia and its
implication for option smirks. Review of Financial Studies 18, 131-164.
London, J., 2004. Modeling derivatives in C++. First edition, Wiley Finance, New
Jersey, United States.
Merton, R.C., 1973. Theory of rational option pricing. Bell Journal of Economics and
Management Sciences 4, 141-183.
Mixon, S., 2002. Factors explaining movements in the implied volatility surface.
Journal of Futures Markets 22, 915-937.
Pan, J., 2002. The jump-risk premia implicit in options: evidence from an integrated
time series study. Journal of Financial Economics 63, 3-50.
Poon, S., and Granger, C., 2003. Forecasting volatility in financial markets: A review.
Journal of Economic Literature 41, 478-539.
Rouah, F., and Vainberg, G., 2007. Option pricing models & volatility using Excel-VBA.
Wiley Finance, New Jersey, United States.
Rubinstein, M., 1994. Implied binomial trees. Journal of Finance 49, 771-818.
Skiadopoulos, G., Hodges, S., and Clewlow, L., 1999, The dynamics of the S&P500
implied volatility surface. Review of Derivatives Research 3, 263-282.
Tompkins, R., 2001. Implied volatility surfaces: uncovering regularities for options on
financial Futures. European Journal of Finance 7, 198-230.
Zhang, J., and Xiang, Y., 2008. The implied volatility smirk. Quantitative Finance 8,
263- 284.
 70

Appendix. Estimated implied volatility surfaces with four models for Oct-2005.
 71

Principal component analysis of the IVS for Oct-2005

Average Implied Volatility Surface, October- 2005

90 90%

80 80%

70 70%
Variance Explained (%)

60 60%

50 50%

40 40%

30 30%

20 20%

10 10%

0 0%
1 2 3 4 5 6 7 8 9
Principal Components

Three Factor Loadings of the Implied Surface, October-2005

0.8

0.6
PC1

0.4

0.2

-0.2

-0.4

-0.6 PC3 PC2

-0.8

5 10 15 20 25 30
 72

One PC and Imp Surface Oct-05 Two PCs and Imp Surface Oct-05
30 25
Implied Volatility(%)

Implied Volatility(%)
20
20
15
10
10

0 5
0 20 40 60 0 20 40 60
Term structure Term structure
Three PCs and Imp Surface Oct-05 20 PCs and Imp Surface Oct-05
30 25
Implied Volatility(%)

Implied Volatility(%)
25
20
20
15
15

10 10
0 20 40 60 0 20 40 60
Term structure Term structure
 73

Quantile Regression Analysis of Asymmetric Return-

Volatility Relation

Ihsan Ullah Badshah

Hanken School of Economics, Department of Finance and Statistics, P.O. Box 287,
FIN-65101 Vaasa, Finland. Phone: +358-6-3533 721, Fax: +358-6-3533 703,
Email: ibadshah@hanken.fi

September 20, 2010

Abstract

This essay uses quantile regression to investigate the asymmetric return-volatility

phenomenon with the newly adapted and robust implied volatility indices VIX, VXN,
VDAX and VSTOXX. A particular goal is to quantify the effects of positive and negative
stock index returns at various quantiles of the IV changes’ distribution. As the level of
the new volatility index increases during market declines, we believe that the negative
asymmetric return-volatility relationship should be significantly more pronounced at
upper quantiles of the IV changes’ distribution than is indicated by ordinary least
squares (OLS) regression. We find pronounced negative and asymmetric return-
volatility relationships between each volatility index and its corresponding stock
market index. The asymmetry increases monotonically when moving from the median
quantile to the uppermost quantile (i.e., 95%); OLS thereby underestimates this
relationship at upper quantiles. Additionally, the asymmetry is pronounced with a
volatility smirk (skew)-adjusted new volatility index measure in comparison to the old
at-the-money volatility index measure. The VIX volatility index presents the highest
asymmetric return-volatility relationship, followed by the VSTOXX, VDAX and VXN
volatility indices. Our findings have implications for trading strategies, hedging
portfolios, pricing and hedging volatility derivatives, and risk management.

Keywords: Asymmetric return-volatility relationship, implied volatility, index

options, quantile regression, volatility index.

JEL Classifications: C21, G12, G13.

The Author would like to thank Johan Knif, Mika Vaihekoski, George Skiadopoulos, Hossein Asgharian,
and Kenneth Högholm for providing useful comments. The author would also like to thank the participants
of the NFN workshop (2009), Copenhagen, Denmark; FMA conference (2009), Turin, Italy; MFS
conference (2009), Crete, Greece; and EFMA conference (2010), Aarhus, Denmark. The author
acknowledges CEFIR (centre for financial research) and NASDAQ OMX Nordic foundation for providing
financial support.
 74

1 Introduction

It is widely documented that implied volatility (IV) is superior to historical volatility

(HV) when forecasting the future realized volatility (RV) of the underlying asset (e.g.,
Day and Lewis, 1992; Christensen and Prabhala, 1998; Fleming, 1998; Dumas et al.,
1998; Blair et al., 2001; Ederington and Guan, 2002; Poon and Granger, 2003;
Mayhew and Stivers, 2003; and Martens and Zein, 2004). IV can be recovered by
inverting the Black-Scholes (1973) formula. However, Britten-Jones and Neuberger
(2000) and Jiang and Tian (2005) have derived a model-free implied volatility (MFIV)
under the pure diffusion assumption and asset price processes with jumps.49 This MFIV
measure has now been adapted by the major IV indices, which used to employ at-the-
money (ATM) BSIV measures in their methodologies.50

As MFIV index is forward looking, that is, it is implied by the market prices of options,
and as options represent the consensus (aggregated beliefs) of market participants
regarding expected future volatility, MFIV index is the market expectation about the
future RV of the underlying stock index for the 22 subsequent trading days (over 30
calendar days).51 Thus, IV indices are often referred to as the “investors’ fear gauge”
(e.g., Whaley, 2000), as the level of the IV index indicates the consensus view about the
expected future realized stock index volatility. When the level of the IV index increases,
fear increases in the market as a result; alternatively, when the level of the IV index
decreases, run-ups are triggered in the daily stock index prices.52

49 They show that the information content of MFIV is superior to that of the Black-Scholes implied
volatility (BSIV) because the MFIV measure accounts for all strikes when computing IV at a particular
point in time, whereas the BSIV measure is a point-based IV and does not account for all strikes in
computation; i.e., each strike has a separate IV. Moreover, BSIV is subject to both model and market
efficiency, while MFIV is only subject to the market efficiency (see Poon and Granger, 2003).
50 The motives for adopting MFIV measures are the following. First, the MFIV index measure is

economically appealing and robust, as it accounts for out-of-the-money (OTM) options (i.e., volatility
smirk/skew). Second, the previous IV index measure (now called VXO) was upward biased, the bias being
induced by trading-day conversion, which is now omitted from the new VIX measure. Finally, with the new
robust MFIV index measure, it is possible to replicate volatility derivatives (e.g., variance swaps), which
was not possiblewith the previous measure.
51 Major option exchanges, including the Chicago Board of Exchange (CBOE) and the Deutsche Börse, have

launched IV indices, robustly providing information on options using MFIV measures; examples of this are
the VIX index for the S&P 500 index, VXN for the NASDAQ 100 index, VDAX for the DAX 30 index and
VSTOXX for the Dow Jones(DJ) EURO STOXX 50 index.
52 Additionally, the IV index level indicates the degree of willingness of market participants to pay in terms

of volatility in order to hedge the downside risk of their portfolios with put options or long positions in call
options with downside risks remain to the premium of options instead of positions in the underlyingasset
(see Simon, 2003, for a detail on trading strategies).
 75

Likewise, the MFIV index measure incorporates both puts and calls and therefore
moves with changes in their prices; for example, a negative or positive shock to the
market induces adjustments in hedging and trading strategies, consequently triggering
changes in the prices of one type (i.e., put or call) of option. The MFIV index measure
then moves in the direction of the market demand of a particular type of option and the
underlyingasset (see Bollen and Whaley, 2004). Also, Liu et al. (2005) argue that the
rare-event premia play an important role in generating the volatility smirk (skew)
pattern observed for options across moneyness and that these rare events are
embedded in the OTM options.53 Camara and Heston (2008) derive an option model
that accounts for both OTM put and call options. They derive the extreme negative
events from OTM puts and extreme positive events from OTM calls. Thus, the MFIV
index that accounts for OTM options thus contains a broader set of information—it
contain information on the future negative and positive jumps, demand/supply of
options etc. Thereby MFIV measure is robust; as a result, an excellent tool for
examining the relationship between the market perception of volatility and returns.

Furthermore, this relationship is asymmetric, implying that the MFIV index reacts
differently to negative and positive returns. Two main hypotheses exist in the literature
that characterizes the negative asymmetric return-volatility relationship: the leverage
effect and feedback effect hypotheses. However, both the leverage and feedback
hypotheses have been unable to explain the observed strong negative asymmetric
return-volatility relationship at daily frequencies (see, e.g., French et al., 1987; Breen et
al., 1989; Schwert1989, 1990). Similarly, a recent study by Hibbert et al. (2008) has
found a very strong contemporaneous negative asymmetric return-volatility
relationship using daily data, thereby empirically rejecting both the leverage and
volatility feedback hypotheses.54 Further empirical investigations are important to
characterize an asymmetric return-volatility relationship using a volatility smirk
(skew)-adjusted robust MFIV index measure with daily frequency.55 Importantly, we

53 Similarly, Pan (2002) showed that volatility skew is primarily due to investors’ fear of large adverse
jumps.
54 Other studies by Simon (2003) and Giot (2005) have also found a very strong negative asymmetric

return-volatility relationship using data of daily frequency. Nevertheless, the negative asymmetric return-
volatility relationship is too strong at the daily level; these hypotheses might be interesting to characterize
an asymmetric relation at lower frequencies, for instance, monthly or quarterly frequencies, but not at high
frequencies.
55 We also believe that the asymmetric volatility-return relationship should be more pronounced with the

new robust MFIV index in contrast to the old BSIV index measure. A possible explanation for pronounced
asymmetric volatility is that a put option is a downside-hedging instrument and traders are always
 76

believe that the estimation technique currently being used for characterizing this
relationship is ordinary least squares (OLS), which may not completely capture this
relationship because it only describes that how this relationship should be at the mean
of the IV changes’ (or response variable) distribution but not at the other parts of the IV
changes’ distribution. Therefore, to fully characterize the asymmetric return-volatility
relationship, the conditional quantile regression should be preferred over OLS
regression, especially to obtain a complete picture of the relationship at all parts of the
IV changes’ distribution (particularly at the uppermost quantiles) with the negative and
positive stock returns.56 Because we believe that the relationship should be more
pronounced and asymmetric at the uppermost quantiles (e.g., 95th quantile) of the IV
changes’ distribution, which OLS regression (or mean regression) would
underestimate. Few well-known studies exist showing a significant negative and
asymmetric relationship between stock index returns and BSIV index changes using
OLS regressions (or mean regressions) (e.g., Fleming et al., 1995; Whaley, 2000; Giot,
2005; Simon, 2003; Skiadopoulos, 2004; Low, 2004; Dennis et al., 2006); as OLS
ignores the responses at the tails of the IV changes’ distribution, therefore, the
asymmetric return-volatility relationship is underestimated, warrant further
investigation using conditional quantile regression analysis.

Nonetheless, the first study on the relationship between the old VIX (now VXO)
changes and S&P 100 index returns was conducted by Fleming et al. (1995). They
investigated the time-series properties of the VXO, finding a significant negative
contemporaneous asymmetric relationship between VXO changes and stock index
returns. Another well-known study is that conducted by Whaley (2000), who examined
the relationship between the weekly VXO changes and S&P 100 returns. He
documented that when the VXO falls by 100 basis points, the S&P 100 index increases
by 0.469%, whereas when the VXO increases by 100 basis points, the S&P 100 index
falls by -0.707%. He thus finds a large negative asymmetric association between VXO
changes and S&P 100 returns, calling the VXO the “investors’ fear gauge.” Simon
(2003) studied the NASDAQ 100 volatility index (VXN) from January 1995 to May
2002, showing that the VXN is inversely related to both positive and negative index
returns. Furthermore, he found stable results across the bubble and post-bubble

concerned about the downward moments in the market, so traders are always hedging their positions with
OTM puts. Consequently, we find a higher volatility for OTM puts than for calls (see, e.g., Bollen and
Whaley, 2004).
56 Quantile regression estimates are robust to outliers, non-normal error distribution, etc.
 77

periods. A more recent study by Hibbert et al. (2008) used a different approach to
investigate the negative asymmetric return-volatility relationship using the newly
developed VIX index. They found a significant negative and asymmetric association
between VIX changes and stock index returns when incorporating both daily and
intraday data, thereby confirming that the MFIV VIX measure can better explain the
asymmetric relationship than the BSIV VIX or the RV measures.

The purpose of this essay is to investigate the negative asymmetric return-volatility

relationship between the stock market returns and the volatility index changes: (1) to
quantify the degree to which a volatility index is responding to the negative and positive
returns at different quantiles of an IV changes’ distribution; (2) to compare the
asymmetric responses of the two volatility index measures, i.e., the MFIV and BSIV
index measures; and (3) to rank volatility indices according to their asymmetries.
Related studies in terms of the volatility-return relationship include Simon (2003), Giot
(2005) and Hibbert et al. (2008). Simon (2003) studied the relationship between the
NASDAQ 100 index returns and the VXN index changes using the BSIV index measure,
while Giot (2005) and Hibbert et al. (2008) studied the relationship between the S&P
100 and the VIX and between the NASDAQ 100 and the VXN. Giot (2005) used the
BSIV index measure, whereas Hibbert et al. (2008) used the new MFIV index measure.
Our study differs from these three previous studies and therefore contributes to the
literature in a number of ways: first, this study extends their methodologies; for
instance, they used mean-regression models, whereas we use a robust conditional
quantile regression model to investigate the uppermost IV changes’ quantiles’
responses to the negative and positive returns. Second, this study uses a broader set of
data drawn from across the Atlantic, for example, the VIX, VXN, VDAX and VSTOXX
volatility indices (using new robust MFIV measures, thereby incorporating a broader
range of information) and their corresponding stock indices.57 Finally, this study
compares the asymmetries of the MFIV and BSIV volatility index measures.58

Our main findings are that from February 2001 through May 2009, the MFIV indices
VIX,VXN, VDAX and VSTOXX responded in a highly asymmetric fashion; i.e., negative
returns had a much greater impact on volatility than positive returns, particularly in the

57 Previously it was found that each equity option market presented somewhat different IV dynamics;

therefore, this study is the first to investigate and compare the volatility asymmetries across the Atlantic.
58 We compare the new VDAX and old VDAX (denoted here VDAXO) volatility index measures; the former

is based on the MFIV measure and the latter on the BSIV measure.
 78

uppermost regression quantiles (e.g., q=0.95). Our quantile regression model (QRM) of
the asymmetric return-volatility relation thus reveals important information that is
underestimated by the mean-regression model (MRM). The VIX index presents the
highest asymmetry, followed by the VSTOXX, VDAX and VXN indices, respectively.
These volatility indices rise sharply in times of market turmoil and decline in market
rallies. Second, our view that the asymmetry with the MFIV index should have
pronounced responses is confirmed by comparing the asymmetric responses of VDAX
(MFIV) and VDAXO (BSIV); we find that the MFIV index responds in a pronounced
fashion, in contrast to the BSIV index. Third, there is a strong contemporaneous
asymmetry in comparison to the lags, thus rejecting the leverage hypothesis, and
similar conclusions can be drawn for the feedback hypothesis.

This essay is organized as follows. Section 2 discusses the asymmetric return-volatility

relationship. Section 3 discusses the data set, the volatility indices and their
construction. Section 4 describes the methodology. Section 5 presents the empirical
results. Section 6 summarizes and concludes.
 79

2 Asymmetric Return-Volatility Relationship

There are two existing hypotheses that characterize asymmetric volatility: the leverage
effect and the feedback effect hypotheses. The leverage hypothesis proposed by Black
(1976) and Christie (1982) attributes asymmetric volatility to the financial leverage of a
firm; i.e., when a firm’s debt increases, the firm’s value declines, triggering the value of
its equity declines further. Because the equity of a firm is more exposed to the firm’s
total risk, therefore, the volatility of the equity should increase as a result. On the other
hand, the volatility feedback hypothesis proposed by French et al. (1987), Campbell and
Hentschel (1992) and Bakaert and Wu (2000), who attribute asymmetric volatility to a
volatility feedback effect.59 Contrary to the leverage-based justification, the volatility
feedback hypothesis states that increases in volatility trigger negative stock returns. For
instance, increases in volatility imply that the required expected future returns will
increase as well; as a result, current stock prices decline. However, both hypotheses
empirically fail under the daily frequency data, being unable to fully characterize the
asymmetric return-volatility relationship; in that respect, Schwert (1990) argued that it
is too strong for the leverage hypothesis to fully characterize asymmetric volatility.
Furthermore, it is also empirically found that the feedback hypothesis is not always
consistent, and this has become a controversial subject; some studies have found that
there are not always positive correlations between current volatility and expected future
returns (e.g., Breen et al., 1989), but others support the hypothesis (e.g., French et al.,
1987; Campbell and Hentschel, 1992; Ghysels et al., 2005). Nonetheless, the economic
and accounting explanations might be important for characterizing the asymmetric
return-volatility relationship at lower frequencies, for instance, monthly or quarterly
data, but not for daily or higher frequencies. Many prior studies have documented very
strong negative asymmetric return-volatility relationships at higher frequencies,
contrary to the explanations of the two hypotheses (see, e.g., Fleming et al., 1995;
Whaley, 2000; Giot, 2005; Simon, 2003; Skiadopoulos, 2004; Low, 2004; Dennis et
al., 2006; Hibbert et al., 2008).

However, this essay considers new MFIV indices because we believe that the
asymmetric return-volatility relation should be more pronounced using the MFIV
indices. Likewise, the importance of the MFIV index measure increases because it

59 Poterba and Summers (1986) characterized the volatility feedback effect through economic explanation

that time-varying risk premia induces volatility feedback because it represent the linkage between the
fluctuations in volatility and stock returns.
 80

accounts for volatility smirk (skew), which may be induced by the net buying pressure
of the OTM put options (see Bollen and Whaley, 2004). Volatility smirk/skew is an
obvious phenomenon, previously documented by many other researchers and
important to capture in any volatility measure (e.g., Alexander, 2001; Low, 2004;
Goncalves and Guidolin, 2006; Badshah, 2008). Bollen and Whaley (2004)
documented that the net-buying pressure of the put options induces volatility smirk
(skew) patterns, therefore, suggested using a volatility smirk (skew) adjusted volatility
measures: who investigated the relationship between net buying pressure and the
shape of the IV function (IVF) for index options. They showed that the buying pressure
of put options considerably affects the changes in the IV. They asserted that when the
buying pressure of index put options (particularly from institutional investors who seek
to hedge their portfolios) increases and thus limits the ability of arbitrageurs to bring
the price back into alignment, this pressure permanently drives the sloping shape of the
IVF downward. Also, information from trading strategies and other shocks are well
absorbed into the MFIV index, as it accounts for both put and call types of options.60
The MFIV index is a robust and informed measure of stock index volatility and is
therefore a perfect choice for examining the asymmetric volatility-return relationship.

3 Data

First, the VIX, VXN, VDAX, and VSTOXX volatility indices are introduced, and their
construction is discussed. Second, the complete data set is presented, and the
descriptive statistics are thoroughly discussed.

3.1 VIX and VXN

The CBOE introduced a new VIX index in September 2003 based on options on the
S&P 500 index. The VIX index is determined from the bid and ask prices of the options
underlying the S&P 500 index. The new VIX is independent of any option pricing
model usingthe MFIV measure. The VIX thus provides an estimate of expected future
realized stock market volatility for the 22 subsequent trading days (over 30 calendar
days). However, the old VIX index, based on options on the S&P 100 index and
introduced in 1993, has now moved to the new ticker symbol VXO. In contrast to the

60 The MFIV index is informed by both fear and exuberance embedded in option prices, and the majority of

option markets’ traders are very informed and possess high skill levels (see Low, 2004; Chakravarty et al.,
2004).
 81

old VIX (now VXO), which is based on near-the-money BSIV options on the S&P 100
index, the new VIX uses market prices of options on the S&P 500 index.61 This new MF
VIX methodology accounts for both OTM put and call options (i.e., volatility
smirk/skew).62 The new methodology is thus more appealing and robust. The CBOE’s
introduction of the new VIX was motivated by both theoretical and practical
deliberations. First, the new VIX is economically more appealing as it is based on a
portfolio of options, whereas the old VIX was based on the ATM option prices. Second,
the new VIX makes it easy to replicate variance swap payoffs while using static
positions in a range of options and dynamic positions in futures trading. Third, the new
VIX has removed the induced upward bias of the old VIX in the trading day conversion
(see, e.g., Carr and Wu, 2006). Similarly, in September 2003, the CBOE introduced the
VXN using the same MFIV methodology as that of VIX. The CBOE has calculated price
histories for VIX and VXN back to the years 1986 and 2001, respectively.

3.2 VDAX and VSTOXX

The Deutsche Börse and Goldman Sachs jointly developed the methodology for the new
VDAX and VSTOXX indices. The VDAX is based on options on the DAX 30 index,
whereas VSTOXX is based on options on the Dow Jones (DJ) Euro STOXX 50 index,
which consists of the eurozone’s50 largest blue-chip stocks. Options on the DAX and
the DJ Euro STOXX 50 are traded on the EUREX derivatives exchange. The VDAX
measure accounts for IVs across all options of a given time to expiry (accounting for
volatility smirk/skew). The methodology of the VDAX, like that of the VIX, is based on
the MFIV measure.63 However, the main VDAX index is further based on eight
constituent volatility indices, which expire in 1, 2, 3, 6, 9, 12, 18, and 24 months,
respectively. The main VDAX is designed as a rolling index at a fixed 30 days to expiry
via a linear interpolation of the two nearest of the eight available sub-indices. The
VDAX and its eight sub-indices are updated every minute and therefore offer great
advantages in terms of trading, hedging and introducing new derivatives on this index.

61 The options on the S&P 500 index, in comparison with the options on the S&P 100 index, contain a

much broader set of implied information; the new VIX is thus a more informative measure than the old
VIX (now VXO).
62See, for further detail on the VIX construction, the CBOE VIX white paper:
www.cboe.com/micro/VIX/vixwhite.pdf.
63 See, for further detail on the VSTOXX/ VDAX construction, the STOXX website: www.stoxx.com; the

Deutsche Börse and Goldman Sachs methodology for VSTOXX/ VDAX volatility indices.
 82

The price histories for both VDAX and VSTOXX were calculated back to the years 1992
and 1999, respectively.

3.3 Data

This essay employs data from four sources. We obtained the daily time-series price data
for the S&P 500 stock index, the NASDAQ 100 index, the DAX 30 index, and the DJ
Euro STOXX 50 index from Thomson Financial DataStream. The data on the new VIX
and VXN were obtained from the CBOE, and the data on the new VDAX and VSTOXX
were obtained from the Deutsche Börse and STOXX, respectively. The daily frequency
data for both stock and volatility indices cover a period of 8 years and 4 months, from
February 2, 2001 to May 29, 2009, for a total of 2172 trading days.

Figure 1 shows the daily closing levels (%) of the volatility indices, i.e., the VIX, VXN,
VDAX and VSTOXX, and the corresponding stock market indices (levels) from
February 2, 2001 to May 29, 2009. Among the four volatility indices, the VXN index
presents the highest volatility level throughout our sample period, whereas the VIX
shows the lowest volatility level. Similarly, the volatility indices and stock indices move
inversely to one other. From the beginning of 2001 until the beginning of 2003, there
were considerably high volatility levels. However, from 2004 to late 2007, we find
upward movement in the stock market indices, whereas the corresponding volatility
indices moved in the opposite direction to the stock markets, showing the lowest
volatility levels. However, in the latter part of 2007, we again find somewhat increasing
volatility levels, with the stock markets again moving downward due to the beginning of
the credit crunch and liquidity crunch crises, which have caused markets to be
extremely volatile and the volatility indices to reach historically high levels
(particularly, in October and November of 2008, the VIX level twice surpassed 80%),
and the corresponding stock markets crashed afterward, therefore moving in
completely opposite directions. This phenomenon is evident until the end of the sample
period in May 2009.
 83

Figure 1. Stock indices versus MFIV indices from February 2, 2001, to May 29, 2009.

Table 1 reports the summary statistics for the daily percentage continuously
compounded returns of four stock indices and the daily percentage changes of five
 84

volatility indices, as well as tests for normality, autocorrelations and unit roots. The
mean values for all nine stock index returns and volatility index changes series are not
statistically different from zero. The tests for skewness and kurtosis confirm that the
stock indices returns are positively skewed except for the S&P 500 returns, whereas all
five volatility indices’ changes are positively skewed, as they should be. Furthermore, all
nine series are highly leptokurtic with respect to the normal distribution. Likewise, the
Jarque-Bera statistics reject normality for each of the stock index and volatility index
changes series. The autocorrelation coefficients for the three lags show that the VIX,
VDAX and VXN changes series present strong autocorrelations, whereas the rest of the
volatility indices changes’ present significant autocorrelation coefficients at lags 2 and
3. Autocorrelations in the S&P 500, NASDAQ 100, and DJ Euro STOXX 50 returns
series are also evident at all three lags, consequently confirming the property of mean
reversion. We also investigated stationarity in all nine series (i.e., stock and volatility
indices) by applying the augmented Dickey-Fuller (ADF) unit-root test. The results in
Table 1 show the rejection of unit roots in each series at the 1% significance level.
Therefore, all nine series are stationary.

Table 1. Descriptive statistics of the daily data.

S&P500 NASDAQ DAX30 STOXX50 VIX VXN VDAX VDAXO VSTOXX
Mean -0.0184 -0.0274 -0.0140 -0.0300 0.0033 -0.0117 0.0058 0.00511 0.0051
Median 0.0016 0.0060 0.0318 0.0000 -0.0350 -0.0200 -0.0300 0.0000 -0.0600
Maximum 10.95 11.84 10.79 10.43 16.54 12.71 21.92 16.94 22.64
Minimum -9.46 -11.11 -8.87 -8.20 -17.36 -12.96 -15.05 -10.24 -13.98
Std. Dev. 1.392 1.968 1.696 1.605 1.747 1.674 1.718 1.490 1.918
Skewness -0.1095 0.0819 0.0568 0.0115 0.3423 0.2127 1.4111 1.3166 1.9082
Kurtosis 11.916 6.859 7.786 8.001 23.712 14.266 25.513 23.133 30.589
J-Bera 7199 1350 2074 2263 38867 11504 46592 37312 70202
Prob 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

1 -0.10*** -0.07*** -0.04* -0.04* -0.12*** -0.04* 0.04** -0.00 -0.03

2 -0.07*** -0.06*** -0.02 -0.04** -0.12*** -0.11*** -0.06*** -0.11*** -0.10***

3 0.05*** 0.02*** -0.03 -0.07*** 0.03*** 0.03*** -0.10*** -0.02*** -0.09***

ADF -37.51*** -36.50*** -48.55*** -22.71*** -28.14*** -37.60*** -24.08*** -24.40*** -23.85***
No. Obs 2172 2172 2172 2172 2172 2172 2172 2172 2172

This table reports the descriptive statistics for the daily percentage continuously compounded returns on S&P 500, NASDAQ
100, DAX 30, Dow Jones Euro STOXX 50 stock indices and and for the daily percentage changes in the VIX, VXN, VDAX,
VDAXO and VSTOXX volatility indices. The autocorrelation coefficients , the Jarque-Bera and the Augmented Dickey-Fuller
(ADF) (an intercept is included in the test equation) test values are reported.
***, ** and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
 85

4 Methodology
4.1 Quantile Regression

In the ordinary least squares (OLS) regression where the conditional mean function,
the function that explains how the mean of dependent (or response) variable changes
with the vector of covariates (or independent variables)—simply this would be the
relationship we would know between dependent variable and covariates if we use least
squares regression i.e., the relationship at the mean of the dependent variable’s
distribution. What is more important to notice that it assumes that the error has the
same distribution whatever values may be taken by the components of the covariates
vector. Hence, this type of model is known as pure location-shift model because it
assumes that the covariates affect only the location (or mean) of the conditional
distribution of the response variable, not its scale (variance), or even shape (e.g.,
skewness or kurtosis) of the distribution. It implies that the mean regression analysis
would give incomplete picture of the asymmetric return-volatility relationship between
response variable (volatility changes) and covariates (negative and positive stock
returns) variables because it is focusing on estimating rates of change in the mean of
the response variable distribution as function of set of covariate variables.
Consequently, this mean regression method fails for a regression model with
heterogeneous variances implies that there are different rates of changes not just a
single rate of change in the mean which characterizes the probability distributions i.e.,
in our case volatility changes. Therefore, the mean regression methods which accounts
only changes in the means would underestimate or overestimate a response variable
with heterogeneous distributions ( see e.g., Cade et al., 1999; Koenker and Hallock,
2001; Cade and Noon, 2003; Koenker, 2005).

To account for the mentioned issues in regression method, Koenker and Bassett (1978)
proposed the quantile regression method as an extension of the mean regression
method for estimating rates of change in all parts of the distribution of a response
variable. The quantile regression model is advantageous because it does not assume
that the random part of the model, that is, the error of the model follow any specific
distribution. Hence, it overcomes most of the estimation issues related to the mean
regression model for instance it produces robust estimates in presence of non-normal
error distribution, outliers, and also account for omitted variables bias. In fact, the
 86

quantile regression model can be viewed as the generalization of mean regression

model to a collection of models to different conditional quantile functions.

The mathematical illustration is, as we know that the sample mean as the solution to
the problem of minimizing a sum of squared residuals in ordinary least squares
regression. For example least squares regression model if we have random sample
^y1 , y 2 ,˜ ˜ ˜, y n `, we solve

n 2

min ¦  y i  P  , (1)
Pƒ
i 1

The sample mean is obtained which is an estimate of the unconditional population

mean EY . However, the estimate of conditional expectation function E Y x   is

obtained by replacing the scalar P by a function P  x, E  in the equation 1 and solving

n 2

min ¦  y i  P  xi , E  (2)
E ƒ
i 1

Similarly, in a quantile regression, the median (quantile) as the solution to the problem
of minimizing a sum of absolute residuals, where the weights are symmetric. However,
the other conditional quantile functions are estimated by minimizing asymmetrically
weighted sum of absolute errors, where the weights are functions of the corresponding
quantile (see Koenker and Hallock, 2001; Barnes and Hughes, 2002; Koenker, 2005).
Thus, to obtain the conditional quantile regression estimator as the solution to the
minimization problem

n
min ¦ U q  y i  [  x i , E  (3)
E ƒ
i 1

Or equivalently

ª º
min « ¦ q y i  [  xi , E   ¦ 1  q  y i  [ xi , E  » (4)
E ƒ
¬ i : y i t[ i: yi d[ ¼

The minimization can be done very efficiently by linear programming methods. The
beauty of quantile regression is that all observations are used to estimate each quantile.
Finally, the standard errors can be computed by employing bootstrap method.
 87

4.2 Quantile Regression Model for Asymmetric Return-Volatility

Relationship

We present a quantile regression model for assessing the negative asymmetric

relationship between the returns on the stock index and changes in the volatility index.
This model is the generalization of the standard mean-regression models of Simon
(2003), Giot (2005) and Hibbert et al. (2008), who have empirically confirmed the
asymmetric return-volatility relationship.64 However, this essay extends these standard
mean-regression models (MRM) by modeling the asymmetric return-volatility
relationship using the conditional quantile regression model (QRM) to examine how
negative and positive stock index returns vary across different quantiles of IV changes,
i.e., how much this asymmetric relationship tends to change across different quantiles
of IV changes. Before specifying our quantile-regression model for the asymmetric
return-volatility relationship, we first specified a MRM model similar to that of Simon
(2003), Giot (2005) and Hibbert et al. (2008), which is considered a standard model in
our analysis.65

For instance, we regressed the daily volatility index changes (denoted VI it , where

i=VIX, VXN, VDAX, VSTOX) on the daily percentage continuously compounded

returns of the stock market index (denoted R it , where i=S&P 500, NASDAQ, DAX, DJ

Euro STOXX50), where Rit was used for positive returns and Rit for negative returns.

For the positive returns, Rit Rit if Rit ! 0 , and Rit 0 otherwise. On the other

hand, for the negative returns, Rit Rit if Rit  0 , and Rit 0 otherwise. The
standard MRM for assessing the negative asymmetric return-volatility relation thus has
the form
3 3 3
VI it   ¦  iL VI it  L  ¦
iL Rit L  ¦ iL Rit L  u t , (5)
L 1 L 0 L 0

64They showed that the relationship behaves differently for negative and positive stock index returns.
65Hibbert et al. (2008) segmented negative and positive stock returns into quantiles and then used least
squares for each quantile, which could not yield the robust results that we can find using quantile
regression, i.e., the effects of negative and positive returns on the upper and lower quantiles of the
dependent variable would be much different and robust using quantile regression instead of least-squares
regression (for a detailed discussion, see Heckman, 1979; Koenker and Hallock, 2001; Bassett and Chen,
2001).
 88

where  is the intercept,  iL represents the coefficients for the lagged IV changes in a

volatility index i , where L 1 to 3 ,

iL represents the coefficients for positive stock

returns and iL the coefficients for the negative returns of a stock market index i ,

where L 0 to 3 for both types of returns; and the errors u t are independently
identically distributed (iid) with zero means.66 Consequently, the standard MRM
assumes that the effects of both types of returns are static across different IV changes
(i.e., response variables); therefore, an MRM would miss important information across
quantiles of IV changes that we could otherwise detect using a QRM, particularly to
determine how the median or perhaps the 5th or 95th percentiles of the response
variable IV changes are affected by negative and positive stock return variables
(regressor variables).67 Koenker and Bassett (1978) were the first to introduce quantile
regression that could effectively model the uppermost quantiles.68 QRM is a
generalization of the MRM and is thereby a robust regression, especially in situations
where errors are non-normally distributed, i.e., are skewed and leptokurtic.
Nonetheless, the QRM is used for examining the asymmetric return-volatility
relationship; for instance, the qth QRM, which is a generalization of equation (5), has
the form
3 3 3
VI it  q   ¦  iLq  VI it  L  ¦
iLq  Rit L  ¦ iLq  Rit L  u t (6)
L 1 L 0 L 0

Where   q  is the intercept;  iL(q) represents the coefficients for the lagged IV changes in
q 
a volatility index i , where L 1 to 3 ;  iL represents the coefficients for positive

returns and iLq  the coefficients for negative returns of a stock market index i , where

L 0 to 3 for both type of returns; and the errors u t are assumed to be independent

from an error distribution  q (u t ) with the qth quantile equal to zero. Equation (6)

implies that the qth conditional quantile of the dependent variable VI i given

VI it 1 ,Vit  2 ,Vit 3 ,Rit ,Rit1 ,Rit 2 ,Rit3 ,Rit ,Rit1 ,Rit 2 ,Rit3 and denoted

66 A similar specification as of Hibbert et al. (2008) and Fleming et al. (1995) is used, which is consistent
with the dynamics of stochastic volatility. Moreover, the lags of negative and positive returns are inclduded
in order to caputre the leverage effect.
67 See a good discussion on this issue in Meligkotsidou et al. (2009).
68 Koenker (2005) provides mathematical details on the different versions of the quantile regression

models.
 89


Qq VI i VI it 1 ,..... it  3 
,Rit ,..,Rit3 ,Rit ,..,Rit3 , is equal to
3 3 3
 q   ¦  iLq  VI it  L  ¦
iLq  Rit L  ¦ iLq  Rit L . The main feature of this quantile
L 1 L 0 L 0

q 
regression framework is that the effects of the variables captured by  iL(q) ,
iL ,and

iLq  vary for each qth quantile within the range q  0,1  . Furthermore, the framework

allows for heteroskedasticity in error u t , and the coefficients are different for different

quantiles. Consequently, a quantile regression provides a broader set of information

about volatility changes here (i.e., the effects on all parts of the distribution of volatility
changes) than OLS regression would, particularly when the error distribution is not
symmetric.69 QRM is thus estimated for the sample period from February 2, 2001
through May 29, 2009 using the quantile regression method proposed by Koenker and
Bassett (1978), which minimizes the asymmetric sum of absolute residuals and robustly
models the conditional quantiles of the response variable, i.e., in our case, changes in
the volatility index:70

ª
min« ¦  ˆ  q VIit  ˆ  ˆiL VIit L 
ˆiL RitL  ˆiL RitL
¬«t::it tˆ  iL VIit  L 
ˆiL Rit  L  iL Rit  L
ˆ
(7)
º
 ¦ (1  q) VIit  ˆ  ˆ iL VIit  L 
ˆiL RitL  ˆiL Rit L » ˜
t::it ˆ  ˆiL VIit  L 
ˆiL Rit L  ˆiL Rit L ¼»

69 Because the differences between the mean and the median produce asymmetric distributions, see, for a
more detailed explanation, Meligkotsidou et al. (2009).
70 For a discussion of quantile models and their estimation techniques, see Koenker (2005).
 90

5 Empirical Results

Figure2 provides quantile regression results for S&P 500 returns with the VIX index
changes, where we have 11 covariates and an intercept. For each of the 12 coefficients,
19 quantile regression estimates were plotted for q ranging along
q  (0.05,0.1, .....,0.9, 0.95) as the solid curve (blue) with circles. In each plot on the x-
axis, we have a quantile (or q) scale, and the y-axis indicates the covariate effect as a
percentage. For each covariate, these estimates could be interpreted as the effect of a
percentage-point change of the covariate on the volatility, holding other covariates
unchanged. The two red-dotted lines show the conventional 95% confidence level for
the quantile regression estimates.

However, more detailed results for the important upper and lower quantiles of
estimates from Figure 2 are provided in Table 2, including corresponding t-statistics (in
parentheses) for each of the estimates therein. The standard errors were obtained using
the bootstrap method; therefore, robust t-statistics were obtained for each of the
quantile estimates. On the other hand, for the OLS estimates, the standard errors were
made heteroskedasticity-consistent using Newey-West (1987) correction. As the aim
was to quantify the asymmetric return-volatility relationship, we limit our discussion to
the positive and negative returns covariates, especially to capturing the
contemporaneous effects. When we look at the estimated coefficients of covariates

SP500Rt and SP500Rt in Columns 6 and 10, respectively, which represent the
contemporaneous return-volatility relationship; it is apparent from the absolute
difference that there are asymmetric effects for all quantile regression estimates,
including OLS estimates (here, OLS estimates are merely provided for comparative

purposes). The absolute values of SP500Rt

 are higher than the absolute values of

SP500Rt . Moreover, the Wald test for coefficients was applied in order to find the
q  q 
statistical difference between the coefficients
t and t in equation 6. The null

hypothesis (i.e., the coefficients for negative and positive returns are equal) for the
Wald test was significantly rejected for each of the quantile regression estimates.71
These results imply an asymmetric return-volatility relationship, indicating that the
negative returns for the stock index are linked to much higher volatilities for the VIX

71 Wald tests results are not reported here to save space.

 91

index than those linked to positive returns. More specifically, looking at each row of
Table 2 (i.e., each quantile of estimates), the results indicate that the impacts of the
negative and positive S&P 500 index returns on the VIX are highly asymmetric, with
both contemporaneous coefficients being statistically significant at the 1% significance
level. The mean or OLS regression estimates are quite similar to the q = 0.5 (median)-
quantile regression estimates; however, the changing nature of the estimates at the
other quantiles provides an interesting picture of how the distribution of IV changes
depends on the positive and negative returns variables and lagged IV changes variables.

The absolute value of SP500Rt monotonically increases when moving from a median

quantile to an upper quantile; i.e., the marginal effect of the negative returns is larger in
upper quantiles (i.e., q=0.95%), and vice versa for positive returns.72 As a result, OLS
underestimates the magnitude of these effects for the highest quantiles and
overestimates for the lowest quantiles.

In detail, the coefficient estimates with q = 0.5 or median (and OLS) for the

SP500Rt variable imply that a 1% decline in S&P 500 returns is linked to a 1.040%

(1.185%) increase in the VIX level, whereas the coefficient estimates for the SP500Rt

variable imply that a 1% increase in S&P 500 returns is linked to a 0.795% (0.864%)
decrease in the VIX level.73 However, in the coefficient estimates for quantile q=0.95,

the SP500Rt
 variable implies that a 1% decline in S&P 500 returns is linked to a

1.646% increase in the VIX, whereas the coefficient estimates for the SP500Rt variable


imply that a 1% increase in the S&P 500 returns is linked to a 0.401% decrease in the
VIX level. Obviously, it is apparent from the quantile regression results that the
asymmetry is much smaller in the lower and median quantiles of the distribution and
noticeably higher in the upper quantiles of the distribution. Thus, the OLS estimate,
which simply captures the mean effect, does a poor job of accounting for this
asymmetry in the upper quantiles.74

72 The equality of the coefficients across quantiles was formally tested using the Wald test. The test results

significantly rejected the null hypothesis of equality of the coefficients (particularly the contemporaneous
negative and positive returns) across quantiles; the Wald test is reported in Table 2.
73 Mean-regression model (or OLS) estimates are provided in parentheses for comparison.
74 The standard mean-regression models of Simon (2003), Giot (2005) and Hibbert et al. (2008) for the

asymmetric return-volatility relationship ignore the higher effects of negative and positive returns on the
upper quantiles of the IV changes’ distribution.
 92

Dependent Variable VIX-----Quantile Process Estimates

Intercept VIX(-1) VIX(-2) VIX(-3)
.6 .15 .2 .2

.10
.4
.1 .1
.05
.2 .00
.0 .0
.0 -.05

-.10
-.1 -.1
-.2
-.15

-.4 -.20 -.2 -.2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

SP500R+ SP500R(-1)+ SP500R(-2)+ SP500R(-3)+

0.0 .4 .6 .4

.4
.2 .2
-0.4
.2
.0 .0
-0.8 .0
-.2 -.2
-.2
-1.2
-.4 -.4
-.4

-1.6 -.6 -.6 -.6

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

SP500R- SP500R(-1)- SP500R(-2)- SP500R(-3)-

-0.4 .6 .6 .4

.4
-0.8 .4 .2

.2
-1.2 .2 .0
.0

-1.6 .0 -.2
-.2

-2.0 -.4 -.2 -.4

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

Figure 2. Quantile Regression Estimates: Dependent variable VIX changes.

Table 2. Quantile Regression Results: Response variable VIX changes
Quantile Intercept VIX t 1 VIX t  2 VIX t 3 SP 500 R 
t SP 500 R 
t 1 SP 500 R 
t 2 SP 500 R 
t 3 SP 500 R 
t SP 500 R 
t 1 SP 500 R 
t 2 SP 500 R 
t 3

0.05 -0.231*** -0.068 -0.025 -0.021 -1.301*** -0.279*** -0.242** -0.281*** -0.636*** 0.313*** 0.344*** 0.151
(-3.21) (-1.39) (-0.37) (-0.45) (-11.65) (-3.17) (-2.51) (-3.19) (-6.72) (3.33) (2.86) (1.63)
0.10 -0.219*** -0.030 -0.016 -0.005 -1.045*** -0.190*** -0.145* -0.224** -0.800*** 0.271*** 0.314*** 0.166**
(-4.19) (-0.68) (-0.29) (-0.11) (-13.56) (-2.94) (-1.95) (-2.82) (-12.98) (3.48) (3.07) (2.31)
0.15 -0.157*** -0.046 -0.024 0.044 -1.066*** -0.126** -0.182*** -0.089 -0.802*** 0.186*** 0.260*** 0.153**
(-3.32) (-1.12) (-0.52) (1.16) (-17.64) (-2.04) (-2.84) (-1.38) (-15.42) (2.98) (3.13) (2.31)
0.20 -0.134*** -0.062* -0.035 0.027 -0.970*** -0.125** -0.141** -0.078 -0.860*** 0.119** 0.233*** 0.147***
(-3.03) (-1.81) (-0.88) (0.82) (-14.11) (-2.56) (-2.49) (-1.42) (-17.51) (2.28) (3.75) (2.61)
0.25 -0.115*** -0.078** -0.006 0.032 -0.943*** -0.137*** -0.068 -0.061 -0.888*** 0.097** 0.182*** 0.167***
(-3.04) (-2.56) (-0.17) (1.16) (-16.18) (-3.41) (-1.42) (-1.36) (-22.78) (2.00) (3.70) (3.53)
Median -0.010 -0.066** 0.012 0.012 -0.795*** -0.087** 0.029 0.031 -1.040*** 0.047 0.090* 0.061
(-0.35) (-2.22) (0.38) (0.48) (-23.86) (-2.14) (0.87) (1.02) (-29.45) (1.03) (1.87) (1.44)
0.75 0.065* -0.054* -0.023 0.024 -0.579*** 0.064* 0.076 0.063 -1.285*** 0.012 0.008 0.058
(1.70) (-1.69) (-0.61) (0.83) (-12.71) (1.72) (1.44) (1.25) (-32.60) (0.22) (0.16) (1.42)
0.80 0.072* -0.074** -0.007 0.009 -0.541*** 0.047 0.116** 0.111** -1.335*** -0.043 -0.004 0.034
(1.92) (-2.23) (-0.17) (0.29) (-9.99) (1.42) (2.01) (2.14) (-37.98) (-0.76) (-0.07) (0.72)
93

0.85 0.149*** -0.090** 0.020 0.013 -0.472*** 0.023 0.145** 0.105** -1.371*** -0.073 -0.002 0.051
(3.25) (-2.39) (0.42) (0.36) (-8.80) (0.58) (2.34) (2.17) (-24.23) (-1.06) (-0.03) (0.76)
0.90 0.267*** -0.079* 0.018 0.021 -0.451*** 0.011 0.185** 0.035 -1.533*** -0.185** -0.026 0.055
(4.90) (-1.69) (0.36) (0.52) (-8.76) (0.21) (2.54) (0.64) (-23.02) (-2.11) (-0.38) (0.70)
0.95 0.373*** -0.025 0.048 0.066 -0.401*** 0.133 0.243*** 0.105 -1.646*** -0.162 -0.005 0.013
(5.02) (-0.39) (0.86) (1.16) (-5.67) (1.57) (2.98) (1.43) (-14.93) (-1.45) (-0.06) (0.12)
OLS 0.000 -0.101** -0.033 -0.034 -0.864*** -0.089* -0.039 -0.079 -1.185*** -0.017 0.066 0.100
(0.01) (-2.27) (-0.48) (-0.69) (-10.75) (-1.71) (-0.70) (-1.15) (-24.19) (-0.18) (0.59) (1.61)
Quantile Slope Equality Test Results: Only significant results of asymmetry are reported with the corresponding quantiles q  1  q  .
0.2-0.4*** 0.2-0.4** 0.2-0.4*** 0.2-0.4*** 0.2-0.4*
0.4-0.5** 0.4-0.5*
0.5-0.6* 0.5-0.6** 0.5-0.6*** 0.5-0.6**
0.6-0.8*** 0.6-0.8*** 0.6-0.8*** 0.6-0.8*** 0.6-0.8**
The table reports results from the Quantile Regression and OLS Regression of the VIX changes on a set of variables; specifications 2 and 1 are estimated. T-statistics are provided in parentheses. ***,
**, and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
 94

Figure 3 presents quantile regression results for NASDAQ 100 returns with the VXN
index changes, and the important full-sample daily upper and lower quantiles’ results
are presented in Table 3.75 The results are qualitatively similar to those found for the
S&P 500 and VIX asymmetric relationship.76 The major difference lies in the lower
asymmetric responses of the covariates (i.e., negative and positive returns) across
different quantiles of the VXN distribution in comparison with the VIX results.77
Furthermore, the significance of covariates is lower for the VXN than for the VIX. The
finding is consistent with both Giot (2005) and Hibbert et al. (2008) in that during
volatile periods, option traders react less aggressively to negative returns. As the
NASDAQ is a tech index, it inherently presents a higher volatility than the S&P 500;
therefore, the conclusion drawn by Giot (2005) and Hibbert et al. (2008) can be
applied to the NASDAQ results.78

On the other hand, Figures 4 and 5 provide quantile regression results for DAX 30
returns with the VDAX changes and the VDAXO changes, respectively; the important
daily upper and lower quantile results are reported in Table 4 and Table 5.79 The
quantile results for the DAX 30 returns with both the VDAX changes and VDAXO
changes are discussed simultaneously in order to compare the asymmetric responses of
both volatility indices to the same negative and positive returns of the DAX 30 index.80
 
The coefficients of the covariates DAXRt and DAXRt are shown in Columns 6 and

10, respectively, in both tables; the coefficients represent contemporaneous return-

volatility relationships. Based on the absolute difference in the coefficients’ values, it is
clear that there are asymmetric effects for all quantile regression estimates, including

the OLS estimates. The absolute values of DAXRt are higher than the absolute values


of DAXRt . Similarly, the null hypothesis of the Wald test that the coefficients
t and
q 

75 Figure 3 and Table 3 are provided in Appendix A.

76 A detailed discussion on these results is avoided merely for space considerations.
77The Wald test for equality of the coefficients across quantiles is formally tested and reported in Table 3.

Here, too, the test results significantly reject the null hypothesis of equality of the coefficients (particularly
the contemporaneous negative and positive returns) across quantiles.
78 For more discussion on this point, see Hibbert et al. (2008).

79 Figures 4 and 5 and Tables 4 and 5 are provided in Appendix A.

80 Remember that VDAX is the MFIV index that incorporates volatility smirk (skew), whereas VDAXO is

the BSIV index that does not account for volatility smirk (skew). Unfortunately, for the comparison of the
two measures we are restricted to only the DAX 30 stock index. For the other stock indices, S&P 500,
NASDAQ 100 and DJ Euro STOXX, we have no active BSIV volatility index; although the VXO, an active
BSIV volatility index, is available, it cannot be compared because it is implied from the options on the S&P
100 index.
 95

tq  in equation 6 are equal is significantly rejected for each of the quantile regression
estimates. Furthermore, looking at each row of Tables 4 and 5 (i.e., each quantile
result), the results indicate that the impacts of the negative and positive DAX 30 index
returns on VDAX and VDAXO are asymmetric, with both contemporaneous coefficients
being statistically significant at the 1% significance level. The coefficient estimates for

q=0.5 or the median (and OLS) for the DAXRt covariate imply that a 1% decline in

DAX 30 returns is linked to a 0.837% (1.022%) increase in the VDAX level and that a
similar decline in DAX 30 returns is linked to a 0.712% (0.778%) increase in VDAXO

level. On the other hand, the coefficient estimates for the DAXR t covariate imply that

a 1% increase in DAX 30 returns is linked to a 0.515% (0.400%) decrease in the VDAX

level, and a similar increase in DAX 30 returns is linked to a 0.543% (0.370%) decrease
in the VDAXO level. However, the coefficient estimates for quantile q=0.95 of the

DAXRt variable imply that a 1% decline in DAX 30 returns is linked to a 1.389%

(1.115%) increase in the VDAX (VDAXO), whereas the coefficient estimates for the

DAXRt covariate imply that a 1% increase in DAX 30 returns is linked to a 0.130%

(0.156%) decrease in the VDAX (VDAXO) level. For comparison, the coefficients of
 
covariates DAXRt and DAXRt are listed in Tables 4 and 5. It is very clear that the

effects of the negative and positive returns are considerably different. The VDAX (MFIV
index) responds in a very asymmetric fashion in comparison to its older counterpart,
the VDAXO (BSIV index), to similar negative and positive returns, implying
pronounced asymmetric return-volatility relationship with the MFIV index.
Furthermore, the asymmetric responses are most apparent in upper-quantile
estimates, where the asymmetry is very pronounced; i.e., the asymmetries in the
absolute differences are smaller in the lower and median quantiles of the distribution
and noticeably larger in the upper quantiles of the distributions.

Figure 6 presents’ quantile regression estimates for the DJ Euro STOXX 50 returns
with the VSTOXX changes, and the important daily upper and lower quantile results
are presented in Table 6.81 Similarly, the results here are qualitatively similar to those
found for the DAX 30 and VDAX asymmetric relationships. The major difference is the
slightly more asymmetric responses of the covariates (i.e., negative and positive
returns) across different quantiles of the VSTOXX in comparison to the VDAX results.

81 Figure 6 and Table 6 are provided in Appendix A.

 96

The main conclusion drawn from Tables 2 to 6 is that negative and positive stock index
returns trigger the volatility index to move in completely opposite directions and in an
asymmetric fashion; i.e., negative returns have a much greater impact on volatility than
do positive returns, particularly at the uppermost regression quantiles (e.g., q=0.95).
Our quantile regression model for the asymmetric return-volatility relation thus reveals
important missing information underestimated by the mean-regression model (OLS).
Second, our argument that the asymmetry with smirk (skew)-adjusted volatility (MFIV)
should present pronounced responses is confirmed by comparing the asymmetric
responses of VDAX (MFIV) and VDAXO (BSIV), where we found that the MFIV index
responded in a pronounced fashion in comparison with the BSIV index.82 Third, if we
look at the coefficients of the lag covariates of negative and positive returns, they are
mostly insignificant; we thus assert that at the daily level, the leverage hypothesis is
unable to quantify this strong asymmetric return-volatility relationship and that similar
conclusions could be drawn for the feedback hypothesis. Finally, the VIX volatility
index presents the strongest asymmetric return-volatility relationship, followed by the
VSTOXX, VDAX and VXN volatility indices, respectively.

82 As the MFIV volatility indices account for OTM puts, the asymmetry should be pronounced with each
MFIV volatility index. Because investors hedge their downside risk by taking positions in the OTM put
options, in periods of market turmoil there is greater buying demand for put options than for call options,
which leads to higher volatilities than those found during market rallies. Consequently, negative stock
index returns induce an increase in the levels of the volatility indices. Our results are also consistent with
the net-buying-pressure hypothesis of Bollen and Whaley (2004).
 97

6 Conclusion

We investigated the asymmetric return-volatility phenomenon in the newly adapted

robust volatility indices (i.e., the VIX, VXN, VDAX, VDAXO, and VSTOXX) using
quantile regression. In particular, we quantified the effects of positive and negative
stock index returns at different quantiles of IV changes’ distributions, asking about the
degree to which the asymmetric responses at the uppermost quantiles are comparable
with the responses of median (or mean) regressions. Additionally, as Bollen and
Whaley (2004) have documented, the net buying pressure for stock index put options
from institutional investors seeking to hedge their portfolios induces increases in IVs.
Likewise, new IV indices incorporate both OTM put and call options and are thus
highly informed and robust measures. Accordingly, they should present more
pronounced asymmetric return-volatility relationships in comparison to their older
counterparts.

There is noticeable evidence that the volatility indices VIX, VXN, VDAX, VDAXO and
VSTOXX from February 2001 through May 2009 responded in a pronounced
asymmetric fashion to the negative and positive returns of their corresponding stock
indices: the asymmetry monotonically increases when moving from the median
quantile to the uppermost quantile (i.e., 95%); therefore, OLS underestimates this
relationship at upper quantiles. These IV indices thus sharply rise during market
declines (fear) and fall during market rallies (exuberance). The VIX presents the
highest asymmetry, followed by the VSTOXX, VDAX and VXN volatility indices,
respectively. Second, our argument that asymmetry with the volatility smirk(skew)-
adjusted volatility index measure (MFIV) should be pronounced is confirmed by
comparing the asymmetric responses of VDAX (MFIV) and VDAXO (BSIV); the MFIV
index responds in a pronounced fashion in comparison with the BSIV index. Third, we
also confirmed that a significant amount of asymmetry occurs contemporaneously
rather than with a lag, thus rejecting the leverage hypothesis, and that a similar
conclusion can be drawn for the feedback hypothesis.

Our results have a number of implications. First, as we found that newly adapted
volatility indices are strongly negatively correlated with their corresponding stock
indices, the new volatility indices are important instruments for hedging stock
portfolios. Derivatives exchanges provide liquid markets for the futures and options
 98

underlying these volatility indices. Therefore, a position in futures or options on a

volatility index can more accurately hedge a stock portfolio position without
considering complicated stock index option trading strategies. Second, when the stock
index drops, the volatility index rises sharply. Therefore, new volatility indices are
useful not only for assessing potential risks, but also for speculative transactions by
risk-seeking investors. Third, since the new volatility indices are based on the robust
MFIV concept and provide better tradability, it is easier for issuers of derivatives to
engineer structured products based on the volatility indices. Fourth, trading strategies
with regard to range could generate profits; an example of this could be a volatility-long
position in decreasing volatility markets paired with a volatility-short position in
increasing volatilitymarkets.
 99

References

Alexander, C., 2001. Principal component analysis of implied volatility smiles and
skews. ISMA Discussion Paper, University of Reading,UK.
Badshah, I., 2008, Modeling the dynamics of implied volatility surfaces. European
Financial Management Association (EFMA)-2008, Athens, Greece, Conference
Proceedings.
Barnes, M., and Hughes, A., 2002. A quantile regression analysis of the cross section of
stock market returns. Working paper, Federal Reserve Bank of Boston, United
States.
Bassett,G., Chen, H., 2001. Quantile style: return-based attribution using regression
quantiles. Empirical Economics 26, 293-305.
Bekaert, G., and Wu, G., 2000. Asymmetric volatility and risk in equity markets.
Review of Financial Studies 13, 1-42.
Black, F., and Scholes, M., 1973. The pricing of options and corporate liability, Journal
of Political Economy 81, 636-654.
Black, F., 1976. Studies of stock market volatility changes. Proceedings of the American
statistical association, Business and Economic Statistics Section, 177-181.
Blair, B., Poon, S.H., and Taylor, S., 2001. Forecasting S&P 100 volatility: The
incremental information content of implied volatilities and high frequency
index returns. Journal of Econometrics 105, 5-27.
Bollen, N., and Whaley, R., 2004. Does net buying pressure affect the shape of the
implied volatility functions? Journal of Finance 59, 711-753.
Breen, D., Glosten, L., and Jagannathan, R., 1989. Economic significance of predictable
variations in stock index returns. Journal of Finance 44, 1177-1189.
Britten-Jones, M., and Neuberger, A., 2000. Option prices, implied processes, and
stochastic Volatility. Journal of Finance 55, 839-866.
Cade, B., and Noon, B., 2003. A gentle introduction to quantile regression for
ecologists. Frontiers in Ecologies and the Environment 1, 412-420.
Cade, B., Terrell, J., and Schroeder R., 1999. Estimating effects of limiting factors with
regression quantiles. Ecology 80, 311-323.
Camara, A., Heston, S., 2008. Closed form option pricing formulas with extreme
events. Journal of Futures Markets, 28, 213-230.
Campbell, J., and Hentschel, L., 1992. No news is good news: an asymmetric model of
changing volatility in stock returns. Journal of Financial Economics 31, 281-318.
Carr, P., and Wu, L., 2006. A tale of two indices. Journal of Derivatives 13, 13-29.
Chakravarty, S., Gulen, H., and Mayhew, S., 2004. Informed trading in stock and
option markets. Journal of Finance 59, 1235-1258.
Christensen, B.J., and Prabhala, N.R., 1998. The relation between implied and realized
volatility. Journal of Financial Economics 50, 125-150.
Christie, A., 1982. The stochastic behaviour common stock variances: Value, leverage,
and interest rate effects. Journal of Financial Economics 10, 407-432.
Day, T., and Lewis, C., 1992. Stock market volatility and the information content of
stock index options. Journal of Econometrics, 52, 267-287.
Dennis, P., Mayhew, S., and Stivers, C., 2006. Stock returns, implied volatility
innovations, and the asymmetric volatility phenomenon. Journal of Financial
and Quantitative Analysis 41, 381-406.
Dumas, B., Fleming, J., and Whaley, R., 1998. Implied volatility functions: Empirical
tests. Journal of Finance 53, 2059-2106.
Ederington, L., and Guan, W., 2002. Is implied volatility an informationally efficient
and effective predictor of future volatility? Journal of Risk, 4, 29-46.
Fleming, J., Ostdiek, B., and Whaley, R., 1995. Predicting stock market volatility: A new
Measure. Journal of Futures Markets 15, 265-302.
 100

Fleming, J., 1998. The quality of market volatility forecasts implied by S&P100 index
option prices. Journal of Empirical Finance 5, 317-345.
French, K., Schwert, W., and Stambaugh, R., 1987. Expected stock returns and
volatility. Journal of Financial Economics 19, 3-29.
Ghysels, E., Santa-Clara, P., and Valkanov, R., 2005. There is a risk-return tradeoff
after all. Journal of Financial Economics, 76, 509-548.
Giot, P., 2005. Relationships between implied volatility indices and stock index returns.
Journal of Portfolio Management 31, 92-100.
Goncalves, S., and Guidolin, M., 2006. Predictable dynamics in the S&P 500 index
options implied volatility surfaces. Journal of Business 79, 1591-1635.
Heckman, J., 1979. Sample selection bias as a specification error. Econometrica 47,
153-161.
Hibbert, A., Daigler, R., and Dupoyet, B., 2008. A behavioural explanation for the
negative asymmetric return-volatility relation. Journal of Banking and Finance
32, 2254-2266.
Jiang, G., and Tian, Y., 2005. The model-free implied volatility and its information
contents. Review of Financial Studies 18, 1305-1342.
Koenker, R., 2005. Quantile regressions. First edition, Cambridge University Press,
New York, United States.
Koenker, R., Bassett, G., 1978. Regression quantiles. Econometrica 46, 33-50.
Koenker, R., Hallock, K., 2001. Quantile Regression. Journal of Economic Perspectives
15, 143-156.
Liu, J., Pan, J., and Wang, T., 2005. An equilibrium model of rare-event premia and its
implication for option smirks. Review of Financial Studies 18, 131-164.
Low, C., 2004. The fear and exuberance from implied volatility of S&P 100 index
options. Journal of Business 77, 527-546.
Martens, M., and Zein, J., 2004. Predicting financial volatility: High-frequency time-
series forecasts vis-à-vis implied volatility. Journal of Futures Markets, 24,
1005-1028.
Mayhew, S., and Stivers, C., 2003. Stock return dynamics, option volume, and the
information content of implied volatility. Journal of Futures Markets, 23, 615-
646.
Meligkotsidou, L., Vrontos, I., Vrontos, S., 2009. Quantile regression analysis of hedge
fund strategies. Journal of Empirical Finance 16, 264-279.
Newey, W., and West, K., 1987. A simple positive definite heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica 55, 703-708.
Pan, J., 2002. The jump-risk premia implicit in options: evidence from an integrated
time series study. Journal of Financial Economics 63, 3-50.
Poon,S.,and Granger,C.,2003. Forecasting volatility in financial markets: A review.
Journal of Economic Literature 41, 478-539.
Poon, S., and Granger, C., 2005. Practical issues in forecasting volatility. Financial
Analyst Journal 61, 45-56.
Poterba, J., and Summers, L., 1986. The persistence of volatility and stock market
fluctuations. American Economic Review 76, 1142-1151.
Schwert, G.W., 1989. Why does stock market volatility change over time? Journal of
Finance 44, 1115-1153.
Schwert, G.W., 1990. Stock volatility and the crash of “87.” Review of Financial Studies
3, 77-102.
Simon, D., 2003. The Nasdaq volatility index during and after the bubble. Journal of
Derivatives 11, 9-24.
Skiadopoulos, G., 2004. The Greek implied volatility index: construction and
properties. Applied Financial Economics 14, 1187-1196.
Whaley, R., 2000. The investor fear gauge. Journal of Portfolio Management 26, 12-17.
 101

Appendix A: Quantile Regression Estimates

Dependent Variable VXN-----Quantil e Process Estimates

Intercept VXN(-1) VXN(-2) VXN(-3)
.8 .2 .20 .3

.15
.4 .1 .2
.10

.05
.0 .0 .1
.00

-.05
-.4 -.1 .0
-.10

-.8 -.2 -.15 -.1

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

NASDAQR+ NASDAQR(-1)+ NASDAQR(-2)+ NASDAQR(-3)+

0.4 .4 .4 .3

.2
0.0
.2 .2
.1
-0.4
.0 .0 .0
-0.8
-.1
-.2 -.2
-1.2
-.2

-1.6 -.4 -.4 -.3

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

NASDAQR- NASDAQR(-1)- NASDAQR(-2)- NASDAQR(-3)-

0.0 .6 .6 .6

.4
-0.4 .4 .4

.2
-0.8 .2 .2
.0

-1.2 .0 .0
-.2

-1.6 -.4 -.2 -.2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

Figure 3. Quantile Regression Estimates: Dependent Variable VXN changes.

Dependent Vari abl e VDAX-----Quantil e Process Esti mates

Intercept VDAX(-1) VDAX(-2) VDAX(-3)
.8 .3 .2 .2

.2 .1
.4 .1

.1 .0
.0 .0
.0 -.1

-.4 -.1
-.1 -.2

-.8 -.2 -.2 -.3

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

DAXR+ DAXR(-1)+ DAXR(-2)+ DAXR(-3)+

0.4 .4 .4 .4

.2 .2 .2
0.0

.0 .0 .0
-0.4
-.2 -.2 -.2

-0.8
-.4 -.4 -.4

-1.2 -.6 -.6 -.6

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

DAXR- DAXR(-1)- DAXR(-2)- DAXR(-3)-

-0.4 .6 .6 .6

-0.6
.4 .4
.4
-0.8
.2 .2
-1.0 .2
.0 .0
-1.2
.0
-.2 -.2
-1.4

-1.6 -.4 -.2 -.4

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

Figure 4. Quantile Regression Estimates: Dependent Variable VDAX changes.

 102

Dependent Variabl e VDAXO-----Quantil e Process Esti mates

Intercept VDAXO(-1) VDAXO(-2) VDAXO(-3)
.6 .2 .3 .3

.1
.4 .2 .2

.0
.2 .1 .1
-.1
.0 .0 .0
-.2

-.2 -.1 -.1

-.3

-.4 -.4 -.2 -.2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

DAXR+ DAXR(-1)+ DAXR(-2)+ DAXR(-3)+

0.2 .3 .3 .4

0.0 .2 .2
.2
-0.2 .1 .1

-0.4 .0 .0 .0

-0.6 -.1 -.1

-.2
-0.8 -.2 -.2

-1.0 -.3 -.3 -.4

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

DAXR- DAXR(-1)- DAXR(-2)- DAXR(-3)-

-0.2 .4 .6 .4

-0.4 .2
.3
.4
-0.6 .0
.2
-0.8 -.2 .2
.1
-1.0 -.4
.0
.0
-1.2 -.6

-1.4 -.8 -.2 -.1

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

Figure 5. Quantile Regression Estimates: Dependent Variable VDAXO changes.

Dependent Variabl e VST OXX-----Quantil e Process Esti mates

Intercept VSTOXX(-1) VSTOXX(-2) VSTOXX(-3)
.6 .3 .2 .2

.4
.2 .1
.1
.2
.1 .0
.0 .0
.0 -.1
-.2
-.1
-.1 -.2
-.4

-.6 -.2 -.3 -.2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

STOXXR+ STOXXR(-1)+ STOXXR(-2)+ STOXXR(-3)+

0.0 .4 .4 .4

-0.2 .2
.2 .2

-0.4 .0
.0 .0
-0.6 -.2
-.2 -.2
-0.8 -.4

-.4 -.4
-1.0 -.6

-1.2 -.6 -.8 -.6

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

STOXXR- STOXXR(-1) STOXXR(-2)- STOXXR(-3)-

-0.6 .8 .8 .6

-0.8 .6 .6
.4
-1.0 .4 .4

-1.2 .2 .2 .2

-1.4 .0 .0
.0
-1.6 -.2 -.2

-1.8 -.4 -.4 -.2

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Quantile Quantile Quantile Quantile

Figure 6. Quantile Regression Estimates: Dependent Variable VSTOXX changes.

Table 3. Quantile Regression Results: Dependent Variable VXN changes
Quantile Intercept        
VXN t 1 VXN t  2 VXN t 3 NASDAQRt NASDAQR t 1 NASDAQRt  NASDAQRt 3 NASDAQRt NASDAQR t 1 NASDAQRt 2 NASDAQRt 3

0.05 -0.320** -0.004 0.001 0.036 -1.049*** -0.204** -0.190** -0.145* -0.240*** 0.258** 0.333*** 0.258**
(-2.50) (-0.06) (0.02) (0.58) (-10.97) (-2.17) (-1.97) (-1.84) (-3.01) (2.51) (4.07) (2.36)
0.10 -0.265*** 0.027 -0.002 0.037 -0.854*** -0.147** -0.114* -0.107*** -0.281*** 0.150** 0.253*** 0.125**
(-3.48) (0.70) (-0.04) (0.97) (-16.82) (-2.19) (-1.66) (-2.60) (-6.36) (2.47) (3.95) (2.23)
0.15 -0.191*** 0.015 -0.009 0.021 -0.825*** -0.143** -0.068 -0.077 -0.330*** 0.124** 0.193*** 0.098**
(-2.79) (0.39) (-0.24) (0.61) (-14.72) (-2.49) (-1.53) (-1.60) (-7.57) (2.22) (3.45) (2.23)
0.20 -0.194*** 0.002 -0.027 0.050 -0.741*** -0.110*** -0.056* -0.036 -0.392*** 0.101** 0.103** 0.092**
(-3.44) (0.05) (-0.71) (1.59) (-13.01) (-2.69) (-1.72) (-0.82) (-9.22) (2.06) (2.26) (2.49)
0.25 -0.166*** 0.027 -0.025 0.039 -0.675*** -0.096*** -0.045 -0.027 -0.420*** 0.081* 0.078* 0.075**
(-3.23) (0.73) (-0.72) (1.35) (-15.10) (-2.64) (-1.44) (-0.78) (-11.49) (1.92) (1.93) (2.12)
Median -0.022 -0.029 -0.009 0.019 -0.487*** -0.057*** -0.012 0.014 -0.562*** -0.020 0.038 0.025
(-0.63) (-1.25) (-0.40) (0.84) (-16.02) (-2.60) (-0.53) (0.65) (-21.45) (-0.77) (1.58) (0.74)
0.75 0.059 -0.041 -0.035 0.072** -0.301*** -0.012 0.071* 0.083** -0.812*** -0.068 -0.047 0.014
(1.06) (-1.18) (-1.09) (2.36) (-7.56) (-0.35) (1.80) (2.04) (-18.79) (-1.59) (-1.24) (0.36)
0.80 0.100 -0.084** -0.022 0.081*** -0.264*** -0.003 0.112*** 0.091** -0.849*** -0.143*** -0.056 0.002
(1.62) (-2.23) (-0.58) (2.63) (-5.36) (-0.09) (2.76) (2.11) (-17.23) (-2.91) (-1.37) (0.05)
0.85 0.212*** -0.098** -0.006 0.058* -0.229*** -0.014 0.108*** 0.127*** -0.888*** -0.123*** -0.043 -0.045
(3.34) (-2.47) (-0.16) (1.83) (-4.89) (-0.35) (3.25) (3.18) (-13.42) (-2.59) (-1.05) (-1.16)
0.90 0.298*** -0.087** -0.012 0.055 -0.149*** 0.031 0.085** 0.112** -1.098*** -0.119** -0.053 -0.063
103

(3.84) (-2.07) (-0.26) (1.41) (-2.85) (0.68) (2.06) (2.41) (-15.75) (-2.06) (-1.12) (-1.26)
0.95 0.528*** -0.053 0.012 0.094 -0.106* 0.100 0.185** 0.098 -1.167*** -0.169** -0.057 -0.024
(4.65) (-0.97) (0.17) (1.54) (-1.70) (1.58) (1.99) (1.12) (-19.99) (-1.99) (-0.86) (-0.28)
OLS -0.048 -0.035 -0.052 0.045 -0.497*** -0.060 -0.036 0.009 -0.642*** -0.029 0.000 0.057
(-0.91) (-1.07) (-1.31) (1.27) (-8.38) (-1.40) (-0.93) (0.23) (-12.59) (-0.59) (0.00) (1.47)
Quantile Slope Equality Tests Results: Only significant results of asymmetry are reported with the corresponding quantiles q  1  q  .
0.2-0.4*** 0.2-0.4* 0.2-0.4* 0.2-0.4*** 0.2-0.4** 0.2-0.4**
0.4-0.5* 0.4-0.5*** 0.4-0.5*** 0.4-0.5*
0.5-0.6** 0.5-0.6*** 0.5-0.6**
0.6-0.8*** 0.6-0.8* 0.6-0.8** 0.6-0.8*** 0.6-0.8*** 0.6-0.8***

The table reports results from the Quantile Regression and the OLS Regression of the VXN changes on a set of variables; specifications 2 and 1 are estimated. T-statistics are provided in parentheses.
***, ** and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
Table 4. Quantile Regression Results: Dependent Variable VDAX changes
Quantile Intercept VDAX t 1 VDAX t  2 VDAX t  3 DAXR 
t DAXR 
t 1 DAXR 
t 2 DAXR 
t 3 DAXR 
t DAXR 
t 1 DAXR 
t 2 DAXR 
t 3

0.05 -0.335*** 0.147*** -0.048 -0.113** -0.880*** -0.284*** -0.252*** -0.312*** -0.642*** 0.277*** 0.293*** 0.249**
(-3.84) (2.86) (-0.78) (-2.12) (-10.23) (-3.06) (-3.47) (-5.09) (-13.22) (2.82) (3.15) (2.34)
0.10 -0.276*** 0.102** -0.010 -0.143*** -0.774*** -0.156** -0.183*** -0.298*** -0.661*** 0.198*** 0.263*** 0.155*
(-3.70) (2.25) (-0.21) (-2.88) (-12.55) (-2.13) (-2.69) (-6.12) (-12.92) (2.75) (3.76) (1.83)
0.15 -0.201*** 0.064 0.002 -0.125*** -0.710*** -0.164*** -0.127** -0.298*** -0.699*** 0.119** 0.242*** 0.069
(-2.95) (1.58) (0.06) (-2.74) (-12.51) (-2.60) (-2.15) (-5.37) (-17.34) (2.14) (4.29) (1.01)
0.20 -0.179*** 0.046 -0.020 -0.092** -0.666*** -0.146** -0.114*** -0.185*** -0.711*** 0.104** 0.185*** 0.088*
(-3.19) (1.32) (-0.56) (-2.56) (-15.55) (-2.43) (-2.65) (-3.83) (-21.31) (2.35) (3.98) (1.89)
0.25 -0.156*** 0.031 0.012 -0.098*** -0.601*** -0.114** -0.088** -0.169*** -0.732*** 0.089** 0.201*** 0.059*
(-3.27) (1.02) (0.39) (-3.33) (-15.15) (-2.17) (-2.43) (-3.86) (-22.10) (2.52) (4.92) (1.66)
Median -0.003 0.015 0.010 -0.044 -0.515*** -0.004 -0.023 -0.028 -0.837*** 0.052 0.121*** 0.041
(-0.08) (0.48) (0.34) (-1.46) (-15.59) (-0.11) (-0.76) (-0.80) (-27.26) (1.52) (3.21) (1.04)
0.75 0.087** 0.043 0.027 -0.033 -0.361*** 0.103** -0.010 0.012 -1.044*** -0.007 0.132*** -0.020
(2.07) (1.18) (0.87) (-1.08) (-8.74) (2.26) (-0.33) (0.29) (-28.52) (-0.15) (2.89) (-0.56)
0.80 0.095* 0.021 0.040 -0.038 -0.291*** 0.143*** 0.005 0.017 -1.126*** -0.043 0.109** -0.003
(1.86) (0.54) (0.97) (-1.02) (-6.70) (3.35) (0.12) (0.37) (-23.52) (-0.83) (2.04) (-0.08)
0.85 0.161*** 0.009 0.021 -0.032 -0.288*** 0.127*** 0.037 0.036 -1.198*** -0.073 0.058 0.020
(2.95) (0.24) (0.51) (-0.78) (-6.55) (3.27) (0.81) (0.76) (-22.19) (-1.40) (1.04) (0.43)
0.90 0.246*** 0.022 0.026 -0.036 -0.259*** 0.125*** 0.065 0.035 -1.271*** -0.049 0.010 -0.018
104

(4.07) (0.49) (0.56) (-0.68) (-4.51) (3.08) (1.18) (0.60) (-20.48) (-0.83) (0.16) (-0.26)
0.95 0.457*** 0.022 0.038 -0.048 -0.130 0.099* 0.051 0.025 -1.389*** -0.136 0.028 -0.029
(4.75) (0.33) (0.59) (-0.70) (-1.28) (1.70) (0.62) (0.27) (-17.19) (-1.22) (0.33) (-0.25)
OLS -0.111 0.080 0.018 -0.083* -0.400*** -0.019 -0.063 -0.080 -1.022*** 0.031 0.160*** 0.080
(-1.57) (1.48) (0.32) (-1.92) (-4.91) (-0.44) (-1.30) (-1.00) (-14.55) (0.32) (3.27) (1.26)
Quantile Slope Equality Tests Results: Only significant results of asymmetry are reported with the corresponding quantiles q  1  q  .
0.2-0.4*** 0.2-0.4** 0.2-0.4*** 0.2-0.4***
0.4-0.5*** 0.4-0.5* 0.4-0.5* 0.4-0.5* 0.4-0.5**
0.5-0.6** 0.5-0.6*
0.6-0.8*** 0.6-0.8*** 0.6-0.8*** 0.6-0.8*

The table reports results from the Quantile Regression and the OLS Regression of the VDAX changes on a set of variables; specifications 2 and 1 are estimated. T-statistics are provided in
parentheses. ***, ** and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
Table 5. Quantile Regression Results: Dependent Variable VDAXO changes
Quantile Intercept VDAXO t 1 VDAXO t  2 VDAXOt 3 DAXR 
t DAXR 
t 1 DAXR 
t 2 DAXR 
t 3 DAXR 
t DAXR 
t 1 DAXR 
t 2 DAXR 
t 3

0.05 -0.192*** 0.099** 0.050 -0.082 -0.766*** -0.207*** -0.144** -0.181** -0.509*** 0.221*** 0.436*** 0.171**
(-2.64) (2.57) (0.66) (-1.38) (-11.78) (-4.96) (-2.24) (-2.52) (-7.13) (3.57) (5.02) (2.27)
0.10 -0.109** 0.023 0.091** -0.111*** -0.759*** -0.167*** -0.087** -0.207*** -0.517*** 0.123*** 0.370*** 0.115**
(-2.37) (0.73) (2.20) (-3.08) (-19.04) (-4.25) (-2.53) (-6.12) (-11.91) (2.83) (6.66) (2.39)
0.15 -0.085* 0.006 0.036 -0.126*** -0.743*** -0.112** -0.060 -0.193*** -0.528*** 0.111** 0.242*** 0.087*
(-1.82) (0.15) (0.87) (-3.61) (-20.72) (-2.57) (-1.41) (-5.46) (-13.72) (2.45) (4.89) (1.91)
0.20 -0.061 0.021 0.018 -0.114*** -0.719*** -0.097* -0.044 -0.187*** -0.564*** 0.098** 0.208*** 0.044
(-1.28) (0.47) (0.46) (-3.49) (-18.66) (-1.88) (-1.26) (-4.80) (-15.21) (2.02) (4.94) (1.15)
0.25 -0.061 0.011 0.004 -0.091** -0.654*** -0.081 -0.038 -0.144*** -0.594*** 0.086* 0.163*** 0.038
(-1.44) (0.26) (0.11) (-2.58) (-16.79) (-1.62) (-1.24) (-3.13) (-15.52) (1.86) (4.28) (0.98)
Median 0.025 -0.035 0.006 -0.063** -0.543*** -0.039 -0.013 -0.066** -0.712*** -0.010 0.120*** -0.007
(0.84) (-1.17) (0.27) (-2.20) (-18.76) (-1.43) (-0.52) (-2.29) (-29.71) (-0.32) (4.65) (-0.24)
0.75 0.083** -0.026 0.027 -0.028 -0.366*** 0.044 0.045 -0.029 -0.892*** -0.088*** 0.094*** 0.028
(2.47) (-0.89) (0.92) (-0.91) (-10.29) (1.37) (1.54) (-0.98) (-31.64) (-2.60) (2.61) (0.81)
0.80 0.102*** -0.049 0.034 -0.041 -0.307*** 0.041 0.056* -0.030 -0.943*** -0.112*** 0.086** 0.019
(2.81) (-1.55) (1.06) (-1.24) (-8.39) (1.49) (1.77) (-0.95) (-33.95) (-2.83) (2.19) (0.50)
0.85 0.127*** -0.070* 0.020 -0.041 -0.270*** 0.029 0.068* -0.022 -0.979*** -0.156*** 0.062 0.013
105

(3.00) (-1.80) (0.60) (-1.05) (-6.46) (0.95) (1.79) (-0.53) (-26.64) (-2.94) (1.56) (0.31)
0.90 0.154*** -0.076 -0.008 -0.016 -0.193*** 0.038 0.084* 0.039 -1.030*** -0.195** 0.050 0.022
(2.83) (-1.52) (-0.22) (-0.34) (-3.95) (0.85) (1.95) (0.59) (-25.30) (-2.57) (1.18) (0.46)
0.95 0.255*** -0.162* -0.032 0.078 -0.156* 0.074 0.075 0.110 -1.115*** -0.388** 0.028 0.069
(2.78) (-1.80) (-0.57) (1.20) (-1.69) (1.17) (0.97) (0.89) (-15.81) (-2.45) (0.49) (1.00)
OLS -0.120 -0.073 -0.005 0.008 -0.370*** -0.049 -0.029 -0.043 -0.778*** -0.196 0.190*** 0.089
(-1.46) (-0.90) (-0.11) (0.11) (-3.60) (-1.42) (-0.73) (-0.57) (-10.67) (-1.49) (4.71) (1.29)
Quantile Slope Equality Tests Results: Only significant results of asymmetry are reported with the corresponding quantiles q  1  q  .
0.2-0.4** 0.2-0.4*** 0.2-0.4* 0.2-0.4*** 0.2-0.4*** 0.2-0.4* 0.2-0.4*
0.4-0.5* 0.4-0.5**
0.5-0.6*** 0.5-0.6***
0.6-0.8*** 0.6-0.8** 0.6-0.8*** 0.6-0.8**

The table reports results from the Quantile Regression and the OLS Regression of the VDAXO changes on a set of variables; specifications 2 and 1 are estimated. T-statistics are provided in
parentheses. ***, ** and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
Table 6. Quantile Regression Results: Dependent Variable VSTOXX changes
Quantile Intercept VSTOXX t 1 VSTOXX t  2 VSTOXX t  3 STOXXR 
t STOXXR 
t 1 STOXXR 
t 2 STOXXR 
t 3 STOXXR 
t STOXXR 
t 1 STOXXR 
t 2 STOXXR 
t 3

0.05 -0.266** 0.058 -0.017 -0.077 -0.956*** -0.238** -0.412*** -0.310** -0.806*** 0.302** 0.479*** 0.308**
(-2.54) (0.86) (-0.25) (-1.20) (-12.70) (-2.28) (-3.92) (-2.46) (-9.25) (2.03) (3.77) (2.33)
0.10 -0.217*** 0.025 0.036 -0.072 -0.879*** -0.174** -0.231** -0.299*** -0.788*** 0.216*** 0.387*** 0.245***
(-2.98) (0.53) (0.69) (-1.53) (-13.09) (-2.41) (-2.55) (-4.32) (-16.33) (2.71) (4.13) (3.23)
0.15 -0.201*** 0.020 -0.013 -0.048 -0.821*** -0.120* -0.125 -0.248*** -0.793*** 0.182** 0.263*** 0.238***
(-3.16) (0.46) (-0.30) (-1.00) (-15.14) (-1.74) (-1.90) (-4.44) (-15.54) (2.51) (3.63) (3.10)
0.20 -0.153*** -0.020 0.004 -0.056 -0.788*** -0.124** -0.100** -0.205*** -0.843*** 0.109* 0.250*** 0.196***
(-2.77) (-0.47) (0.11) (-1.56) (-13.33) (-2.17) (-2.38) (-5.63) (-18.80) (1.80) (4.25) (2.98)
0.25 -0.117*** -0.016 -0.026 -0.042 -0.755*** -0.097** -0.118*** -0.175*** -0.850*** 0.092** 0.195*** 0.177***
(-2.59) (-0.42) (-0.67) (-1.40) (-13.86) (-2.00) (-3.64) (-5.90) (-22.25) (2.07) (3.40) (3.11)
Median 0.010 0.003 -0.018 -0.015 -0.607*** -0.015 -0.044 -0.101** -0.968*** 0.069 0.083** 0.091**
(0.24) (0.11) (-0.61) (-0.54) (-14.37) (-0.38) (-1.38) (-2.29) (-24.55) (1.55) (1.97) (2.18)
0.75 0.042 -0.001 -0.020 0.013 -0.448*** 0.071 0.023 -0.013 -1.235*** -0.002 -0.015 0.052
(1.05) (-0.04) (-0.62) (0.45) (-12.33) (1.64) (0.53) (-0.34) (-31.73) (-0.05) (-0.30) (1.27)
0.80 0.049 -0.007 -0.004 0.022 -0.401*** 0.078* 0.058 -0.010 -1.297*** -0.058 -0.021 0.020
(1.13) (-0.20) (-0.14) (0.71) (-10.17) (1.80) (1.25) (-0.25) (-28.06) (-1.04) (-0.44) (0.50)
0.85 0.119** 0.031 -0.013 0.027 -0.347*** 0.103** 0.079* 0.003 -1.348*** -0.005 -0.086 0.028
(2.29) (0.71) (-0.33) (0.66) (-7.65) (2.04) (1.66) (0.06) (-27.13) (-0.07) (-1.26) (0.51)
0.90 0.207*** 0.021 -0.092* 0.001 -0.354*** 0.114** 0.067 0.068 -1.387*** -0.070 -0.149* -0.016
106

(3.07) (0.43) (-1.67) (0.02) (-5.91) (2.16) (1.24) (0.88) (-16.21) (-0.84) (-1.76) (-0.20)
0.95 0.277*** 0.082 -0.063 -0.007 -0.240*** 0.235*** 0.071 0.143 -1.576*** -0.106 -0.186** -0.029
(3.16) (1.32) (-0.98) (-0.15) (-2.93) (3.07) (0.82) (1.35) (-14.23) (-1.14) (-2.06) (-0.34)
OLS -0.117* -0.011 -0.063 -0.060 -0.598*** -0.069 -0.072 -0.111** -1.199*** 0.038 0.017 0.116*
(-1.75) (-0.16) (-1.19) (-1.06) (-10.52) (-0.78) (-1.01) (-2.11) (-12.96) (0.47) (0.28) (1.80)
Quantile Slope Equality Tests Results: Only significant results of asymmetry are reported with the corresponding quantiles q  1  q  .
0.2-0.4*** 0.2-0.4** 0.2-0.4** 0.2-0.4***
0.4-0.5**
0.5-0.6* 0.5-0.6*** 0.5-0.6*
0.6-0.8*** 0.6-0.8** 0.6-0.8** 0.6-0.8* 0.6-0.8*** 0.6-0.8**

The table reports results from the Quantile Regression and the OLS Regression of the VSTOXX changes on a set of variables; specifications 2 and 1 are estimated. T-statistics are provided in
parentheses. ***, ** and * denote rejection of the null hypothesis at the 1% , 5% and 10% significance levels, respectively.
 107

The Information Content of the VDAX Volatility Index and

Backtesting Daily Value-at-Risk Models

Ihsan Ullah Badshah

Hanken School of Economics, Department of Finance and Statistics, P.O. Box 287,
FIN-65101 Vaasa, Finland. Phone: +358-6-3533 721, Fax: +358-6-3533 703,
Email: ibadshah@hanken.fi

September 20, 2010

Abstract

This essay examines the information content of the new VDAX volatility index to
forecast daily value-at-risk (VaR) estimates and compares its VaR forecasts with the
VaR forecasts of the filtered historical simulation (FHS) and RiskMetrics models. The
daily VaR models were backtested from January 1992 through May 2009 using
unconditional coverage, independence, and conditional coverage tests. A quadratic
score was also estimated for each of the models. We found that the information content
of implied volatility was superior to that of the historical volatility for the daily VaR
forecasts of a portfolio of the DAX 30 stock index: implied volatility (VDAX) and
combined (implied volatility plus GJR) VaR models outperformed the VaR models of
the FHS (GJR) and RiskMetrics. Finally, the quadratic score also supports the use of
implied volatility VaR models. Our findings have implications for traders, risk
managers, and regulators.

Keywords: backtesting, filtered historical simulation, implied volatility, value at risk,

VDAX

JEL Classifications: G13, C52, C53

 108

1 Introduction

Volatility forecasting is important to both financial practitioners and academics. This is

very much true in option pricing, option trading, hedging derivative positions, risk
management activities, stock selection and portfolio diversification, all of which require
accurate volatility modeling and forecasting. Fortunately, we have seen tremendous
development in volatility forecasting models since the introduction of the
autoregressive conditional heteroskedasticity (ARCH) model by Engle (1982). Since
then, many ARCH models have been developed that attempt to forecast volatility using
historical information. In 1986, Bollerslev (1986) extended the ARCH model proposed
by Engel (1982), to a generalized ARCH model called GARCH. Many extensions have
been made to the GARCH model (see Poon and Granger, 2003; Poon and Granger,
2005); for example, Glosten, Jagannathan and Runkle (hereafter GJR)(1993) proposed
the GJR model, an extension of the GARCH model that is used to account for
asymmetry. However, the growing literature supports using implied volatility (IV)
instead, calling IV the best forecast of future realized volatility (RV). These studies
empirically document that the information content of IV is richer and superior to that
of historical volatility (HV) when forecasting the future RV of the underlying asset (e.g.,
Day and Lewis, 1992; Christensen and Prabhala, 1998; Fleming, 1998; Dumas, Fleming
and Whaley, 1998; Blair et al., 2001; Ederington and Guan, 2002; Poon and Granger,
2003; Mayhew and Stivers, 2003; and Martens and Zein, 2004).

IV can be recovered by inverting the Black-Scholes (1973) formula. However, Britten-

Jones and Neuberger (2000) and Jiang and Tian (2005) derived model-free implied
volatility (MFIV) under the assumptions of pure diffusion and asset price processes
with jumps. They showed that the information content of MFIV is richer and superior
to that of Black-Scholes implied volatility (BSIV): the MFIV measure accounts for all
available strike prices, whereas the BSIV measure is a point-based IV, where each strike
price has a separate IV. Additionally, BSIV is subject to both model and market
efficiency, while MFIV is only subject to market efficiency (see Poon and Granger,
2003).

Deutsche Börse has launched a new VDAX volatility index for the German DAX 30
stock index that incorporates more robust information on options by using MFIV
 109

measures.83 In addition, the new VDAX measure aligns with the consensus view of
option traders (who are usually professionals) about the future direction of the
volatility of the stock market over the next 30 days. The rationale for adopting MFIV
measures was to account for both out-of-the-money (OTM) put and call options (i.e.,
volatility smirk/skew). The new VDAX measure incorporates both types of options;
therefore, it moves with changes in option prices. For example, a negative or positive
shock to the market induces adjustments in hedging and trading strategies; this
consequently triggers changes in the prices of one type of option (i.e., put or call). It
then moves in the direction of the market demand for particular types of options and
underlying stocks (see, e.g., Bollen and Whaley, 2004).84 In addition, Liu et al. (2005)
found that the rare-event premiums play an important role in generating the volatility
skew patterns observed for options across moneyness and that these rare events are
embedded in the OTM options.85 Camara and Heston (2008) derived an option pricing
model that accounted for both OTM put and call options. They derived the extreme
negative events from OTM puts and the extreme positive events from OTM calls. Thus,
The new VDAX index should be an excellent instrument for forecasting volatility (or
VaR) for the underlying DAX 30 stock index portfolio as it is implied by OTM options,
which are informed on future events and contain information on negative and positive
jumps ( see, e.g., Pan, 2002; Liu et al., 2005; Camara and Heston, 2008; Bates, 2008).

There are some famous studies related to the topic of our essay that use the information
content of IV in their volatility forecasting models and find that IV subsumes almost all
information for forecasting future RV. For example, Blair et al. (2001) studied the
information content of the VIX (now VXO) and intraday returns for the S&P 100 stock
index and then compared them for the sample period from 1987 to 1999. Their in-
sample forecasting showed that all relevant information is provided by the IV (i.e.,
implied from the VIX) and that there is not much incremental information in intraday
index returns. However, for out-of-sample forecasts, the VIX provides superior and
accurate forecasts for all forecast performance measures and horizons (i.e., from 1 to 20
days); moreover, the incremental forecasting information in the intraday returns is
insignificant. Ederington and Guan (2002) studied the importance of IV forecasts
while using S&P 500 future options, finding that IV has superior forecasting power and

83 Chicago Board Options Exchange uses a similar MFIV measure for the VIX and VXN, which are derived

from S&P 500 and NASDAQ options, respectively.

84 Whaley (2000) referred to the volatility index as an “investors’ fear gauge”.
85 Similarly, Pan (2002) showed that volatility skew is primarily due to investors’ fear of large adverse

jumps.
 110

subsumes the information in HV. Similarly, Martens and Zein (2004) confirmed that
the GARCH model extended with IV provides better forecasts than the GARCH model
extended with RV (obtained from high-frequency intraday returns). Finally, the study
by Giot (2005), which is slightly different from the above two studies, assesses the
information content of IV indices such as the VIX (now the VXO) and the VXN in a
daily VaR framework while studying time periods that include both bull and bear
markets. The performances of the VaR models were evaluated using unconditional
coverage, independence and conditional coverage tests. Backtesting results showed that
IV indices provide superior VaR forecasts and thus fewer VaR violations over time
relative to the time-series-based VaR forecasts. Moreover, the performance was stable
during market turmoil.

This essay contributes to the literature on market risk management, volatility

modeling, and forecasting. The field of volatility modeling and forecasting, especially
when predicting future RV by using IV, HV or a combination of the two, represents an
interesting and less trite area of research in finance. In this essay, we incorporated the
information content of the new VDAX volatility index into daily VaR models, and the
resulting VaR forecasts were compared with the VaR forecasts from the FHS model
(GJR) and the RiskMetrics model at confidence levels of 99%, 97.5% and 95%. The
daily VaR models were then backtested using unconditional coverage, independence,
and conditional coverage tests; furthermore, a quadratic score was estimated for each
of the VaR models for the period from January 1, 1992 through May 29, 2009.
However, this essay differs from the above-cited studies in several ways. First, they
focus on using the BSIV index to forecast future RV for U.S. stock markets, whereas we
focused on using a robust MFIV volatility index to forecast daily VaR for the German
stock market. Second, we compared the implied daily VaR forecasts with the forecasts
from the famous approach by Barone-Adesi et al. (1998), the filtered historical
simulation; to our knowledge, no other study has ever compared it with implied VaR
forecasts. Finally, we used much more detailed and robust backtesting techniques on
the very long data series to evaluate different VaR models for daily market risk
management.

Our main finding was that the MFIV (the new VDAX) index subsumed almost all
information required for actual volatility in VaR models. The backtesting results show
that the new VDAX index contained significant information regarding volatility in VaR
models and that the number of implied VaR violations was not significantly different
 111

from the set coverage rates, thereby yielding fewer VaR exceptions and clusters of
exceptions. Consequently, the null hypotheses regarding independence and conditional
coverage tests were never rejected for implied volatility, implied volatility plus GJR,
and FHS VaR models. However, the null hypotheses for the RiskMetrics model were
rejected at lower confidence levels. Finally, the quadratic scores also favored implied
VaR models.

This essay is organized as follows. In Section 2, we discuss the new VDAX index and its
construction. In Section 3, we discuss the data. In Section 4, we specify daily VaR
models. In Section 5, we discuss backtesting techniques. The empirical results are
presented in Section 6. In Section 7, we summarize and conclude.

2 The New VDAX Volatility Index

The Deutsche Börse and Goldman Sachs jointly developed the methodology for the new
VDAX index. It is based on options on the DAX 30 index; options are traded on the
derivatives exchange EUREX. Like VIX the VDAX measure is also based on MFIV
measure. Hence, it accounts for IVs across all options of a given time to expiry (it
accounts for volatility smirk/skew). In fact, the main VDAX index is further based on
eight constituents volatility indices that expire in 1, 2, 3, 6, 9, 12, 18, and 24 months,
respectively. The main VDAX is designed as a rolling index at a fixed 30 days to expiry
via a linear interpolation of the two nearest of the eight available sub-indices. The price
history for the VDAX is calculated back to the years 1992. The VDAX and its eight sub-
indices are updated every minute. The sub-indices are calculated according to the
formula below:

VDAX i 100 ˜  i2 , (1)

2
§ Fi ·
˜ Ri ˜ M K i,j  
2 K i,j 1
 i2
Ti
¦ K i,j2 Tj
¨
¨K  1 ¸ , i
¸ 1,2,...8. (2)
j © i,0 ¹
Where Ti is time to expiry of the ith DAX 30 option (best bid and best ask of all DAX

30 options) , Fi is forward price derived from the prices of the ith DAX 30 options, for
which the absolute difference between call and put prices is the smallest as
Fi K min C  P  Ri ˜ C  P  , K i,j the exercise price of the OTM option of the ith DAX
 112

K i,j 1  K i,j 1
30 option expiry month , K i,j is the interval between the strike
2
prices, K i,0 the highest exercise price below forward price, Ri is the refining factor

equal to Ri e ri .Ti , r is the risk free rate, M K i,j  is the price of the option K i,j ,

whereby K i,j z K i,o , and M K i,0  is the average of the put and call prices at exercise

price K i,0 .

Finally, the main VDAX is designed as a rolling index at a fixed 30 days to expiry via a
linear interpolation of the two nearest of the eight available sub-indices as follows.

§ § N  NT · § N  N Ti · · N 365
VDAX ¨ T ˜ VDAX 2 ˜ ¨ Ti 1 ¸  Ti 1 ˜ VDAX i21 ˜ ¨ T ¸¸ ˜ , (3)
¨ i i
¨ NT  NT ¸ ¨ NT  NT ¸ ¸ NT
© © i 1 i ¹ © i 1 i ¹¹

where N Ti time to expiry of the ith DAX 30 index option, N Ti  1 time to expiry of the i+

1th option, N T time for next days, and N 365 time for standard year.86

3 Data

This essay employs data from two sources. We obtained the daily time-series price data
for the DAX 30 stock index from Financial DataStream. The data on the new VDAX
were obtained from the Deutsche-Börse. The daily data for both the DAX 30 and the
VDAX indices cover a period of 17 years and 5 months, from January 1, 1992, to May
29, 2009, for a total of 4,543 trading days.

Figure 1 shows the daily continuously compounded returns (%) of the DAX 30 stock
index and the daily closing level (%) of the VDAX volatility index from January 1992
through May 2009. As can be seen, from January 1992 through May 1997, the VDAX
showed the lowest volatility levels and an upward moving stock index (a period in a
low-volatility bull market). From June 1997 through April 2000, there were high VDAX
levels (a bubble period where we find a high-volatility bull market). A spike in the

86 See, for further detail on the VDAX construction, the Deutsche Börse and Goldman Sachs methodology

for VSTOXX on the STOXX website: www.stoxx.com, the same methodology is also used for the
construction of the VDAX volatility index.
 113

VDAX can be seen in 1998. This was followed by a stock market crash due to the Long-
Term Capital Management (LTCM) crisis; during this crash, the VDAX level reached as
high as 57%. A crash can also be observed from April 2000 through January 2003
when the bubble burst (a period in a high-volatility bear market).

DAX 30 stock index returns

15

10

-5

-10
92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09

VDAX volatility index level

100

80

60

40

20

0
92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09

Figure 1. DAX 30 stock index returns (%) and the VDAX volatility index level (%) from
January 1992 through May 2009.

However, from 2004 to August 2007 (a period in a low-volatility bull market), we

found positive returns, and the level of the VDAX decreased, showing low volatility
 114

levels. From September 2007 through May 2009 (a period of extremely high volatility
and an extreme bear market), we found the highest VDAX levels, decreasing stock
index returns and some jumps due to the credit crunch and liquidity crises. In this
particular period, a spike in the VDAX level was observed in October 2008 when the
market crashed and the VDAX level reached a historical peak of 74%.

Table 1 provides the summary statistics for the daily continuously compounded returns
(in %) of the DAX 30 index as well as tests for normality and unit roots.

Table 1. Descriptive Statistics of the DAX 30 returns series.

DAX 30
Mean 0.000251
Median 0.000501
Maximum 0.107975
Minimum -0.088747
Std.Dev. 0.014682
Skewness -0.111999
Kurtosis 8.228179
Jarque-Bera 5183.57**
ADF -68.75**
No. Obs 4543
Table1 report the descriptive statistics of the returns series of DAX30 index. The Jarque-Bera and the Augmented Dicky Fuller
(ADF) (an intercept is included in the test equation) test values are reported.
** and * denote rejection of the null hypothesis at the 1% and 5% significance levels respectively.

The mean return on the DAX 30 index is not statistically different from zero. The tests
for skewness and kurtosis confirm that returns on the DAX 30 index are negatively
skewed and are highly leptokurtic with respect to a normal distribution. Likewise, the
Jarque-Bera statistics reject normality for the return series. The stationarity in the
return series was investigated by applying the augmented Dickey-Fuller (ADF) test for
a unit root. The ADF results reject the hypothesis of a unit root in the return series at
the 1% significance level.

Figure 2 shows a Q-Q plot of the theoretical quantiles of the normal distribution
(vertical axis) against empirical quantiles of the returns on the DAX 30 (horizontal
axis). As can be seen from this plot, the data are not normally distributed, indicated by
the fact that the empirical quantiles do not lie on a straight line
 115

.06

.04

Quantiles of Normal .02

.00

-.02

-.04

-.06
-.10 -.05 .00 .05 .10 .15

Quantiles of returns on DAX 30

Figure 2. Normal Q-Q plots returns on DAX 30.

As can be seen from the plot, there is a significant deviation from the straight line in the
tails, particularly in the lower tail, indicating that the distribution of returns on the
DAX 30 index is more heavily tailed than the normal distribution. The Q-Q plot thus
reinforces the findings of our earlier statistical tests for normality, leading to the use of
a heavy-tail distribution rather than a normal distribution in the rest of our analysis.
 116

4 Daily Value-at-Risk Models

4.1 Value-at-Risk

Value at Risk (VaR) is defined as the maximum expected loss in the value of a portfolio.
It has a certain probability over a certain holding period (for details on VaR, see, e.g.,
Duffie and Pan, 1997; Jorion, 2000; Dowd, 2005; Christoffersen, 2009). VaR forecasts
are fundamental to financial risk management and risk regulation. However, the
importance and recognition of VaR as a risk management tool spring from the Market
Risk Amendment (1996) to the Basel Capital Accord of 1998 and also because of the
popularity of the RiskMetrics introduced by J.P. Morgan (see Jorion, 2000 and Dowd,
2005). After these events, VaR became widely accepted by banks and was also imposed
by regulators. The aim of these two groups was to supervise and manage market risk—
the market-exposure risk arising due to unfavorable movements in equity prices,
interest rates, exchange rates, commodity prices etc.87 Because VaR has become a
standard measure for market risk due to the trading activities and market positions
taken by large banks, most financial institutions and trading houses currently use VaR
models to assess their daily portfolio losses from significant trading activities. They also
backtest VaR models by observing when the portfolio returns exceed the VaR forecasts;
that is, the VaR forecasts match their expectations. As a result, accurate VaR forecasts
are crucial for market risk management. Given that accurate VaR forecasts rely heavily
on the accurate forecasting of the volatility of a portfolio, this is an important
parameter for any VaR model. For instance, the level of the VaR over a one-day holding
period is defined as the solution

P(rt P  VaR tP  t 1 ) , (4)

where  is 1 minus the VaR confidence level (e.g. 99%), and rt P is the return of a

portfolio over a one-day holding period. Having conditional volatility specification ht

such that rt P ht  t , where the residuals are distributed as  t ~N 0,ht  , then a one-

period VaR at time t is

VaR t  1   ht , (5)

87 The Basel committee for banking supervision allows banks to use VaR as a benchmark in order to
determine how much additional capital is needed to cover daily market risk beyond that required for credit
risk.
 117

where  denotes the standardized normal cumulative distribution function. In fact, we

need to accommodate heavy tails in VaR estimation; therefore, VaR needs to be
estimated using a Student’s t density function.88 We assume that  t are distributed as

follows:
 t ht    2  ~ ,
12
(6)

where  is the standardized Student’s t-distribution with degrees of freedom.

To estimate the %daily VaR using Student’s t density function, we followed the
method described by Dowd (2005):89

 1  §¨   2  ·¸
1
2
VaRt, ht , (7)
© ¹
where the shape is to be estimated. Here, we used four different types of volatility
forecasts (implied volatility (VDAX), implied volatility (VDAX) with GJR, FHS (GJR),
and RiskMetrics) as parameters for the daily VaR models; these are specified and
discussed in the following subsections.

4.2 Implied Volatility

We considered the new VDAX index (MFIV index), which is a robust and more
informed measure. The rationale for considering the new VDAX is that we believe VaR
modeling is about extreme events; that is, implying volatility from the extreme
outcomes in the options market is of paramount importance to VaR forecasting. These
extreme events are embedded in the IV derived from OTM options (see, for instance,
Liu et al., 2005; Camara and Heston, 2008; Bates, 2008). Likewise, the importance of
the new VDAX measure increases because it accounts for volatility smirk/skew, which
may be induced by the net buying pressure of the OTM put options (see Bollen and
Whaley, 2004). Volatility skew is an obvious phenomenon previously documented by
many other researchers and is important to capture in any volatility measure (e.g.,
Alexander, 2001; Low, 2004; Goncalves and Guidolin, 2006; Badshah, 2008). In
addition, information from trading strategies and other shocks is well absorbed in the

88 Most previous studies have concluded that distribution functions accounting for fat tails are
fundamental for VaR modeling. See, for instance, Huisman et al. (1998), and Alexander and Sheedy (2008)
for VaR forecasts obtained through different density functions.
89 Because the returns on the DAX 30 are not normally distributed, the VaR forecasts are estimated using

Student’s t density function.

 118

new VDAX index because it accounts for both OTM put and call types of options.
Nevertheless, the new VDAX is implied from both the fear and exuberance embedded
in option prices, even though the majority of option traders are very informed and
possess strong skills (see, e.g., Low, 2004 and Chakravarty et al., 2004). As a result, the
new VDAX is a perfect choice to be used as a volatility parameter in the daily VaR
model to quantify a daily VaR forecast for the DAX 30 stock index portfolio. However,
for the daily VaR model, a daily-variance parameter is needed in place of the standard
deviation because VaR uses variance as an input, as the VDAX is expressed in
annualized standard deviation units. As a result, transformation is essential; for
instance, at time t, we insert VDAX t 1 into the daily VaR model as 90

2
ht  imp,1,t 1 , (8)

while
2
 imp,1,t 1  252
.
VDAX t  1 2
(9)

However, Granger and Poon (2003, 2005) point out that BSIV is biased and is always
higher than the actual volatility. They suggest using HV for calibration as

ht    imp,1,t
2
1 , (10)

where and  need to be estimated. However, we assert that VDAX is calculated

using the MFIV measure, whose IV value should not be subject to model risk, and that
this bias should thus not be of great concern; therefore, any of the above two variance
measures can be used equally. Furthermore, we have a combined specification for
variance using GJR-GARCH (1,1) extended with the lagged IV as

ht  0   1  t2 1   2  t2 1 d t  1  h t  1   imp,1,t

2
1 . (11)

This equation has a dummy variable to capture asymmetry. For instance, the dummy
variable d t 1 is equal to 1 when  t 1  0 and is equal to 0 otherwise. Therefore,

estimating an %daily implied (VDAX) VaR forecast or the combined VaR forecast
using implied volatility (VDAX) plus GJR can be done with the following specification:

 1  §¨   2  ·¸
1
2
VaRt, ht , (12)
© ¹

90 A similar transformation scale is used by Blair, Poon, and Taylor (2001) for the VIX index.
 119

where  is the standardized Student’s t-distribution function, the shape parameter

needs to be estimated and  is the quantile (which in our case is 99%, 97.5% or 95%).

4.3 Filtered Historical Simulation

As we know, historical simulation (bootstrapping) is a model-free approach that uses

the past historical returns of each asset in a portfolio in order to generate future
scenarios; therefore, a historical simulation (HS) VaR forecast is based on the empirical
distribution of asset returns, which is more realistic and economically appealing for
measuring a portfolio’s risk. Pritsker (2006) and Boudoukh et al. (1998) thoroughly
discussed the assumptions and limitations of HS, pointing out that HS has many
disadvantages. HS-VaR forecasting may be misleading; for instance, the asset returns
are not independently, identically distributed (iid), leading to the present fat-tailed
distribution with time-varying conditional moments and volatility clustering. Barone-
Adesi et al. (1998) extend the HS and propose Filtered Historical Simulation (FHS),
which overcomes limitations of the HS and is consistent with the empirically observed
features in the financial market data.91 FHS has a great advantage in that we do not
need to make any assumptions about the distribution in advance; instead, the past data
directly tell us about the distribution. Here, we fit a GARCH-type model to the data. We
based the conditional volatility model on the assumption that the model must produce
iid-standardized returns. Finally, we came up with the following GJR-GARCH (1,1)
model specification:

rt t 1   t ,  t ~N 0,ht , (13)

ht  0   1  t21   2  t21 d t 1  ht 1 . (14)

This model has a dummy variable to capture asymmetry. For instance, the dummy
variable d t 1 is equal to 1 when  t 1  0 and equals 0 otherwise.92 However, for

forecasting using FHS-VaR, the i.i.d.-standardized residual returns are essential. The
residuals are divided by the corresponding daily conditional volatility obtained from
the GJR model:

91 FHS combines the characteristics of parametric and non-parametric methods, which account for

leptokurtosis, time-varying volatility, asymmetry, and serial correlations.

92 The MA term is inserted to remove serial correlation in the returns series.
 120

t
zt ht
. (15)

The daily FHS-VaR forecast is estimated by using the VaR specification:

VaR t >
 ht *Perc  z t  i i 1 , ,
n
@ (16)

where n represents the standardized residuals and  is the quantile (which in our case
is 99%, 97.5% or 95%).

4.4 RiskMetrics

RiskMetricsTM is a set of methodologies for measuring market risk (e.g., VaR)

developed by J.P. Morgan. The RiskMetrics model assumes that the daily return of a
portfolio follows a conditional normal distribution. It assigns greater weight to recent
observations and less weight to more distant observations. This declining weighting
scheme, through the assertion that volatility tends to change over time in a stable way,
might be more reasonable than assuming that it is constant (see e.g., Christoffersen,
2003). The specification of the RiskMetrics model is

ht rt 21  (1  )h t 1 , (17)

where  is weight, which is usually fixed at 0.94 by the RiskMetrics group. The
RiskMetrics model’s forecast for one-day-ahead volatility at time t is therefore a
weighted average of volatility at time t  1 times the squared return at t  1 . Moreover,
this model has several advantages, some of which are worth mentioning here. First, it
accounts for time variations in the variance in a manner consistent with observed
returns. Second, recent observations are given more importance than the older
observations. Third, fewer observations are needed in order to forecast a one-day-
ahead variance. Finally, there is an agreement on the parameter value O 0.94 across
assets for one-day-ahead variance forecasting (for more details, see, e.g.,
Christoffersen, 2003). Nevertheless, we improve the RiskMetrics VaR forecast by
accommodating a heavy-tailed Student’s t-distribution function instead of a normal
distribution function. Hence, the daily D % RiskMetrics VaR forecast is estimated as

 1  §¨   2  ·¸
1
2
VaRt, ht , (18)
© ¹
 121

where  is the standardized Student’s t-distribution function, the shape parameter

needs to be estimated, and  is the quantile (which in our case is 99%, 97.5% or 95%).

5 Backtesting Daily VaR Models

The current regulatory framework requires financial institutions with massive trading
activities to have enough capital, called a market capital requirement (MRC), in order
to cover excessive portfolio losses. This MRC requirement is determined in terms of the
portfolio VaR measure. The regulatory frameworks also necessitate that financial
institutions should use their own internal VaR models to compute and provide their
99% VaR. As a result, the MRC requirements are directly linked to both the portfolio
risk level and the internal VaR model’s performance on backtests. However, based on
VaR model forecasts, the regulatory body has increased (decreased) MRC requirements
(for more details on MRC, see Campbell, 2006). Consequently, the accuracy of internal
VaR models is of paramount importance to both regulators and financial institutions.93
When examining the accuracy of a VaR model, the significance of the backtesting
technique used increases. In this respect, Kupiec (1995) was the first to propose a test
of unconditional coverage; later, Christoffersen (1998) extended the Kupiec test to the
test of conditional coverage.

Let us assume we want to backtest a VaR model; to do that, a hit sequence needs to be
defined. For instance, if ex-ante VaR forecasts and ex-post returns are observed in a
time-series manner, then a hit sequence of VaR violations (or exceptions) are
represented with the following indicative function:

­°1, if rt P  VaRtP  
I t   ® ˜ (19)
°̄0, if rt P ! VaRtP  
If a portfolio’s loss on day t is greater than the forecasted VaR for day t, then the hit
sequence reports 1, implying a VaR violation; otherwise, the hit sequence reports 0,

implying no VaR violation. In this way, the hit sequence ^I t `Tt 1 across T days is

constructed, indicating the violation rate of the VaR model. In fact, Christoffersen

93 For instance, if a VaR model indicates less risk, then the regulatory body could reduce MRC

requirements, and the financial institution could then use that capital for other financial activities. This is
why accurate VaR forecasting is very important for a financial institution.
 122

(1998) suggests that the VaR model is adequate when its ‘hit sequence’ satisfies both
unconditional coverage and independence properties.94

The unconditional coverage property: The probability that an ex-post loss exceeds VaR
forecasts must be exactly equal to the coverage rate.

The independence coverage property: The VaR violations observed at two different
periods must be independently distributed over time. The observed VaR violations do
not carry information to forecast current and future VaR violations, and this property
also holds for any other variable in the information set, for instance, past returns, past
VaR levels, etc. Therefore, the null hypothesis of the unconditional coverage property
that E >I t @ P can formally be tested using the log-likelihood ratio test is

§§ X· §X· ·
T X

2ln¨ ¨ 1  ¸ ¨ ¸ ¸  2ln 1  P  P X ~ 12 ,
T X
LRUC (20)
¨© T ¹ © T ¹ ¸
© ¹
which is asymptotically  2 -distributed with one degree of freedom. P is the target
coverage rate (i.e., 1% for a 99% VaR), T is the total number of observations, and X is
the number of violations. Unfortunately, the unconditional coverage test does not
account for the clustering of violations.95 To overcome a clustering problem, the
independence coverage property must be satisfied, thereby leading to the correct
conditional coverage test proposed by Christoffersen (1998): joint tests for both
independence and unconditional coverage tests.96 The likelihood ratio tests for these
tests are

§§ X · 00 10 § X · 01 11 ·¸ 2
T T T T

LR IND  T T

2ln 1   0  00  0T01 1   1  01  1T11  2ln¨ ¨ 1  ¸
¨© T ¹
¨ ¸ ~ , (21)
¸ 1
© ©T ¹ ¹
LR CC LR UC  LR IND , (22)

94 Christoffersen (2009) thoroughly reviews backtesting techniques.

95 This is because the unconditional coverage test has a major problem in that it does not account for the
clustering of ones or zeros in a time-dependent fashion. For instance, if a value of 1% gave exactly 1%
violations, but all of these violations occurred during one-month period, then the probability of bankruptcy
of a financial institution would be much higher than if the VaR violations are scattered over a longer
backtesting period, such as a one or two-year period.
96 For some technical details, see, for instance, Christoffersen (1998) and Campbell (2005).
 123

where Tij is the number of observations, with value i followed by j , for i,j 0,1 ;

 ij Tij ¦ j Tij are the corresponding probabilities; ij 1 indicates that a violation

has occurred; and ij 0 indicates the opposite. The LR CC statistics are asymptotically

 2 distributed with two degrees of freedom. The null hypothesis in the independence
test LR IND states that the probability of a violation on a given day does not depend on

the previous day’s violation.

An alternative method for evaluating VaR forecasts could be based on the use of loss
functions that are more aligned to the regulatory requirements. Lopez (1999) was the
first to propose a loss function for evaluating VaR forecasts, here called Lopez-I.
However, this loss function has a major drawback in that it ignores the magnitude of
tail losses. To remedy this drawback, he himself proposed a second loss function called
Lopez-II. In practice, Lopez-II has a size-adjusted loss function that makes it hard to
interpret in monetary terms because it employs squared monetary returns.97 Later,
Dowd (2005) proposed a size-based loss function that avoids terms in the denominator
and the squared term in Lopez-II; therefore, we prefer Dowd’s size-based function over
the others. It takes the following form:

­ Lt if Lt ! VaRt
Ct ® (23)
¯0 if Lt d VaRt ,

where Lt is the loss incurred over period t and VaRt is the forecasted VaR estimate for

that period. We then calculated the expected tail loss ET by using C t , and ET was

used as a benchmark. Finally, the quadratic score function takes the following form:

2
2 n
QS ¦ Ct  ETt  .
nt1
(24)

This function has advantages over others because it punishes deviations of the tail
losses from the expected value and because of its quadratic form; it gives very high tail
losses much greater weight than normal tail losses.

97 See Dowd (2005) for detailed discussions on different loss functions.

 124

6 Results
6.1 Conditional Variances

Figure 3 provides a comparison of the conditional variances for implied volatility

(VDAX), GJR plus implied volatility (VDAX), GJR-GARCH (1, 1), and RiskMetrics
specifications and the squared returns that they forecast. The comparison period for
the five time series is from January 1, 1992 through May 29, 2009.

Implied(VDAX) GJR-Implied(VDAX)

.0028 .0028

.0024 .0024

.0020 .0020

.0016 .0016

.0012 .0012

.0008 .0008

.0004 .0004

.0000 .0000
92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

GJR RiskMetrics

.0028 .0028

.0024 .0024

.0020 .0020

.0016 .0016

.0012 .0012

.0008 .0008

.0004 .0004

.0000 .0000
92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

Squared Returns

.012

.010

.008

.006

.004

.002

.000
92 94 96 98 00 02 04 06 08

Figure 3. Implied, conditional variances and squared returns from January 01, 1992
to May 29, 2009.
 125

It can be observed that the squared returns are very noisy and that the four variances
move closely together.

6.2 Backtesting Results

Table 2 provides backtesting results of VaR models for the DAX 30 stock index
portfolio from January 1, 1992 to May 29, 2009.98 Implied volatility (VDAX), implied
volatility–GJR, FHS (GJR) and RiskMetrics VaR models were statistically tested using
tests of unconditional coverage, independence, and conditional coverage and the
quadratic score test (using the expected tail loss) at different confidence levels, for
instance, 99%, 97.5% and 95%.99 There are three panels in Table 2. In rows 1, 2, and 3
of each panel, the results of the unconditional coverage, independence and conditional
coverage tests are provided, respectively. In rows 4, 5, 6, 7, and 8 of each panel, the
number of exceptions, the %VaR exceptions, the number of successive exceptions, the
% expected tail loss, and the quadratic score are presented, respectively.

First, we will discuss the backtesting results for our four specified VaR models at the
99% confidence level, which are provided in panel A of Table 2.100 Because the null
hypothesis of the unconditional coverage test states that the observed frequency of
exceptions should equal the frequency of expected exceptions, testing this is the first
step to take when considering any VaR model. As can be seen, the chi-square values for
the null hypothesis of the unconditional coverage tests at 10% significance cannot be
rejected for any of the four VaR models.101 Unfortunately, the unconditional test only
considers the frequency of exceptions, and it ignores the time-series independence of
those exceptions, i.e., the clustering of exceptions. As a result, the independence test is
of paramount importance for any VaR model. Conducting the independence test is the
second step to take when considering any VaR model.

98 All VaR models except FHS are estimated in conjunction with the Student’s t density function.
99 We do recognize that the daily VaR forecasts of the FHS model were based on using in-sample GJR-
GARCH forecasts. However, a more appropriate way would be to estimate the daily FHS-VaR forecasts
using out-of-sample GJR-GARCH model forecasts. However, this issue should not be of concent in the rest
of the daily VaR models’ forecasts such as using Riskmetrics or Implied Voalatility forecasts.
100 The daily 99% VaR forecasts for each of the four VaR models are compared with the corresponding daily

returns in a time-series fashion in Figure 4.

101 A chi-square critical value of 2.70 (for one degree of freedom) for a 10% significance level is used to test

the null hypothesis that the frequency of observed exceptions is consistent with the frequency of expected
exceptions.
 126

Table 2. Backtesting results of VaR models for the DAX 30 from January 1, 1992 to May
29, 2009.
Panel A: 99% VaR-t forecasts
Model Implied (VDAX) Implied (VDAX)- FHS (GJR) RiskMetrics

LRU 1.299692 0.663502 0.919916 2.317618

LRI - -6.763572 2.115715 1.698050
LRCC - -6.100069 3.035631 4.015668
No. of Exceptions 38 51 52 56
%VaR Exceptions 0.843507 1.122606 1.144618 1.232666
No. Succ. Exceptions 0 2 2 2
Expected Tail Loss (%) -4.323369 -3.5817592 -3.71932664 -3.48212508
Quadratic Score 0.000371 0.000594 0.000635 0.000740

Panel B: 97.5% VaR-t forecasts

Model Implied(VDAX) Implied(VDAX)-GJR FHS(GJR) RiskMetrics

LRU 0.285262 0.018263 0.781421 3.912869

LRI -6.094480 -6.926530 -7.03222 -6.738680
LRCC -5.809210 -6.908270 -6.25080 -2.825810
No. of Exceptions 108 115 123 135
%VaR Exceptions 2.435175 2.531367 2.707462 2.971605
No. Succ. Exceptions 1 4 4 5
Expected Tail Loss (%) -3.728668 -3.221964 -3.099948 -3.176138
Quadratic Score 0.000897 0.001082 0.001288 0.001403

Panel C: 95% VaR-t forecasts

Model Implied(VDAX) Implied(VDAX)-GJR FHS(GJR) RiskMetrics

LRU 0.123797 1.605154 1.786875 8.434941

LRI -4.187958 -1.675680 1.828643 0.047557
LRCC -4.064160 -0.070530 3.615519 8.482498
No. of Exceptions 222 246 247 271
%VaR Exceptions 5.137699 5.414924 5.436936 5.965221
No. Succ. Exceptions 7 7 9 17
Expected Tail Loss (%) -3.295459 -2.841407 -2.783163 -2.751958
Quadratic Score 0.001844 0.002288 0.002334 0.002613
Boldface indicates that we cannot reject the null hypothesis at 10% significance level.
 127

The chi-square values at the 10% significance level for the independence tests confirm
that the null hypothesis of the first-order independence of exceptions is not rejected for
the three specified VaR models, i.e., implied volatility (VDAX)-GJR, FHS (GJR), and
RiskMetrics; however, this test could not be conducted for the implied volatility
(VDAX) model because there are no successive exceptions implying that the model is
correct. Finally, the conditional coverage test that needs to be tested in order to endorse
any VaR model is the joint null hypothesis that the VaR model has an adequate
frequency of independence exceptions. This test also provides important information
regarding the likelihood of the successive exceptions and the average number of days
between exceptions; this is the third and final step for a VaR model to be considered for
use. As can be observed, the chi-square values at the 10% significance level for the joint
null hypothesis tests verify that the three VaR models have an adequate frequency of
independence exceptions; as a result, the joint null hypothesis could not be rejected.
Similarly, for the implied volatility (VDAX) model, the conditional coverage test could
not be conducted because there are no successive exceptions implying that this model is
correct under the conditional coverage test.102

For comparison, the backtesting results for the VaR models at 97.5% and 95%
confidence levels are provided in panels B and C of Table 2.103 As can be seen, the chi-
square values for the null hypothesis of the unconditional coverage tests at the 10%
significance level also could not be rejected for three of the VaR models. However, in
the case of the RiskMetrics model, we found a higher unconditional coverage than
expected, leading to a rejection of the null hypothesis of correct unconditional coverage
at the 10% significance level. In comparison, the chi-square values at 10% significance
for the independence tests show that the null hypothesis of the independence of the
exceptions was not rejected for our VaR models. As can be observed, for the conditional
coverage test, the chi-square values at the 10% significance level for the joint null
hypothesis tests verify that the four VaR models have an adequate frequency of
independence exceptions; therefore, the joint null hypothesis could not be rejected.
However, the null hypothesis for the conditional coverage test was rejected at the 10%
significance level for the RiskMetrics model (at the 95% confidence level).

102 Chi-square critical value of 4.60 (for two degree of freedom) for 10% significance level is used to test the

joint null hypothesis of the conditional coverage test.

103 The daily 97.5% and 95% VaR forecasts for each of the four specified VaR models are compared with

corresponding daily returns in a time-series fashion in Figure 5, and 6, respectively.

 128

In order to select a model, we can also use estimated quadratic scores for each VaR
model at different confidence levels such as 99%, 97.5% and 95%. However, the
quadratic scores are estimated using a loss function proposed by Dowd (2005), which
states that a lower minimum score level indicates a better VaR model. The quadratic
scores are presented in the last rows of each panel of Table 2. If we compare the
quadratic scores for each of the VaR models at the 99% confidence level, the rank of the
VaR models would be, from better to worse. implied volatility (VDAX), implied
volatility (VDAX)-GJR, FHS (GJR), and RiskMetrics. Alternatively, in a more
traditional way, one could also select a VaR model by evaluating it based on the number
of exceptions and the number of successive exceptions it has previously made. By these
criteria, the VaR models’ rank is similar to the rank we determined from the quadratic
score estimates. In addition, ranks for the VaR models at lower confidence levels
(97.5% and 95%) support the quadratic score estimates of the VaR models at the 99%
confidence level.

We conclude from the results presented in Table 2 that the information content of IV is
superior to that of HV, which is consistent with the earlier research on the information
content (e.g., Day and Lewis, 1992; Christensen and Prabhala, 1998; Fleming, 1998;
Dumas, Fleming and Whaley,1998; Blair et al., 2001; Ederington and Guan, 2002;
Poon and Granger, 2003; Mayhew and Stivers, 2003; Martens and Zein; 2004; and
Giot, 2005). The unconditional coverage, independence, and conditional coverage tests
and the quadratic scores suggest that the VDAX and a combination of the VDAX with
GJR-GARCH (1,1) VaR models always outperform filtered historical simulation (FHS)
and RiskMetrics models during our sample period from January 1, 1992 through May
29, 2009.
 129

7 Conclusion

This essay examined the information content of the newly adopted VDAX (MFIV) index
for the daily VaR forecasts. The information content of the new VDAX was incorporated
into daily VaR forecasts and compared with the VaR forecasts from the FHS (GJR) and
the RiskMetrics models at various confidence levels (i.e., 99%, 97.5% and 95%), using
unconditional coverage, independence, and conditional coverage tests for backtesting
each of the VaR models. Furthermore, a quadratic score was estimated for each VaR
model for the period from January 1, 1992 through May 29, 2009. The backtesting
results showed that the new VDAX index contains significant information about actual
volatility in VaR models. The null hypotheses of independence and conditional
coverage backtests were never rejected for implied volatility, implied volatility plus
GJR, and FHS (GJR) VaR models. The number of VaR exceptions was not significantly
different from the set coverage rates. However, the null hypotheses of the RiskMetrics
model were rejected for lower confidence levels. It was also found that implied volatility
and implied volatility-GJR VaR models presented the fewest VaR exceptions and
clusters of exceptions, in contrast to the FHS (GJR) and RiskMetrics models. On the
other hand, the quadratic score for each model suggests the following ranking of VaR
models: implied volatility (VDAX), combined (implied volatility plus GJR), FHS, and
RiskMetrics. Our findings have implications for traders who hold long positions, risk
managers (internal), and regulators (external).
 130

References

Alexander, C., 2001. Principal component analysis of implied volatility smiles and
skews. ISMA Discussion Paper, University of Reading, UK.
Alexander, C., Sheedy, E.,2008. Developing a stress testing framework based on market
risk Models. Journal of Banking and Finance 32, 2220-2236.
Badshah, I., 2008, Modeling the dynamics of implied volatility surfaces. European
Financial Management Association (EFMA)-2008, Athens, Greece, Conference
Proceedings.
Barone-Adesi, G., Bourgoin, F., Giannopoulos, K., 1998. Don’t look back. Risk 11, 100-
103.
Bates, D., 2008. The market for crash risk. Journal of Economic Dynamics & Control
32, 2291-2321.
Black, F., and Scholes, M., 1973. The pricing of options and corporate liability. Journal
of Political Economy 81, 636-654.
Black, F., 1976. Studies of stock market volatility changes. proceedings of the American
statistical association, Business and Economic Statistics Section, 177-181.
Blair, B., Poon, S.H., and Taylor, S., 2001. Forecasting S&P 100 volatility: The
incremental information content of implied volatilities and high frequency
index returns. Journal of Econometrics 105, 5-27.
Bollen, N., Whaley, R., 2004. Does net buying pressure affect the shape of the implied
volatility functions? Journal of Finance 59, 711-753.
Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. Journal
of Econometrics 31, 307-327.
Boudoukh, J., Richardson, M.,Whitelaw, R.,1998. The best of both worlds. Risk 11, 64-
67.
Britten-Jones, M., and Neuberger, A., 2000. Option prices, implied processes, and
stochastic Volatility. Journal of Finance 55, 839-866.
Camara, A., Heston, S., 2008. Closed form option pricing formulas with extreme
events. Journal of Futures Markets 28, 213-230.
Campbell, S., 2006. A review of backtesting and backtesting procedures. Journal of
Risk 9, 1-17.
Carr, P., Wu, L., 2006. A Tale of Two indices. Journal of Derivatives 13, 13-29.
Chakravarty, S., Gulen, H., and Mayhew, S., 2004. Informed trading in stock and
option markets. Journal of Finance 59, 1235-1258.
Christensen, B., and Prabhala, N., 1998. The relation between implied and realized
volatility. Journal of Financial Economics 50, 125-150.
Christoffersen, P., 1998. Evaluating interval forecasts. International Economic Review
39, 841-862.
Christoffersen, P., 2003. Elements of financial risk management. Elsevier Science,
Academic
Press, San Diego, United States.
Christoffersen, P., 2009. Value-at-Risk models. Handbook of Financial Time Series,
First edition, Springer Verlag, Berlin, Germany.
Day, T., and Lewis, C., 1992. Stock market volatility and the information content of
stock index options. Journal of Econometrics 52, 267-287.
Dennis, P., Mayhew, S., Stivers, C., 2006. Stock returns, implied volatility innovations
and the asymmetric volatility phenomenon. Journal of Financial and
Quantitative Analysis 41, 381-406.
Dowd, K., 2005. Measuring market risk. Second edition, Wiley Finance, Sussex,
England.
Duffie, D., Pan, J., 1997. An overview of Value at Risk. Journal of Derivatives 4, 7-49.
 131

Dumas, B., Fleming, J., and Whaley, R.E., 1998. Implied volatility functions: Empirical
tests. Journal of Finance 53, 2059-2106.
Ederington, L., and Guan, W., 2002. Is implied volatility an informationally efficient
and effective predictor of future volatility? Journal of Risk 4, 29-46.
Engle, R., 1982. Autoregressive conditional heteroskedasticity with estimates of the
variance of UK inflation. Econometrica 50, 987-1008.
Fleming, J., Ostdiek, B., and Whaley, R., 1995. Predicting stock market volatility: A new
Measure. Journal of Futures Markets 15, 265-302.
Fleming, J., 1998. The quality of market volatility forecasts implied by S&P100 index
option prices. Journal of Empirical Finance 5, 317-345.
Giot, P., 2005. Implied volatility indices and daily value at risk models. Journal of
Derivatives 12, 54-64.
Glosten, L., Jagannathan, R., and Runkle, D., 1993. On the relation between expected
excess return on stocks. Journal of Finance 48, 1779-1801.
Goncalves, S., and Guidolin, M., 2006. Predictable dynamics in the S&P 500 index
options implied volatility surfaces. Journal of Business 79, 1591-1635.
Huisman, R., Koedijk, K., Pownall, R., 1998. VaR-x: Fat tails in financial risk
management. Journal of Risk 1, 47-61.
Jiang, G., and Tian, Y., 2005. The model-free implied volatility and its information
contents. Review of Financial Studies 18, 1305-1342.
Jorion, P., 2000. Value at Risk: the benchmark for controlling market risk. Second
edition, McGraw-Hill, New York, United States.
Kupiec, P., 1995. Techniques for verifying the accuracy of risk measurement models.
Journal of Derivatives 2, 173-184.
Liu, J., Pan, J., and Wang, T., 2005. An equilibrium model of rare-event premia and its
implication for option smirks. Review of Financial Studies 18, 131-164.
Lopez, A., 1999. Regulatory evaluation of value-at-risk models. Journal of Risk 1, 37-64.
Low, C., 2004. The Fear and exuberance from implied volatility of S&P 100 index
options. Journal of Business 77, 527-546.
Martens, M., and Zein, J., 2004. Predicting financial volatility: High-frequency time-
series forecasts vis-à-vis implied volatility. Journal of Futures Markets 24,
1005-1028.
Mayhew, S., and Stivers, C., 2003. Stock return dynamics, option volume, and the
information content of implied volatility. Journal of Futures Markets 23, 615-
646.
Pan, J., 2002. The jump-risk premia implicit in options: evidence from an integrated
time series study. Journal of Financial Economics 63, 3-50.
Poon, S., and Granger, C., 2003. Forecasting volatility in financial markets: A review.
Journal of Economic Literature 41, 478-539.
Poon, S., and Granger, C., 2005. Practical issues in forecasting volatility. Financial
Analyst Journal 61, 45-56.
Pritsker, M.,2006. The hidden dangers of historical simulation. Journal of Banking and
Finance 30, 561-582.
Skiadopoulos, G., 2004. The Greek implied volatility index: construction and
properties. Applied Financial Economics 14, 1187-1196.
Whaley, R., 2000. The investor fear gauge. Journal of Portfolio Management 26, 12-17.
 .12 .12

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 99% VaR-Implied(VDAX)-t RT 99% VaR-GJR+Implied(VDAX)-t

.12 .12
132

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 99% VaR- FHS(GJR) RT 99% VaR-RiskMetrics-t

Figure 4. Daily 99% VaR forecasts versus returns on DAX 30 index from January 01, 1992 to May 29, 2009.
 .12 .12

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 97.5% VaR-Implied(VDAX)-t RT 97.5%VaR-GJR+Implied(VDAX)-t

.12 .12
133

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 97.5% VaR-FHS(GJR) RT 97.5% VaR-RiskMetrics-t

Figure 5. Daily 97.5% VaR forecasts versus returns on DAX 30 index from January 01, 1992 to May 29, 2009.
 .12 .12

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 95% VaR-Implied(VDAX)-t RT 95% VaR-GJR+Implied(VDAX)-t

.12 .12
134

.08 .08

.04 .04

.00 .00

-.04 -.04

-.08 -.08

-.12 -.12

92 94 96 98 00 02 04 06 08 92 94 96 98 00 02 04 06 08

RT 95% VaR-FHS(GJR) RT 95% VaR-RiskMetrics-t

Figure 6. Daily 95% VaR forecasts versus returns on DAX 30 index from January 01, 1992 to May 29, 2009.
 135

Modeling Risk Factors Driving the EUR, USD, and GBP

Swaption Volatilities

Ihsan Ullah Badshah

Hanken School of Economics, Department of Finance and Statistics, P.O. Box 287,
FIN-65101 Vaasa, Finland. Phone: +358-6-3533 721, Fax: +358-6-3533 703,
Email: ibadshah@hanken.fi

September 20, 2010

Abstract

The purpose of this essay is twofold: First, we use principal component analysis to
identify the common implied risk factors affecting implied volatility (IV) movements in
the EUR, USD and GBP swaption markets. We then examine the dynamic interactions
between the implied factors (i.e., factors in each of the EUR, USD, and GBP swaption
IVs) by using techniques such as Granger causality and the generalized impulse
response function (IRF). Second, we calibrate the string market model (SMM) for each
of the swaption markets using multivariable nonlinear optimization in order to
reproduce the swaption volatility matrix. We find that the first three implied factors
explain 94 -97% of the variance in each of the EUR, USD and GBP swaption IVs. All of
the significant factors present high degrees of correlation across swaption markets.
Consequently, there are significant linkages across the factors and bi-directional
causality is at work between the factors implied by EUR and USD swaption IVs.
Furthermore, in innovation accounting investigations, the factors implied by the EUR
and USD swaption IVs respond to each others’ shocks, and the factors implied by the
GBP IVs respond to their shocks as well. However, a shock to the GBP factors does not
affect their factors. On the other hand, calibration results show that the SMM can
efficiently reproduce the whole swaption volatility matrix for each of the EUR, USD,
and GBP markets. We obtain optimal solutions for the EUR, GBP and USD markets.
Finally, we find similar characteristics between the EUR and the GBP swaption
markets, but different characteristics in the USD swaption market. The implications of
these results are important for the pricing of interest rate derivatives, Vega-hedging,
risk management, and policy making.

Keywords: calibration, implied volatility, linkages, principal component analysis,

swaption

JEL Classification: G12, G13, E4

The Author would like to thank Johan Knif, Mika Vaihekoski, George Skiadopoulos, Seppo Pynnonen,
Gregory Koutmos, Kenneth Högholm, James Kolari, and Jussi Nikkinen for providing useful comments.
The author would also like to thank to the participants of the Graduate School of Finance (GSF) Research
workshops for their useful comments. The author acknowledges Evald & Hilda Nissi and HANKEN
foundations for providing financial support.
 136

1 Introduction

The interest rate (IR) derivatives market has grown exponentially since the trading of
the first IR swap security gave birth to many complex (exotic) derivatives products
afterwards. The 1990s, in particular, saw phenomenal pace in the introduction of new
types of products. Besides the plain-vanilla derivatives, there were many other products
such as Bermudan swaptions, indexed-principal swaps knockout caps, ratchet caps,
callable inverse floaters, index accruals and many more (see e.g., Rebonato, 2002). In
fact, according to the International Swaps and Derivatives Association (ISDA), the IR
derivatives market is the largest derivatives market in the world. ISDA reported that
the notional amount of IR derivatives outstanding at the end of 2007 was 382 trillion
US dollars. Another detailed report was published by Bank of International Settlements
(BIS) at the end of 2007, showing that euro-denominated IR derivatives form a
substantial amount of this proportion (about $146 trillion), in addition to US dollar
(about $129 trillion), Japanese yen (about $53 trillion), and UK pound sterling (about
28 trillion). Thus, IR derivatives saw phenomenal growth in outstanding notional
amounts and consequently, huge development of different products and models to
support traders. Hence, this area has attracted many professionals from finance,
physics, mathematics and statistics.

However, IR markets present a high degree of correlation—a high dependence among

the market risk factors. When a few important sources of information are common to
market risk factors, there is a high degree of correlation among the risk factors. For
traders required to price and hedge huge portfolios, such portfolios might consist of
different IR securities and depend on hundreds of underlying risk factors. In order to
enhance efficiency and achieve parsimony, dimensionality reduction is necessary and
can be achieved through principal component analysis (PCA). PCA is a technique that
extracts the significant underlying risk factors from correlated IR securities (see e.g.,
London, 2004). Similarly, swaption implied volatilities (IVs) can be viewed as highly
correlated and can be explained by only two to three independent factors. The risk
factors implied by IVs could be used for a variety of purposes: vega-hedging, generating
implied volatility surface (IVS), model calibration, and risk management. For example,
plain-vanilla options are actively traded (i.e., in hundreds) with a variety of maturities
in order to estimate sensitivities of a portfolio with respect to each of these sources of
risk--the sensitivities can be consistently computed by using the significant principal
 137

components (PCs).104 Second, the IVS can be projected onto the PCs. IVS are important
for traders in quoting options for thinly traded over-the-counter (OTC) market. Third,
the pricing model can be calibrated using only a few important PCs instead of a bunch
of underlying securities.105 Finally, the value-at-risk for the IR portfolios can be
estimated using the PCs. There is an extensive literature that identifies the common
PCs in term structure movements (see e.g., Litterman and Scheinkman, 1991; Knez,
Litterman, and Scheinkman, 1994; and a recent study on the swap curve dynamics by
Huang et al.,2008). However, besides Wadhwa (1999) we have not found any other
study identifying the PCs affecting volatility movements in the IR swaption markets.
Wadhwa (1999) studied USD LIBOR IVs movements, who found that four significant
PCs explained a large proportion of the total variation in the cap and swaption IVs
while using daily data from March 1992 to March 1999. Further, it was found that the
first PC was stable over time and never changed its shape, and the PCs beyond the first
PC were unstable. Bruggemann et al. (2008) modeled and analyzed the stochastic
properties of factors of IVs (i.e., implied from the options on the German stock market)
in a VAR framework. The VAR framework nicely described the dynamic linkages
between factors that determine the movements of the IVs. It was suggested that the
interaction between the first two factors (representing level and maturity slope) carries
useful risk management information. The relationship of the factors to movements in
economic variables was also investigated, and it was found that factor loadings are
linked to U.S. stock market returns and volatility. U.S. stock market returns and
volatility affect German stock market IV and the effect was found to be bi-directional.

On the other hand, let us consider the market models and their calibration procedures.
For instance, a trader of vanilla options (such as swaptions and caps) would expect his
model to price the required hedges (swaps and bonds) for each trade accurately and in
line with the vanilla options market. Equally, the exotic trader would expect his model
to price hedges (swaptions and caplets) in line with the market.106 A pricing model is
then calibrated to the vanilla market, facilitating to determine the price of the exotic
product and hedging positions of the exotic product.107 Thus, for accurate and efficient
pricing of exotic products, it is of paramount importance to find a parsimonious model
that accurately matches the market quoted prices. In this respect, Heath, Jarrow and

104 We use risk factors and PCs equivalently throughout this essay.
105 A good example is the LIBOR market model.
106 For detail on the calibration procedures, see Rebonato (2002).
107 Since quoted prices of the plain-vanilla products are available in the market, calibrating the model to the

market prices allows us to extract information about the distribution of the underlying interest rates.
 138

Morton (1992) (HJM) were the first to develop a model on the firm ground of
mathematical finance; this model is a framework for the continuously compounded
instantaneous forward rate (FR) models. However, this tractable model does not
guarantee that forward LIBOR rates have lognormal volatility structures. Therefore, the
HJM models conflict with the lognormal Black (1976) model commonly used in the
market. Later, Brace, Gatarek and Musiela (1997) (BGM) developed the LIBOR market
model (LMM) to overcome this shortcoming and were thereby consistent with both the
Black (1976) model and the HJM framework. LMM models the dynamics of observable
forward LIBOR rates directly, in contrast to instantaneous FRs. However, these models
share many of the same calibration issues and can be implemented with a small
number of factors. In practice, models require frequent recalibration, i.e., the constant
parameters of the model are often required to change in order to suit current market
conditions. The recalibration of the above models may ignore some hidden sources of
risks, because the models do not account for the dynamics of every FR. The string
market model (SMM) accounts for all hidden sources of risks in the market. The
stochastic string models were first developed by Kennedy (1994, 1997), who modeled
the evolution of the term structure of FRs as a stochastic string. Later, Goldstein
(2000), Longstaff and Schwartz (2001), Santa-Clara and Sornette (2001) and Longstaff
et al. (2001) developed similar models that allow for correlated strings of shocks to the
FR curve. The dynamics of each FR are modeled in the SMM model. Moreover, it
accounts for the correlations along each point of term structure. This model is also
called high dimensional (continuum of FRs). The calibration of the SMM model is more
straightforward than that of models with fewer factors (i.e., the LIBOR market model).
The parameters of the SMM model are directly estimated from the covariance matrix
and the parameterization of the SMM becomes more advantageous and parsimonious
than the LMM as the number of factors driving the term structure of FRs increases (see
e.g., Longstaff et al., 2001).

This essay has two purposes. First, we aim to examine the dynamics of swaption IVs for
each of the EUR, USD and GBP markets. Our goal is to characterize the number of
significant implied PCs required to explain the whole swaptions maturity surface and to
examine the dynamic linkages of the factor loadings implied from the three swaptions
markets. Second, we aim to calibrate the SMM model to the swaptions markets in a
discrete fashion, thereby reproducing the whole swaption volatility matrix for each the
 139

EUR, USD, and GBP swaption markets.108 Thus, this study is related to that of Wadhwa
(1999) in terms of PCA and USD cap/swaptions, whereas the second part of the essay is
related to that of Longstaff et al. (2001) and Kerkhof and Pelsser (2002), who
calibrated the SMM to USD cap/swaption data. However, our study differs in several
important respects. First, we apply PCA to the EUR, USD and GBP swaption IVs,
whereas Wadhwa (1999) studied only the dynamics of USD cap/swaptions. Moreover,
we compare the significant PCs implied from the three markets and examine their
dynamic linkages. Second, we calibrate the SMM to the swaptions matrix for each of
the EUR, USD, and GBP swaption markets using a nonlinear multivariable
optimization, whereas the studies of Longstaff et al. (2001) and Kerkhof and Pelsser
(2002) only calibrated to the USD swaption data.

The main findings are that only three PCs explain 94-97% of the variations in each of
the EUR, USD, and GBP swaption IVs. The PC1s (which stand for the parallel shifts) are
found to be highly correlated across the three markets. The remaining PC2s and PC3s
are also correlated. As a result, we find considerable linkages across the swaption
markets. Bi-directional causality is at work between the EUR and USD factors. In
innovation accounting investigations, unit shocks to the EUR_PC1 and USD_PC1 have
significant influences on each other. Both influence GBP_PC1 as well, but not vice versa
(thus GBP_PC1 is found to be the least influential among the PC1s). Similar patterns
are found for the PC2s. On the other hand, after calibration of the SMM model to
swaption matrices, each of the three markets provides different results. The EUR
swaption optimization requires 1446 iterations (and 2000 function evaluations) to
obtain an optimal solution, the USD optimization requires 1433 iterations (and 2000
function evaluations), and the GBP optimization requires the least iterations among the
three optimization process: 1429 iterations (and 2000 function evaluations). Thereby
the smallest level of RMSE is observed for the EUR, followed by the GBP and USD
markets. Our results imply that the EUR and GBP swaption IVs contain identical
characteristics both for the PCA and calibration processes, but are somewhat different
from those of the USD market.

This essay is organized as follows. Section 2 discusses the swaptions market. Section 3
discusses the data. Section 4 examines factor loadings dynamics in a VAR framework.

108Note that we use swaption volatility surface and swaption volatility matrix equivalently throughout this
essay.
 140

Section 5 calibrates the SMM model for each of the EUR, USD, and GBP swaption
markets. Section 6 summarizes and concludes.

2 The Swaptions Market

A swaption is an OTC-traded option that grants its owner the right to enter the
underlying swap contract, where the contract can be traded on a variety of IR swap
securities. There are two types of swaptions. In one type, the payer swaption, the owner
of the swaption is granted the right to enter into the IR swap, where he receives the
floating rate and pays the fixed rate. The second type of swaption is the receiver
swaption, which grants the owner of the swaption the right to enter into the IR swap
where he pays the floating rate and receive the fixed rate. Therefore, a swaption can be
viewed as call or put on coupon bonds.109 The market makers for the swaptions in the
major currencies (i.e., US dollar, euro, Japanese yen, and UK pound) are the big
investment and commercial banks. Whereas, the main participants in the swaption
markets are banks, firms, hedge funds, and other financial institutions. Banks and
firms normally use swaptions to manage IR risk arising from their borrowings or
businesses. For instance, if a bank holds a mortgage portfolio and wants to safeguard
against lower IRs (as lower IR will lead to early prepayment of the mortgages), the bank
may buy a receiver swaption. Likewise, a firm may want to manage the risk of higher
IRs by buying a payer swaption. A hedge fund that expects that the future IR will rise by
a certain percentage will sell a payer swaption, earning money in the form of swaption
premium. However, in practice, swaptions are quoted in terms of at-the-money (ATM)
IVs while using the Black (1976) model, the practitioner’s standard swaption model.
European swaptions are priced using the forward swap rate as an input to the Black
(1976) model. Jamshidian (1996) showed that the Black model is arbitrage-free under
the assumption of a lognormal swap rate.

Payer ¨¨

§ 1  1 (1  F/n) txn  ·¸e >FN(d1)  KN(d2)@ ,
 rT
(1)
F ¸
© ¹

Receiver

§ 1  1 (1  F/n) txn
¨¨
 ·¸e >KN(  d2)  KN(d1)@ ,
 rT
(2)
F ¸
© ¹
where

109 For more discussion see for instance, Longstaff et al. (2001).
 141

ln(F/K)  ( 2 /2)T
d1 , (3)
 T
d2 d1   T . (4)
F is the forward swap rate, K is the strike rate of the swaption, t is the tenor of the swap,
 is the volatility of the forward swap rate, n is the yearly compounding of the swap
rate, and T is the time to expiration. Thus, swaption IVs can be obtained by inverting
the Black (1976) formula as shown above to obtain the IVs associated with the selected
pair of maturity and tenor.

3 Data

We use EUR-, USD- and GBP-denominated IR datasets for the analysis, as well as data
on money market rates and swap rates, and Black (1976)-IVs of swaptions.110 We have
weekly data on money market rates with maturities of 1, 3, 6, 9 and 12 months, and
data on EUR, USD, and GBP swap rates with maturities from 2 years up to 15 years.
The IR datasets are used to construct the FR curve and then the bootstrap technique is
used to arrive at the discount factors. This is essential for calibrating the SMM model
later on. On the other hand, the second type of datasets is the time series ATM Black-
IVs using weekly data backed out of swaptions for expiries of 3 months, 6 months and
1-8 years, and struck on the underlying swaps of 1-10 years maturities. The time series
data of swaption IVs for the EUR, USD and GBP markets from April 1, 2006 through
October 31, 2008.

110 Data are provided by ICAP, London, United Kingdom.

 142

4 Factor Loading Dynamics

4.1 Factor Loadings

We apply PCA to reduce the swaptions maturity surface to only a few risk factors and
model the implied dynamics in a vector-autoregressive (VAR) framework (i.e., for the
EUR, USD, and GBP swaption markets). The rationale is to extract the independent
risk factors that explain most of the dynamics of the swaption IVs and to examine the
dynamic linkages across markets. Previously many researchers attempted to extract
those risk factors that explains most the dynamics in the volatility smirk (skew) or
surfaces for equity index options; for FTSE 100 index IV by Badshah (2008), and Cont
and da Fonseca (2002); for S&P 500 index IV, we consider Cont and da Fonseca
(2002), and Skiadopoulos et al. (1999); for DAX 30 index IV, Fengler et al. (2003),
whose findings confirmed that about 70-90% of the total variation in the surface or
smirk (skew) can be attributed to only three PCs. A closer study to ours is of
Bruggemann et al. (2008) who analyzed and modeled the stochastic properties of the
factors of IVs (i.e., obtained from options on German stock market) in a VAR
framework and found significant interactions between the factor loadings.

Thus, significant PCs are extracted for each of the EUR, USD, and GBP swaption IVs,
and their implied linkages across markets are examined in a VAR framework: 111 First,
weekly time-series data from April 1, 2006 to October 31, 2008 are assembled into
three groups: EUR, USD, and GPB. Then, the swaption IVs are incorporated as follows:
3M×1Y, 3M×2Y,…, 3M×10Y; 6M×1Y, 6M×2Y,…,6M×10Y; 9M×1Y, 9M×2Y,…,9M×10Y;
1Y×1Y, 1Y×2Y,…,1Y×10Y and so on up to 8Y×1Y, 8Y×2Y,…,8Y×10Y.112 Second, we
compute the first log-difference IVs across different expiries and time-to-maturities.
Finally, we extract the significant PCs from swaption IVs for each market.

PCA is a method of matrix decomposition into eigenvectors and eigenvalue matrices. It

is applied to a pool of ATM swaption IVs across different expiries and maturities,
T
effectively decomposing the covariance matrix as  , where the diagonal elements
of  are the eigenvalues and the columns  are the associated eigenvectors. The
choice of column labeling in  allows PC ordering such that e1 belongs to the largest

111Bruggemann et al., 2008 used a similar VAR model for factor loading dynamics of IVs implied from the
German option market.
112 M symbolizes month, while Y symbolizes year.
 143

eigenvalue 1 , e2 belongs to second largest eigenvalue 2 and so forth. In highly

correlated IVs, the first eigenvalue would be much larger than the others.
Consequently, the first PC can explain much of the variation in the swaption IVs. If the
first three PCs explain most of the variations in the swaption IVs, then these PCs can
replace the original volatility variables without loss of much information (see e.g.,
Skiadopoulos et al.,1999; London, 2004). That is, the original volatility input data may
be written as a linear combination of the PCs, which reduces the dimensions of the
system.

Table 1 provides results for the PCA on the swaption IVs for the swaption markets
considered from April 1, 2006 through October 31, 2008. Each panel of Table 1
presents the cumulative proportion of variance explained by the first eight PCs.
Looking at the eigenvalues, we find that three PCs are significant, i.e., with eigenvalues
greater than 1, in EUR swaption IVs. The eigenvalues in the fourth PC in the USD and
GBP swaption IVs are also greater than 1. The proportions of the fourth PC that
contribute to the cumulative values are negligible, implying that only three PCs are
sufficient to capture much of the systematic variations for each of the three swaption
markets.113

The interpretations of the PCs are intuitively similar to those of Longstaff et al. (2001)
and Wadhwa (1999). The first PC generates parallel shifts in the term structure for
swaption IVs and captures about 86%, 88% and 80% for EUR, USD, and GBP,
respectively. The second PC that generates shifts in the slope of the term structure of
IVs captures an additional 9%, 6%, and 12% for EUR, USD, and GBP, respectively.
Finally, the third PC is the curvature factor, which generates movements in the term
structure of the IVs such that short-term and long-term IVs move in opposite directions
from the mid-term IVs and account for 1.5%, 2%, and 2% for EUR, USD, and GBP,
respectively. Thus, the first three PCs cumulatively explain about 96%, 97%, and 94% of
the variation in the IVs for each the EUR, USD and GBP swaption markets.

113 Here our results are consistent with the results of Wadhwa (1999), who found three significant PCs in

the n .
 144

Table 1. Principal Component Analysis of EUR, USD, and GBP swaption IVs

Panel A: EUR Swaption IVs

Number Eigenvalues Difference Proportion Cumulative Cumulative Prop.

1 86.09613 77.15441 0.8610 86.09613 0.8610
2 8.941718 7.387048 0.0894 95.03785 0.9504
3 1.554670 0.684989 0.0155 96.59252 0.9659
4 0.869682 0.350749 0.0087 97.46220 0.9746
5 0.518933 0.227666 0.0052 97.98113 0.9798
6 0.291267 0.014989 0.0029 98.27240 0.9827
7 0.276279 0.110757 0.0028 98.54868 0.9855
8 0.165522 0.038358 0.0017 98.71420 0.9871
Panel B: USD Swaption IVs

Number Eigenvalues Difference Proportion Cumulative Cumulative Prop.

1 88.43159 82.02862 0.8843 88.43159 0.8843
2 6.402971 4.413318 0.0640 94.83456 0.9483
3 1.989653 0.828376 0.0199 96.82421 0.9682
4 1.161277 0.756448 0.0116 97.98549 0.9799
5 0.404829 0.150173 0.0040 98.39032 0.9839
6 0.254655 0.039919 0.0025 98.64498 0.9864
7 0.214736 0.050626 0.0021 98.85971 0.9886
8 0.164110 0.037694 0.0016 99.02382 0.9902
Panel C: GBP Swaption IVs

Number Eigenvalues Difference Proportion Cumulative Cumulative Prop.

1 80.28617 67.94028 0.8029 80.28617 0.8029
2 12.34590 10.36036 0.1235 92.63207 0.9263
3 1.985536 0.916064 0.0199 94.61760 0.9462
4 1.069472 0.284469 0.0107 95.68708 0.9569
5 0.785003 0.235437 0.0079 96.47208 0.9647
6 0.549566 0.269289 0.0055 97.02165 0.9702
7 0.280277 0.034339 0.0028 97.30192 0.9730
8 0.245938 0.054895 0.0025 97.54786 0.9755

In sum Table 1 shows that the first three PCs cumulatively account for about 94-97% of
the total variation in each of the EUR, USD, and GBP swaption IVs. The parallel shifts
are evident in the USD, followed by the EUR and GBP swaption IVs, respectively. This
 145

implies that the participants in the EUR swaption market are more concerned about
parallel shifts than about the other two underlying risk factors. The remaining two risk
factors are more important for the GBP and EUR market participants, respectively.

Figure 1 reports factor loadings for the first three PCs extracted from the time series
data for each the EUR, USD, and GBP swaption IVs from April 1, 2006 to October 31,
2008. The x-axis contains variables (i.e., IVs for a variety of maturities for instance
1M*1Y… 8Y*10Y), whereas the y-axis contains loadings for the three PCs.114 The PC1s
that correspond to parallel shifts in the swaption IVs are more stable and carry positive
weights throughout the sample period. PC1s are important for the shorter and medium
maturity swaptions, which are increasing, while the longer maturity swaptions remain
more stable and thus show a decreasing effect. Imply that the parallel shifts are
increasing and an important factor for the participants in the swaption markets during
the sample period is influential for all swaptions. In particular, short and medium
maturity swaptions are the most sensitive to parallel shifts. On the other hand, the
PC2s correspond to shifts in the slope of the swaption IVs. Here, the PC2s carry positive
weights for short maturity swaptions and gradually enter into negative territory,
moving downward as the term increases. This implies that the PC2s have generated
positive shifts in the term structure slope of shorter maturity swaption IVs and negative
shifts in the term structure slope of longer maturity swaption IVs. Finally, the PC3s are
viewed as the curvature factor that essentially influences swaption IVs. They carry
positive weights for the very short and very long maturity swaption IVs, and show
negative jumps for the medium term swaption IVs.

Figure 1 provides interesting results. EUR and GBP swaption IVs are affected by the
first three PCs and show somewhat similar characteristics, while the USD swaption IVs
are heavily affected by the first PC (i.e., parallel shifts are important for USD
swaptions) and to some extent by the second PC. Market participants in the USD
swaption market are concerned with short-term future changes in the interest rates,
whereas the additional two factors are essential for EUR and GBP markets.

114 About 100 variables are included in the principal component analysis.
 146

.11

.10

.09

.08

.07
10 20 30 40 50 60 70 80 90 100

EUR_PC1 USD_P C1 GB P_PC1

.24

.20

.16

.12

.08

.04

.00

-.04

-.08

-.12
10 20 30 40 50 60 70 80 90 100

E UR_P C2 US D_P C2 GB P _P C2

.3

.2

.1

.0

-.1

-.2

-.3
10 20 30 40 50 60 70 80 90 100

E UR_P C3 US D_PC3 GBP _P C3

Figure 1. Factor loadings on the first three PCs for the EUR, USD, and GBP Swaption
IVs.
 147

4.2 Correlation Analysis

Correlation analysis helps to quantify the extent to which important PCs are correlated
in and across markets. Table 2 presents the correlation matrix for the EUR, USD, and
GBP swaption PCs.

Table 2. Correlation matrix for the three PCs of the EUR, USD, and GBP swaption
EUR USD GBP
PC1 PC2 PC3 PC1 PC2 PC3 PC1 PC2 PC3
EUR PC1 1

PC2 -0.770 1
[-11.9]
(0.00)

PC3 -0.221 -0.001 1

[-2.24] [-0.01]
(0.03) (0.99)

USD PC1 0.969 -0.690 -0.231 1

[39.23] [-9.43] [-2.34]
(0.00) (0.00) (0.02)

PC2 -0.760 0.993 -0.005 -0.679 1

[-11.5] [85.49] [-0.05] [-9.16]
(0.00) (0.00) (0.959) (0.00)

PC3 -0.414 0.0212 0.916 -0.441 -0.001 1

[-4.50] [0.20] [22.59] [-4.86] [-0.00]
(0.00) (0.834) (0.000) (0.00) (0.99)

GBP PC1 0.969 -0.721 -0.1975 0.952 -0.698 -0.427 1

[38.9] [-10.29] [-1.99] [30.81] [-9.65] [-4.67]
(0.00) (0.00) 0.0488 (0.00) (0.00) (0.00)

PC2 -0.721 0.991 -0.0627 -0.642 0.992 -0.052 -0.6677 1

[-10.2] [74.99] [-0.62] [-8.28] [76.10] [-0.51] [-8.88]
(0.00) (0.00) (0.53) (0.000) (0.00) (0.61) (0.00)

PC3 -0.265 0.0551 0.950 -0.271 0.048 0.8756 -0.225 -0.001 1

[-2.71] [0.54] [30.21] [-2.78] [0.47] [17.94] [-2.28] [-0.01]
(0.00) (0.58) (0.00) (0.00) (0.63) (0.00) (0.02) (0.99)

In the square brackets t-statistics whereas in parenthesis probability values.

As can be seen from the correlation coefficients for the PC1s (which generate parallel
shifts in swaption IVs), the three swaption markets are highly correlated and
 148

statistically significant. For example, the correlation between the PC1s for the EUR and
USD swaption IVs is positive, with the corresponding coefficient high at 0.97. A
similarly high correlation coefficient is found between the EUR and the GBP. A
correlation coefficient of 0.95 is observed between the USD and the GBP. Second, PC2s
that generate shifts in the slope of the swaption IVs are negative, correlations with the
PC1s ranging from a maximum of -0.72 to a minimum of -0.69. Third, the PC3
curvature factors in the swaption IVs. As can be observed, the PC3s are negative,
correlations with PC1s ranging from a maximum of -0.43 to a minimum of -0.23.
Finally, we found that PC2s and PC3s are uncorrelated for the three considered
swaption markets, and that the corresponding coefficients are not significant even at
the 5% level.

The conclusion drawn from the results in Table 2 is that that implied PC1s for the three
swaption markets are highly positively correlated. However, PC2s and PC3s are
negatively correlated with PC1s. The highest correlation is between the USD and the
EUR, whereas the USD and the GBP are the least correlated among the three swaption
markets.
 149

4.3 VAR Model for Factor Loading Dynamics

The VAR model is used to investigate the dynamic interactions between the first two
PCs (i.e., PC1s, PC2s extracted from each of the EUR, USD, and GBP Swaption IVs).115
The VAR model was developed by Sims (1980) that successfully treats each endogenous
variable in a system as a function of the lagged values of all endogenous variables in the
dynamic simultaneous equation system. Hence, the VAR model for the factor loadings
dynamics can be represented mathematically as:

L
Ft   ¦  i Ft  i  u t , (5)
i 1

where Ft (z t1EUR ,z t2EUR ,z t1USD ,z t2USD ,z t1GBP ,z t2GBP )c is an m u 1 vector of endogenous variables

representing the factors in the EUR, USD, and GBP swaption IVs; ^i , i 1, 2, 3, ˜ ˜ ˜ , L`

is an m u m matrix of coefficients; and u t is an m u 1 vector of innovations that can be

contemporaneously correlated and uncorrelated with its own lagged values and with
other variables. Accordingly, the VAR model is estimated by the OLS estimation
method. However, for selecting a suitable number of lags for the dynamic VAR system,
we use a parsimonious Schwartz information criterion (SIC) that suggests two lags.
Consequently, analysis with Equation 5 is conducted using two lags.

To investigate the direction between PCs, we apply the Granger (1969) causality test,
which establishes the causal relationship between the PCs and hence confirms the lead-
lag relationship. The Granger causality test establishes that PC i is Granger-caused by

PC j if the information in the past and present values of PC j helps to improve the

forecasts of PC i . The results for the PCs are reported in Table 3 for lag 2 with

corresponding values of the F-tests. Bi-directional causality is at work between

EUR_PC1 and USD_PC1, and EUR_PC1 and USD_PC1 Granger-causes GBP_PC1, but
not vice-versa. On the other hand, there is significant bi-directional causality between
EUR_PC2 and USD_PC2, and they both influence GBP_PC2. Lastly, EUR_PC2 is
found to Granger-cause EUR_PC1 and USD_PC1 and similar results are found for
USD_PC2.

115 The first two PCs are selected because they are able to explain an average of 94% of the variation in each

of the market IVs. We do not consider PC3s in our VAR framework. This also helps to maintain a
reasonable number of parameters, increase efficiency, and achieve parsimony.
 150

Table 3. Granger causality for implied PCs.

Null Hypothesis 2 lags
F-Statistics P-Value
EUR_PC2 does not Granger Cause EUR_PC1 1.2394 0.294
EUR_PC1 does not Granger Cause EUR_PC2 1.5349 0.220
USD_PC1 does not Granger Cause EUR_PC1 8.5325*** 0.000
EUR_PC1 does not Granger Cause USD_PC1 6.5000*** 0.002
USD_PC2 does not Granger Cause EUR_PC1 0.2017 0.817
EUR_PC1 does not Granger Cause USD_PC2 0.5404 0.584
GBP_PC1 does not Granger Cause EUR_PC1 0.0425 0.958
EUR_PC1 does not Granger Cause GBP_PC1 3.3994** 0.037
GBP_PC2 does not Granger Cause EUR_PC1 0.0117 0.988
EUR_PC1 does not Granger Cause GBP_PC2 1.0068 0.369
USD_PC1 does not Granger Cause EUR_PC2 1.6799 0.192
EUR_PC2 does not Granger Cause USD_PC1 7.1236*** 0.001
USD_PC2 does not Granger Cause EUR_PC2 2.8396* 0.063
EUR_PC2 does not Granger Cause USD_PC2 1.2269 0.297
GBP_PC1 does not Granger Cause EUR_PC2 1.4088 0.249
EUR_PC2 does not Granger Cause GBP_PC1 0.4363 0.647
GBP_PC2 does not Granger Cause EUR_PC2 0.7087 0.494
EUR_PC2 does not Granger Cause GBP_PC2 3.9015** 0.023
USD_PC2 does not Granger Cause USD_PC1 4.8363** 0.010
USD_PC1 does not Granger Cause USD_PC2 2.4546* 0.091
GBP_PC1 does not Granger Cause USD_PC1 1.8424 0.164
USD_PC1 does not Granger Cause GBP_PC1 6.2286*** 0.002
GBP_PC2 does not Granger Cause USD_PC1 0.7008 0.498
USD_PC1 does not Granger Cause GBP_PC2 0.8442 0.433
GBP_PC1 does not Granger Cause USD_PC2 2.2721 0.108
USD_PC2 does not Granger Cause GBP_PC1 0.0189 0.981
GBP_PC2 does not Granger Cause USD_PC2 1.3948 0.253
USD_PC2 does not Granger Cause GBP_PC2 10.381*** 0.000
GBP_PC2 does not Granger Cause GBP_PC1 0.1065 0.899
GBP_PC1 does not Granger Cause GBP_PC2 0.9315 0.397

***, **, and * Denote rejection of the null hypothesis at the 1% , 5%, and 10% significance levels respectively.

Granger causality in a VAR system using the F-test provides information about which
variables impact the future values of each of the variables in the VAR system. However,
F-test values do not provide the sign of the relationship, speed or persistence. The
 151

VAR’s impulse response functions could provide information about this kind of
dynamic relationship (see, e.g., Brooks, 2002). An impulse response function (IRF)
measures the responses of the variables in the dynamic VAR system (in our case, the
first two factors of each the EUR, USD, and GBP swaption IVs) when a shock is given to
each factor. Thus, a one standard error shock is applied to the error of a factor, and the
effect on the dynamical VAR system over a specified period of time is recorded.
Pesaran and Shin (1998) proposed a generalized IRF, which is preferred since it does
not require orthogonalization of shocks and is invariant to the reordering of variables in
the VAR system. A mathematical illustration of IRFs is provided below, the moving
average representation of the Equation (5):

f
Ft   ¦  i u t i , (6)
i 1

where Ft is an m-variate stochastic process;  is the mean of the process; and  i is

m u m MA matrices containing the responses to forecast errors u t that occurred i

periods ago. Hence,  i is computed through looping relationships. The IRF computes

the time profile of the effect of shocks on future values of the factor loadings in a
dynamic system at a given time. Here, the  i matrices are the dynamic multipliers of

the system, representing the model’s response to a unit shock in each of the factors. The
response of Fj to a unit shock in Fk is specified by the following sequence, called the

IRF:
 jk,1 , jk,2 , jk,3 ˜ ˜ ˜ , (7)

where  jk,i is the jkth element of matrix  i . Since we know that (u t u tc ) ¦ , the

components of u t are contemporaneously correlated. Pesaran and Shin (1998) suggest

that whenever these correlations are high, separation of the response of Fj to Fk from

its response to other shocks that are correlated with u kt is unattainable and the

resulting values obtained from simulation will be misleading. Therefore, generalized

IRFs should be considered instead. The generalized IRF shocks only one element of u t ,

e.g.,the kth element, and integrates out the effects of the error distribution of the other
shocks (i.e., the correlations among u t are taken into account). Assume that u t has a

multivariate normal distribution. For non-normal error distribution, stochastic-

simulation is required to obtain the conditional expectations (u t u kt  k ) (see, e.g.,
 152

Koop et al., 1996; Pesaran and Shin, 1998; Mills, 1999). The generalized IRF (in
combination with stochastic-simulations) KU:

(u t u kt k ) (1k , 2k , . .  Kk )c  kk1  k ¦e 

k
1
kk k , (8)

^
Where  jk ,j,k 1, ˜ ˜ ˜ K ` represents the elements of ¦ and e k is a selection vector.

Hence, the response vector to a shock in variable k that occurred i periods ago is:

 i ¦ ek k
. (9)
 kk  kk
Finally, the generalized IRF that measures the impact of a shock of one standard error
in scaled form at time t on future values of f at time t+i as follows:

 kg (i)  kk1/2  i ¦ e k , i 0, 1, 2, ˜ ˜ ˜ ,. (10)

Figure 2 translates the generalized IRFs for each EUR_PC1, EUR_PC2, USD_PC1,
USD_PC2, GBP_PC1, and GBP_PC2 to a unit shock in the innovations of each implied
PC (i.e., the same shock applied in each period starts at period 1 and ends at period 10).
For instance, the EUR_PC1 IRFs to a shock of one standard deviation (STD) in the
innovations of EUR_PC1, EUR_PC2, USD_PC1, USD_PC2,GBP_PC1, and GBP_PC2
are plotted in Figure 2, where the number of periods ahead is on the x-axis and the
IRFs in units are reported on the y-axis.116 A unit shock is applied to EUR_PC1,
USD_PC1 and GBP_PC1 in period 1 and the corresponding IRFs are traced, resulting in
an increase in EUR_PC1 equal to 0.001794 units, 0.001506 units, and 0.001402 units,
respectively. However, a shock in EUR_PC2, USD_PC2 and GBP_PC2 generates
decreases in the EUR_PC1 equal to -0.001240 units, -0.001323 units, and -0.000716
units, respectively. Similarly, the same shock is applied to each PC and we capture the
IRFs in each period ahead. The effects of the PC1s are positive and persistent for
EUR_PC1, though the pattern is decreasing and the effects die out after 10 periods. The
effects of PC2s on EUR_PC1 are negative, and the negative effects decrease and tend to
zero after 10 periods.

Figure 2 shows the IRFs of USD_PC1 to a shock of one STD in the innovations of each
PC considered. Likewise, a unit shock in each EUR_PC1, USD_PC1 and GBP_PC1
induces an increase in USD_PC1 equal to 0.001460 units, 0.001738 units, and

116 The standard deviations of the IRFs are estimated with 10,000 Monte Carlo runs.
 153

0.001270 units, respectively. On the other hand, a shock in EUR_PC2, USD_PC2 and
GBP_PC2 induces a negative change in USD_PC1 equal to -0.001166 units, -0.001316
units, and -0.000856 units, respectively. If we look at the time profile of IRFs up to 10
periods, the effects of PC1s are positive and persistent on the USD_PC1, though
decreasing and tending towards zero after 10 periods. The effects of PC2s on USD_PC1
are negative and the negative effects are decreasing.

Figure 2 plots the IRFs of GBP_PC1 to a shock of one STD in the innovations of each
PC. A unit shock in each EUR_PC1, USD_PC1 and GBP_PC1 induces an increase in
GBP_PC1 equal to 0.001979 units, 0.001850 units, and 0.002533 units respectively.
However, a shock in EUR_PC2, USD_PC2 and GBP_PC2 induces a decrease in
GBP_PC1 equal to -0.001698 units, -0.001785 units, and -0.001322 units, respectively.
Similarly, the same shock in each PC provides IRFs for each period ahead. The effects
of the PC1s are positive and persistent on GBP_PC1, though they decrease and vanish
after 10 periods. The effects of PC2s on GBP_PC1 are negative and the negative effects
are decreasing.

The IRFs of EUR_PC2 to a unit shock in the innovations of each PCs are shown in
Figure 2. A unit shock is applied to each of EUR_PC2, USD_PC2 and GBP_PC2 that
induces a positive change in EUR_PC2 equal to 0.011322 units, 0.009673 units, and
0.008825 units respectively. However, a shock in EUR_PC1, USD_PC1 and GBP_PC1
induces a negative change in EUR_PC2 equal to -0.007828 units, -0.007597 units, and
-0.007592 units, respectively. We capture a time profile of IRFs and observe that the
effects of the PC2s are positive and persistent on EUR_PC2, though decreasing. The
effects of PC1s on EUR_PC2 are negative and decreasing. Second, we consider the IRFs
of USD_PC2 to a unit shock in the innovations of each PC. A shock in EUR_PC2,
USD_PC2 and GBP_PC2 induces a positive increase in USD_PC2 equal to 0.007032
units, 0.008231 units, and 0.006065 units, respectively. However, shocking EUR_PC1,
USD_PC1 and GBP_PC1 generates a decrease in USD_PC2 equal to -0.006070 units, -
0.006232 units, and -0.005802 units, respectively. A time profile of IRFs shows that
the effects of the PC2s are positive and persistent on USD_PC2, although decreasing.
The effects of PC1s on USD_PC2 are negative and decreasing. Finally, the IRFs of
GBP_PC2 to a unit shock in the innovations of each PC are plotted in Figure 2. A unit
shock in EUR_PC2, USD_PC2 and GBP_PC2 induces a positive change in GBP_PC2
equal to 0.011001 units, 0.010401 units, and 0.014114 units, respectively.
 154

Response to Generalized One S.D. Innovations ± 2 S.E.

R e s p o n s e o f EU R _ PC 1 to EU R _ PC 1 R e s p o n s e o f EU R _ PC 1 to EU R _ PC 2 R e s p o n s e o f EU R _ PC 1 to U SD _ PC 1 R e s p o n s e o f EU R _ PC 1 to U SD _ PC 2
.003 .003 .003 .003

.002 .002 .002 .002

.001 .001 .001 .001

.000 .000 .000 .000

-.001 -.001 -.001 -.001

-.002 -.002 -.002 -.002

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f EU R _ PC 1 to G BP_ PC 1 R e s p o n s e o f EU R _ PC 1 to G BP_ PC 2 R e s p o n s e o f EU R _ PC 2 to EU R _ PC 1 R e s p o n s e o f EU R _ PC 2 to EU R _ PC 2
.003 .003 .02 .02

.002 .002
.01 .01
.001 .001
.00 .00
.000 .000
-.01 -.01
-.001 -.001

-.002 -.002 -.02 -.02

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f EU R _ PC 2 to U SD _ PC 1 R e s p o n s e o f EU R _ PC 2 to U SD _ PC 2 R e s p o n s e o f EU R _ PC 2 to G BP_ PC 1 R e s p o n s e o f EU R _ PC 2 to G BP_ PC 2
.02 .02 .02 .02

.01 .01 .01 .01

.00 .00 .00 .00

-.01 -.01 -.01 -.01

-.02 -.02 -.02 -.02

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f U SD _ PC 1 to EU R _ PC 1 R e s p o n s e o f U SD _ PC 1 to EU R _ PC 2 R e s p o n s e o f U SD _ PC 1 to U SD _ PC 1 R e s p o n s e o f U SD _ PC 1 to U SD _ PC 2
.003 .003 .003 .003

.002 .002 .002 .002

.001 .001 .001 .001

.000 .000 .000 .000

-.001 -.001 -.001 -.001

-.002 -.002 -.002 -.002

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f U SD _ PC 1 to G BP_ PC 1 R e s p o n s e o f U SD _ PC 1 to G BP_ PC 2 R e s p o n s e o f U SD _ PC 2 to EU R _ PC 1 R e s p o n s e o f U SD _ PC 2 to EU R _ PC 2
.003 .003 .010 .010

.002 .002
.005 .005
.001 .001
.000 .000
.000 .000
-.005 -.005
-.001 -.001

-.002 -.002 -.010 -.010

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f U SD _ PC 2 to U SD _ PC 1 R e s p o n s e o f U SD _ PC 2 to U SD _ PC 2 R e s p o n s e o f U SD _ PC 2 to G BP_ PC 1 R e s p o n s e o f U SD _ PC 2 to G BP_ PC 2
.010 .010 .010 .010

.005 .005 .005 .005

.000 .000 .000 .000

-.005 -.005 -.005 -.005

-.010 -.010 -.010 -.010

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f G BP_ PC 1 to EU R _ PC 1 R e s p o n s e o f G BP_ PC 1 to EU R _ PC 2 R e s p o n s e o f G BP_ PC 1 to U SD _ PC 1R e s p o n s e o f G BP_ PC 1 to U SD _ PC 2

.004 .004 .004 .004

.002 .002 .002 .002

.000 .000 .000 .000

-.002 -.002 -.002 -.002

-.004 -.004 -.004 -.004

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f G BP_ PC 1 to G BP_ PC 1 R e s p o n s e o f G BP_ PC 1 to G BP_ PC 2 R e s p o n s e o f G BP_ PC 2 to EU R _ PC 1 R e s p o n s e o f G BP_ PC 2 to EU R _ PC 2

.004 .004 .02 .02

.002 .002 .01 .01

.000 .000 .00 .00

-.002 -.002 -.01 -.01

-.004 -.004 -.02 -.02

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

R e s p o n s e o f G BP_ PC 2 to U SD _ PC 1 R e s p o n s e o f G BP_ PC 2 to U SD _ PC 2 R e s p o n s e o f G BP_ PC 2 to G BP_ PC 1 R e s p o n s e o f G BP_ PC 2 to G BP_ PC 2

.02 .02 .02 .02

.01 .01 .01 .01

.00 .00 .00 .00

-.01 -.01 -.01 -.01

-.02 -.02 -.02 -.02

2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10

Figure 2. Generalized IRFs of the first two PCs implied from Swaption IVs for each of
the EUR, USD, and GBP swaption market.
 155

However, a shock in EUR_PC1, USD_PC1 and GBP_PC1 induces a decrease in

GBP_PC2 equal to -0.005632 units, -0.006951 units, and -0.007369 units,
respectively. The time profile of IRFs shows that the effects of the PC2s are positive and
persistent on GBP_PC2. However, the effects of PC1s on GBP_PC2 are negative and
decreasing.

We conclude from the IRFs plotted in Figure 2 that the PCs of EUR, USD and GBP
swaption IVs respond strongly to a unit shock in each PC of EUR and USD swaption
IVs. The effects are contemporaneous (period 1) and die out after 10 periods. However,
the PCs of both the EUR and the USD swaption IVs do not respond significantly to
shocks in any of the PCs implied from the GBP swaption IVs. Thus, the PCs of the GBP
market are found to be the least influential among the PCs. Nevertheless, the IRF
results presented in Figure 2 reconcile our results from the correlation analysis and the
Granger causality tests.
 156

5 Calibration of a Discrete String Market Model

5.1 The String Market Model

We also model and reproduce the swaption volatility matrix on a particular day using
orthogonal PCs in an SMM framework. Santa-Clara and Sornette (2001), and Longstaff
et al. (2001) were the first to introduce the SMM model, also known as the high or
infinite dimensional model. Later, Kerkhof and Pelsser (2002) showed that the SMM
and the LMM are observationally equivalent. The mathematical illustration of the SMM
is as follows. Consider a finite number of forward rates (LIBOR forward rates) with an
increasing maturity structure. We define a set of dates or tenor structure

^T0 ,T1 ,T2 , ˜ ˜ ˜ ˜ ˜ ,TM `, (11)

where T0 represents the current time and T1 , ˜ ˜ ˜ ˜ ˜ ,TM 1 forward tenor dates, and the

corresponding M forward rates are the F(t,T k ,Tk 1 ), where k 1,.....,M . Moreover,

the day count fractions can be defined K Tk 1  Tk , and are computed by the
maturity of LIBOR rate i.e., say three or six calendar months. On day t, the FR can be
represented by Fk (t) { F(t,Tk ,Tk 1 ) extracted as

1 ª B(t,Tk ) º
Fk (t) { «  1» , k Tk  1  Tk , (12)
k ¬ B(t,Tk 1 ) ¼

where B(t,T) denotes the time t price of the discount bond maturing at time T. Thus,
the dynamics of M forward rates in SMM can be specified as follows

dFk (t)
 kP (t)dt   k dW kP (t), k 1,........ ,M, (13)
Fk (t)

^P
where Wk (t) i`k
1
are correlated Wiener processes under probability measure P with

>
d WkP ,Wl P (t)@  kl dt, k,l 1, ˜ ˜ ˜ ˜,M . (14)

However, we are interested in finding the dynamics of Fk (t) under a measure Q k 1

equivalent to measure P associated with numeraire B( x ,t k 1 ) . To exclude arbitrage

 157

k 1
opportunities in the above FR dynamics, we must have a drift  kQ (t) equal to zero,

i.e.,
k 1
dFk (t)  k Ft (t)dWkQ (t), k 1,........,M, (15)

The change of measure does not affect volatility functions ^ k `k

M
1
and corresponding

correlations ^ kl `k,l 1 . Therefore, the volatility functions and correlations determine the
M

covariance matrix of the FRs. There are three widely used approaches for determining
the volatility functions and the covariance/correlation matrix of the FRs. First, they can
be obtained from historical data. Second, they can be obtained by calibrating the model
to the market prices of caps and swaptions. Third, a trader can explicitly specify
volatilities and correlations based on how he believes they will evolve. Here, the second
approach is selected by calibrating the SMM directly to the whole swaptions matrix.

For the SMM calibration, we select a separated approach (a calibration technique for
multifactor interest rate models) used widely in practice in the financial industry.
Separated calibration could be implemented in a straightforward way. For instance, the
SMM model could be calibrated to the whole swaption matrix. In fact, calibrating the
model to a swaption matrix is preferred over the caps because it is essentially important
in a situation where the prices of derivatives instruments are dependent on
covariance/correlation structures in addition to volatilities i.e., the values of exotic
instruments such as the Bermudan type of swaption, which are more dependent on
swaptions prices than on caps.117 We use a calibration approach proposed by Gatarek et
al. (2007). The technical details of separated calibration are provided in Appendix A.
The calibration of the SMM model to each of the EUR, USD and GBP swaption IVs are
provided in the following subsections.

5.2 Calibration of the SMM to the EUR swaption IVs

Because the EUR swaption market is relatively new, few studies have been devoted to
this market. This study attempts to fill the gap, calibrating the SMM model to the EUR
swaption IVs and reproducing the whole swaption matrix. Furthermore, the calibration
results obtained are compared with those of the EUR swaption IVs and GBP IVs. Thus,
a separated calibration technique is used in conjunction with the multivariable
optimization algorithm, where the target function is the calibration procedure that

117 See Gatarek et al. (2007).

 158

minimizes the difference between the theoretical and market swaption volatilities.118
We have initial parameter values, and the time to maturity of swaptions is expressed in
years, i.e., [0.25, 0.5, 1, 2, 3, 4, 5, 6, 7, 8].119 Gatarek et al. (2007) argued that the
variance-covariance matrix must be positive definite. If this restriction is not to be
violated, then the algorithm must first remove eigenvectors associated with negative
eigenvalues.

Figure 3 provides results for the SMM calibration to the EUR market by using a
nonlinear multivariable optimization technique (i.e., fminsearch solver in Matlab). The
subplot at the top-right of Figure 3 shows the number of iterations on the x-axis and the
number of function evaluations on the y-axis. This plot requires 1446 iterations and
2000 function evaluations in order to reach the optimal solution. The target for the
optimization routine is the root mean squared error (RMSE) (i.e., the graph shows its
value ‘current function value’) between theoretical volatilities and market volatilities,
and we obtained RMSE=0.072506.

The subplot at the top-left shows the number of variables on the x-axis and the values
of optimized parameters on the y-axis. The optimized parameter values obtained for
our ten variables are [0.961, 1.681, 2.284, 2.084, 2.437, 1.977, 1.597, 1.21, 0.627, 0.399].
The subplot to the bottom-left shows a graphical representation for function values
during 1446 iterations against 2000 function evaluations. The y-axis function produces
a very high value (around 1800). By 1446 iterations, the RMSE level is reduced to
0.072506. When this value is smaller, we will have better results.

118 Mathematical illustration of the calibration procedure is provided in the Appendix A.

119 The approach is similar to the study of Longstaff et al. (2001).
 159

Figure 3. Nonlinear multivariable optimization results for the EUR swaption market.

The corresponding theoretical volatilities (model volatilities), market volatilities and

their differences for the swaptions matrix on April 4, 2006 are reported in Table 4.120
Panel A of Table 4 provides final theoretical volatilities obtained by calibrating the
SMM to the EUR market data. Many intermediate steps are involved, as illustrated
mathematically in separated calibration procedure.121 Panel C of Table 4 reports
volatility differences between theoretical and market volatilities (i.e., model volatilities
minus market quoted volatilities). Here, the eigenvalues generate very small differences
between the theoretical and market swaption volatilities. The largest differences are
seen for swaptions with very short maturities, such as three, six months, and one year
maturities on one year underlying swaps. There are no significant differences for
maturities longer than one year in the whole swaption matrix. The whole swaption
matrix is reproduced more or less identically to the market swaption matrix. This
implies that SMM calibration using nonlinear multivariable optimization can be used in
practice for valuation of various types of interest rate and credit derivatives.

120Figure 4 graphically translates the theoretical swaption market and difference volatilities in Table 4.
121The variance-covariance and modified variance-covariance matrices are reported in Table 5 of Appendix
B. Moreover, we had two negative eigenvalues in all ten eigenvalues during the optimization process for
EUR swaptions.
 160

Table 4. The EUR swaptions theoretical, market swaption volatilities and their
Panel A : Theoretical volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y

3M 14.2168 14.7102 15.8130 16.1022 16.0104 15.6030 15.1019 14.7036 14.3034 13.9033
6M 15.3588 15.9081 16.0293 16.1319 16.0106 15.6063 15.1081 14.7069 14.4061 14.0000

1Y 16.5356 16.3159 16.2220 16.1055 15.8030 15.3046 14.9040 14.6035 14.3000 14.0000
2Y 16.5100 16.3046 16.0003 15.6002 15.2010 14.8010 14.5010 14.2000 13.9000 13.7000

3Y 16.3177 16.0016 15.5007 15.1018 14.7016 14.4014 14.1000 13.8000 13.6000 13.4000
4Y 16.1020 15.5007 15.1001 14.7001 14.4002 14.1000 13.8000 13.5000 13.3000 13.1000

5Y 15.6001 15.1008 14.7006 14.3005 14.0000 13.7000 13.5000 13.2000 13.0000 12.9000
6Y 16.0020 14.4010 14.2006 13.8000 13.5000 13.3000 13.1000 12.9000 12.7000 12.6000
7Y 14.4001 14.0001 13.7000 13.4000 13.1000 12.9000 12.7000 12.5000 12.4000 12.3000
8Y 17.1000 13.6000 13.3000 13.1000 12.9000 12.5000 12.4000 12.2000 12.1000 12.0000
Panel B: Market volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M 14.0000 14.7000 15.8000 16.1000 16.0000 15.6000 15.1000 14.7000 14.3000 13.9000
6M 15.3000 15.8000 16.0000 16.1000 16.0000 15.6000 15.1000 14.7000 14.4000 14.0000
1Y 16.4000 16.3000 16.2000 16.1000 15.8000 15.3000 14.9000 14.6000 14.3000 14.0000
2Y 16.5000 16.3000 16.0000 15.6000 15.2000 14.8000 14.5000 14.2000 13.9000 13.7000
3Y 16.3000 16.0000 15.5000 15.1000 14.7000 14.4000 14.1000 13.8000 13.6000 13.4000

4Y 16.1000 15.5000 15.1000 14.7000 14.4000 14.1000 13.8000 13.5000 13.3000 13.1000
5Y 15.6000 15.1000 14.7000 14.3000 14.0000 13.7000 13.5000 13.2000 13.0000 12.9000
6Y 16.0000 14.4000 14.2000 13.8000 13.5000 13.3000 13.1000 12.9000 12.7000 12.6000
7Y 14.4000 14.0000 13.7000 13.4000 13.1000 12.9000 12.7000 12.5000 12.4000 12.3000
8Y 17.1000 13.6000 13.3000 13.1000 12.9000 12.5000 12.4000 12.2000 12.1000 12.0000

Panel C: Volatility difference between Theoretical and Market

1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M -0.2168 -0.0102 -0.0130 -0.0022 -0.0104 -0.0030 -0.0019 -0.0036 -0.0034 -0.0033

6M -0.0588 -0.1081 -0.0293 -0.0319 -0.0106 -0.0063 -0.0081 -0.0069 -0.0061 0.0000
1Y -0.1356 -0.0159 -0.0220 -0.0055 -0.0030 -0.0046 -0.0040 -0.0035 0.0000 0.0000
2Y -0.0100 -0.0046 -0.0003 -0.0002 -0.0010 -0.0010 -0.0010 0.0000 0.0000 0.0000

3Y -0.0177 -0.0016 -0.0007 -0.0018 -0.0016 -0.0014 0.0000 0.0000 0.0000 0.0000
4Y -0.0020 -0.0007 -0.0001 -0.0001 -0.0002 0.0000 0.0000 0.0000 0.0000 0.0000
5Y -0.0001 -0.0008 -0.0006 -0.0005 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

6Y -0.0020 -0.0010 -0.0006 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

7Y -0.0001 -0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
8Y -0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
 161

EUR - Theoretical Volatilities

18
Volatility

16
10
14
8
12
10 6
8
4
6
4 2
2 Swap Maturity
Swaption Expiry 0 0

EUR - Market Volatilities

18
Volatility

16
10
14
8
12
10 6
8
4
6
4 2
2 Swap Maturity
Swaption Expiry 0 0

EUR - Vol Diff between Theoretical and Market

0
Difference

-0.2 10

8
-0.4
10 6
8
4
6
4 2
2 Swap Maturity
Swaption Expiry 0 0

Figure 4. EUR theoretical, market swaption volatilities and their differences.

 162

5.3 Calibration of the SMM to the USD swaption IVs

The USD-denominated swaption market is the oldest and largest swaption market,
implying an abundance of research on this particular market relative to the EUR and
GBP swaption markets. Nonetheless, SMM is calibrated to the USD swaption matrix
(i.e., using separated calibration with multivariable nonlinear optimization technique),
so the whole swaption volatility matrix is reproduced for this market as well. For
calibration by optimization, we have initial parameters for [0.25, 0.5, 1, 2, 3, 4, 5, 6, 7,
8]. Optimization results are shown in Figure 5.

Figure 5: Nonlinear multivariable optimization results for the USD swaption market.

The subplot at the top-right of Figure 5 shows that the optimization requires 1433
iterations and 2000 function evaluations in order to produce an optimal solution for
the USD market, whereas for EUR it required 1446 iterations. The target for the
optimization routine is the RMSE level (its value is shown on graph as ‘current function
value’), which is 5.1568674. All ten optimized parameter values are reported in the
subplot at the top-left of Figure 5. The values are [1.063, 1.406, 1.768, 2.239, 2.327,
1.542, 0.971, 0.612, 0.311, 0.155]. The subplot at bottom left presents the function
values obtained according to the number of iterations and function evaluations (the
optimum function value obtained in 1433 iterations and 2000 function evaluations).
 163

The function initially produces a very high function value (around 2300) and ultimately
minimizes at a final value of 5.1568674. If the RMSE level is compared with the EUR
RMSE level of 0.072506, then the EUR solution is found to be relatively optimal in
both, because the smaller RMSE level is the final solution.

Table 6 provides the theoretical, market swaption volatilities and their differences for
the USD swaptions matrix on April 4, 2006.122 Panel A of Table 6 presents the
theoretical volatility matrix obtained using the separated calibration procedure. During
the optimization process, there was one negative eigenvalue out of ten, though very
small in absolute value. Panel C of Table 6 provides the differences between the
theoretical and market volatilities. Like the EUR, here as well the eigenvalues generate
very small differences between the theoretical and market volatilities. The largest
differences are seen for the swaptions with the shorter maturities (three months, six
months, one year, and two years, with underlying one and two year swaps contracts).
However, there are no significant differences for maturities longer than two years. The
whole swaption matrix is reproduced with our calibration and is roughly similar to that
of the market quoted matrix. Compared to the EUR theoretical volatilities, we assert
that the SMM produced an optimal solution for the EUR swaption market rather than
for the USD swaption market. The differences between the theoretical and market
swaption volatilities are noticed more in shorter swaption maturities in the USD
theoretical volatility matrix, whereas this pattern is noticeable for maturities up to one
year in the EUR matrix. Beyond one year, the differences between theoretical and
market swaption volatilities are negligible.

122 Figure 6 graphically translates the theoretical, market swaption volatilities and their differences in Table
6.
 164

Table 6. The USD swaptions theoretical, market swaption volatilities and their differences.
Panel A : Theoretical volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y

3M 13.0115 13.5062 13.9192 14.1714 14.2297 14.1941 14.0561 13.9584 13.8512 13.7472
6M 14.0921 14.0248 14.4009 14.5008 14.7275 14.6138 14.4196 14.3186 14.1186 14.0000

1Y 15.1626 15.6954 15.6701 15.7485 15.6778 15.5728 15.3602 15.2527 15.0000 14.9000
2Y 16.7696 16.6266 16.6045 16.4484 16.3471 16.2386 16.0342 15.8000 15.7000 15.5000

3Y 16.9003 17.0561 16.8212 16.7237 16.5199 16.4180 16.2000 16.1000 15.9000 15.8000
4Y 17.1320 16.8283 16.6245 16.5179 16.3149 16.2000 16.0000 15.9000 15.7000 15.6000

5Y 16.8006 16.6012 16.4014 16.3017 16.1000 15.9000 15.8000 15.6000 15.5000 15.4000
6Y 16.4049 16.2027 16.0023 15.8000 15.6000 15.5000 15.3000 15.2000 15.1000 14.9000
7Y 16.1005 15.9005 15.6000 15.4000 15.2000 15.1000 14.9000 14.8000 14.7000 14.6000
8Y 15.7003 15.4000 15.2000 15.0000 14.8000 14.7000 14.5000 14.4000 14.3000 14.2000
Panel B: Market volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M 11.5000 13.5000 13.9000 14.1000 14.2000 14.1000 14.0000 13.9000 13.8000 13.7000
6M 12.5000 13.9000 14.4000 14.5000 14.7000 14.6000 14.4000 14.3000 14.1000 14.0000
1Y 14.9000 15.4000 15.6000 15.6000 15.6000 15.5000 15.3000 15.2000 15.0000 14.9000
2Y 16.5000 16.6000 16.5000 16.4000 16.3000 16.2000 16.0000 15.8000 15.7000 15.5000
3Y 16.9000 17.0000 16.8000 16.7000 16.5000 16.4000 16.2000 16.1000 15.9000 15.8000

4Y 17.0000 16.8000 16.6000 16.5000 16.3000 16.2000 16.0000 15.9000 15.7000 15.6000
5Y 16.8000 16.6000 16.4000 16.3000 16.1000 15.9000 15.8000 15.6000 15.5000 15.4000
6Y 16.4000 16.2000 16.0000 15.8000 15.6000 15.5000 15.3000 15.2000 15.1000 14.9000
7Y 16.1000 15.9000 15.6000 15.4000 15.2000 15.1000 14.9000 14.8000 14.7000 14.6000
8Y 15.7000 15.4000 15.2000 15.0000 14.8000 14.7000 14.5000 14.4000 14.3000 14.2000
Panel C: Volatility difference between Theoretical and Market
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M -1.5115 -0.0062 -0.0192 -0.0714 -0.0297 -0.0941 -0.0561 -0.0584 -0.0512 -0.0472
6M -1.5921 -0.1248 -0.0009 -0.0008 -0.0275 -0.0138 -0.0196 -0.0186 -0.0186 0.0000
1Y -0.2626 -0.2954 -0.0701 -0.1485 -0.0778 -0.0728 -0.0602 -0.0527 0.0000 0.0000
2Y -0.2696 -0.0266 -0.1045 -0.0484 -0.0471 -0.0386 -0.0342 0.0000 0.0000 0.0000

3Y -0.0003 -0.0561 -0.0212 -0.0237 -0.0199 -0.0180 0.0000 0.0000 0.0000 0.0000
4Y -0.1320 -0.0283 -0.0245 -0.0179 -0.0149 0.0000 0.0000 0.0000 0.0000 0.0000

5Y -0.0006 -0.0012 -0.0014 -0.0017 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
6Y -0.0049 -0.0027 -0.0023 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

7Y -0.0005 -0.0005 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
8Y -0.0003 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
 165

USD - Theoretical Volatilities

18

16
Volatility

14

12
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

USD - Market Volatilities

18

16
Volatility

14

12

10
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

USD - Vol Diff between Theoretical and Market

-0.5
Difference

-1

-1.5

-2
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

Figure 6. USD theoretical, market swaption volatilities and their differences

 166

5.4 Calibration of the SMM to the GBP swaption IVs

The GBP-denominated swaption market is the smallest of the three, though it is older
than the EUR swaption market. More studies have been devoted to this market than to
the relatively new EUR swaption market. Our objective is thus to reproduce swaption
volatility matrix for the GBP swaption market. The calibration steps for the SMM are
similar. First of all, initial parameters are provided to the optimization routine, i.e., the
time to maturity of each swaption contract [0.25, 0.5, 1, 2, 3, 4, 5, 6, 7, 8]. The
optimization results are presented in Figure 7. The subplot at the top-right shows that
optimization requires 1429 iterations and 2000 function evaluations in order to
provide the optimal solution (this is more than the number of iterations required of the
EUR and USD markets). An RMSE level of 0.306703 is obtained for the GBP market, a
smaller level than the USD RMSE level though larger than the EUR RMSE level. The
corresponding optimized values are [1.501, 1.811, 1.97, 2.422, 2.55, 2.021, 1.56, 1.063,
0.546, 0.208]. The different function values are presented in the subplot at the bottom-
left and the optimum solution is found during 1429 iterations and 2000 function
evaluations. The function value reaches around 3300, translating into an optimal value
of 0.306703 after 1429 iterations.

Figure 7. Nonlinear multivariable optimization results for the GBP swaption market.
 167

Table 8 reports the theoretical, market swaption volatilities and their differences for the
GBP swaption matrix on April 4, 2006.123 Panel A of Table 8 shows the theoretical
swaption volatilities obtained by calibrating SMM to the GBP swaption market. As with
the USD, one negative eigenvalue in ten eigenvalues is found to be negative, but very
small in absolute value. The corresponding differences between theoretical and market
volatilities are reported in Panel C of Table 8. These eigenvalues generate very small
differences between the theoretical and market swaption volatilities. The deviations are
observed for the swaptions of shorter maturities (three months, six months and one
year, with underlying contracts of one year lengths). However, there are no significant
differences for maturities longer than one year. The volatility patterns are similar to the
EUR theoretical volatilities, but different from the patterns of the USD theoretical
volatilities.

123 Figure 8 shows the theoretical, market swaption volatilities and their differences.
 168

Table 8. The GBP theoretical, market swaption volatilities and their differences.
Panel A: Theoretical volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y

3M 10.0083 10.4012 10.7066 11.0011 11.3000 11.5003 11.7006 12.0000 12.2001 12.4001
6M 10.9684 11.2259 11.3193 11.5106 11.7034 11.9011 12.0018 12.2010 12.4008 12.6000

1Y 12.4549 12.4056 12.3002 12.2007 12.3010 12.4001 12.5002 12.7001 12.8000 12.9000
2Y 13.5052 13.5024 13.3002 13.2000 13.1002 13.1000 13.1000 13.1000 13.1000 13.1000

3Y 14.1011 14.0000 13.8002 13.5000 13.4000 13.3000 13.3000 13.3000 13.3000 13.3000
4Y 14.4009 14.3008 14.0000 13.8001 13.6000 13.5000 13.5000 13.5000 13.5000 13.5000

5Y 14.5005 14.3000 14.1000 13.9000 13.8000 13.7000 13.6000 13.6000 13.6000 13.6000
6Y 14.6007 14.4001 14.2001 14.0000 13.8000 13.7000 13.7000 13.7000 13.7000 13.7000
7Y 14.7000 14.5000 14.3000 14.0000 13.8000 13.7000 13.8000 13.8000 13.8000 13.8000
8Y 14.7000 14.6000 14.3000 14.0000 13.9000 13.8000 13.8000 13.8000 13.8000 13.8000
Panel B: Market volatilities
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M 9.6000 10.4000 10.7000 11.0000 11.3000 11.5000 11.7000 12.0000 12.2000 12.4000
6M 10.6000 11.2000 11.3000 11.5000 11.7000 11.9000 12.0000 12.2000 12.4000 12.6000
1Y 12.4000 12.4000 12.3000 12.2000 12.3000 12.4000 12.5000 12.7000 12.8000 12.9000
2Y 13.5000 13.5000 13.3000 13.2000 13.1000 13.1000 13.1000 13.1000 13.1000 13.1000
3Y 14.1000 14.0000 13.8000 13.5000 13.4000 13.3000 13.3000 13.3000 13.3000 13.3000

4Y 14.4000 14.3000 14.0000 13.8000 13.6000 13.5000 13.5000 13.5000 13.5000 13.5000
5Y 14.5000 14.3000 14.1000 13.9000 13.8000 13.7000 13.6000 13.6000 13.6000 13.6000
6Y 14.6000 14.4000 14.2000 14.0000 13.8000 13.7000 13.7000 13.7000 13.7000 13.7000
7Y 14.7000 14.5000 14.3000 14.0000 13.8000 13.7000 13.8000 13.8000 13.8000 13.8000
8Y 14.7000 14.6000 14.3000 14.0000 13.9000 13.8000 13.8000 13.8000 13.8000 13.8000
Panel C: Volatility difference between Theoretical and Market
1Y 2Y 3Y 4Y 5Y 6Y 7Y 8Y 9Y 10Y
3M -0.4083 -0.0012 -0.0066 -0.0011 -0.0000 -0.0003 -0.0006 -0.0000 -0.0001 -0.0001

6M -0.3684 -0.0259 -0.0193 -0.0106 -0.0034 -0.0011 -0.0018 -0.0010 -0.0008 0.0000
1Y -0.0459 -0.0056 -0.0002 -0.0007 -0.0010 -0.0001 -0.0002 -0.0001 0.0000 0.0000
2Y -0.0052 -0.0024 -0.0002 -0.0000 -0.0000 -0.0000 -0.0000 0.0000 0.0000 0.0000
3Y -0.0011 -0.0000 -0.0002 -0.0000 -0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000

4Y -0.0009 -0.0008 -0.0000 -0.0001 -0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

5Y -0.0005 -0.0000 -0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
6Y -0.0007 -0.0001 -0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

7Y -0.0000 -0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
8Y -0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
 169

GBP - Theoretical Volatilities

16

14
Volatility

12

10
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

GBP - Market Volatilities

16

14
Volatility

12

10

8
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

GBP - Diff between Theoretical and Market

-0.2
Difference

-0.4

-0.6

-0.8
10
10
8
8
6 6

4 4
2 2
Swaption Expiry 0 0 Swap Maturity

Figure 8. GBP theoretical, market swaption volatilities and their differences.

 170

6 Conclusion

We estimate and extract the important implied risk factors affecting IV movements in
each of the EUR, USD and GBP swaption markets. We then provide intuitive
interpretations of the observed, implied risk factors. Moreover, we examine the
dynamic interactions between the risk factors across markets using techniques such as
the Granger causality test and the impulse response function. In the second part of the
study, we calibrate the string market model to swaption IVs by using multivariable
nonlinear optimization to reproduce the swaption matrices for the EUR, USD, and GBP
swaption markets.

First, we applied PCA to each of the EUR, USD, and GBP swaption IVs to discover the
important risk factors. We found that three risk factors explain about 94 -97% of the
variance in each of the EUR, USD, GBP swaption IVs. Second, the significant implied
factors present high correlations across swaption markets; consequently, there are
strong linkages across the three markets. Bi-directional causality is at work between the
implied factors from each of the EUR and USD swaption markets. The factors from
EUR and USD swaption markets Granger-cause the factors from the GBP swaption
market, but not vice-versa. Furthermore, in innovation accounting investigations,
shocks to both factors implied by the EUR and USD IVs are found to be influential for
the factor implied from the GBP IVs. However, a shock to the GBP factors does not
affect the factors noticed in the other two markets. Finally, there are many similar
characteristics between EUR and GBP markets, in contrast to the USD swaption
market. The whole swaption matrix for the EUR, USD and GBP markets is reproduced
using the SMM model. The least differences are observed between the theoretical and
market volatilities for EUR, GBP, and USD, respectively. Here too, similar
characteristics are found between EUR and GBP markets.

The identification of risk factors is important in practice. These factors can be used for
hedging of the portfolio position, particularly vega-hedging, generating smooth implied
volatility surfaces, risk management, and for model calibration purposes. First, as in
the interest rate market, we find securities with a variety of maturities; therefore, it is
almost impossible for a trader to vega-hedge portfolios against each and every
individual risk. Hence, in the easiest way to account for most of the risks, a trader uses
the major risk factors instead of hedging against every underlying risk. Second, we find
a limited number of OTC-traded swaptions in the swaption markets. When a trader
 171

needs to quote options with different maturities, he is generally restricted. This could
be solved by a swaption volatility surface and thereby we could use the risk factors in
order to generate swaption volatility surface. A trader can easily quote new options by
taking estimates from the volatility surface. Third, by employing these factors, a trader
can manage the overall downside risk for his portfolio i.e., value-at-risk for a portfolio
consisting of interest rate derivatives. Finally, market models could be easily calibrated
to only few important factors i.e., LIBOR market models.

The SMM is an important pricing and hedging tool. First, this model enriches the LMM
by calibrating to the whole swaption matrix, whereas the LMM is calibrated to only few
risk factors. Thus, the SMM calibration accounts for all independent risk factors in a
parsimonious fashion. Second, once a whole swaption matrix is reproduced, the pricing
and hedging of exotic derivatives such as Bermudan swaptions could be done by using
these plain-vanilla products. Finally, in conjunction with SMM, multivariable nonlinear
optimization could be used for increased accuracy and efficiency in the pricing and
hedging activities of exotic interest rate derivatives.
 172

References

Badshah, I., 2008. Modelling the dynamics of implied volatility surfaces. European
Financial Management Association (EFMA)-2008, Athens, Greece, Conference
Proceedings.
Black, F., 1976. The pricing of commodity contracts. Journal of Financial Economics 3,
167-179.
Brace, A., Gatarek,D., and Musiela, M., 1997. The market model of interest rate
dynamics. Mathematical Finance 7, 127-155.
Brooks, C., 2002. Introductory econometrics for finance. First edition, Cambridge
University Press, Cambridge, United Kingdom.
Bruggemann,R., Härdle,W.,Mungo,J.,Trenkler,C., 2008. VAR modeling for dynamic
loadings driving volatility strings. Journal of Financial Econometrics 6, 361-381.
Cont, R., and Da Fonseca, J., 2002. Dynamics of implied volatility surfaces.
Quantitative Finance 2, 45-60.
Fengler, M., Härdle, W., and Villa, C., 2003. The dynamics of implied volatility: A
common principal components analysis approach. Review of derivative research
6, 179-202.
Gatarek, D., Bachert, P., and Maksymiuk, R., 2007. The Libor market model in practice.
First edition, Wiley Finance, Sussex, England.
Goldstein, R., 2000. The term structure of interest rates as a random field. Review of
Financial Studies 13, 365-384.
Granger, J., 1969. Investigating causal relationships by econometric models and cross-
spectral Analysis. Econometrica 37, 424-438.
Huang, Y., Neftci, S.N., and Guo, F., 2008. Swap curve dynamics across markets: Case
of US dollar versus HK dollar. Journal of International Financial Markets,
Institutions & Money 18, 79-93.
Heath, D., Jarrow, R., and Morton, A., 1992. Bond pricing and term structure of
interest rates: a new methodology. Econometrica 60, 77-105.
Knez, P., Litterman, R., and Scheinkman, J., 1994. Explorations into factors explaining
money market returns. Journal of Finance 5, 1861-1882.
Jamshidian, F., 1997. Libor and Swap market models and measures. Finance and
Stochastics 1, 293-330.
Kennedy, D., 1994. The term structure of interest rates as Gaussian random field.
Mathematical Finance 4, 247-258.
Kennedy, D., 1997. Characterizing Gaussian models of the term structure of interest
rates. Mathematical Finance 7, 107-118.
Kerkhof, J., and Pelsser, A., 2002. Observational equivalence of discrete string models
and market models. Journal of Derivatives 10, 55-61.
Koop, G., Pesaran, M., and Potter, S., 1996. Impulse response analysis in nonlinear
multivariatemodels. Journal of Econometrics 74 , 119-147.
Litterman, R., and Scheinkman, J.,1991. Common factors affecting bond returns.
Journal of Fixed Income 1, 54-61.
London, J., 2004. Modeling derivatives in C++. First edition, Wiley Finance, New
Jersey, United States.
Longstaff, F., Santa-Clara, P., Schwartz, E., 2001. The relative valuation of caps and
swaptions: Theory and empirical evidence. Journal of Finance 56, 2067-2109.
Longstaff, F., and Schwartz, E., 2001. Valuing American options by simulation: A
simple least-squares approach. Review of Financial Studies 14, 113-147.
Mills, T., 1999. The Econometric Modelling of Financial Time Series. Second edition,
Cambridge University Press, Cambridge, UK.
Nelder, J., and Mead, R, 1965. A simplex method for function minimization. Computer
Journal 7, 308-313.
 173

Persaran, M., and Shin, Y., 1998. Impulse response analysis in linear multivariate
models. Economics Letters 58, 17-29.
Rebonato, R., 2002. Modern pricing of the interest rate derivatives: The LIBOR market
model and beyond. Princeton University Press, New Jersey, United States.
Santa-Clara, P., and Sornette, D., 2001. The dynamics of the forward interest rate curve
with stochastic string shocks. Review of Financial Studies 14, 149-185.
Skiadopoulos, G., Hodges, S., and Clewlow, L, 1999. The Dynamics of the S&P 500
Implied Volatility Surface. Review of Derivative Research 3, 263-282.
Sims, C.A., 1980. Macroeconomics and reality. Econometrica, 48, 1-48.
Wadhwa, P., 1999. An empirical analysis of the common factors governing U.S. Dollar
LIBOR implied volatility movements. Journal of Fixed Income 9, 61-68.
 174

Appendix A: Separated Calibration Procedure

We use the separated calibration procedure proposed by Gatarek et al. (2007).124 The
calibration steps are as follows. First, we form a matrix of market quoted swaption IVs,
i
 mM , where m represents swaption maturity and M the underlying swap maturity. For
i
example,  1,2 is the swaption IV for a swaption maturing at T1 with the underlying
swap period from (T1 ,T2 ). Second, we consider other market data such as LIBOR rates,
FRA and IRS and create a vector of discount factors B as

B ^B0,T1 , B0,T2 , B0,T3  , ˜ ˜ ˜ , B0,TM ` ,

i
Then the covariance matrix of FRs, C mm , is estimated through the following iterative
process

Ti

C kli ³
inst
(t,Tl 1 ,Tl ) inst (t,Tk 1 ,Tk )dt ,
0

Where i  k and i  l and  inst (t,T l  1 ,T l ) is the stochastic instantaneous volatility of

the FRs Fl (t,Tl 1 ,Tl ) , while assuming that

C kli i C kl , t0.

Then the initial values  i are supplied, given that  i values are positive and particularly
important for obtaining on the diagonal. Hence, the diagonal parameters are

C kk 0,k  i (t,Tk ,Tk 1 ) 2 k .

To determine the non-diagonal elements of the matrix C , the parameters Fi,jk (t) are
required to be estimated as

Fi,jk (0) B(0,Tk 1 )  B(0,Tk ) B(0,Ti )  B(0,T j )

Using the above equation, the complete matrix of parameters F is estimated

iteratively. This is a three-dimension or three-array matrix: the first array elements are
indexed with i, called the outer loop; the second array elements are indexed with j,
called the second loop; and the final array is indexed with k, called the inner most loop.
Consequently, the off-diagonal parameters of matrix C are estimated using the
estimated parameters of matrix F . With the help of the following iterative process, the
off-diagonal parameters are

124 Using similar notations as Gatarek et al. (2007).

 175

§ n n
·
C k,n1 k  kn2 ¨ ¦ ¦F i
k,n (0)M i 1l 1 FkNl (0)  2Fknk 1(0)Ck,n1 Fk,nn (0)¸ 2k Fknk 1(0)Fknn (0) .
© l k 1 i k 1 ¹

The whole matrix is computed using this iterative process, starting as k 1,..........,m
and n k  2,......... . Finally, we obtain the matrix C . If negative eigenvalues are found,
then we remove the associated eigenvectors and compute a new PCA-approximated
PCA
matrix C . Finally, the theoretical volatility matrix is obtained by modifying the
initial covariance matrix C as
i
C klPCA i C klPCA
§ n n
·
2
and k, kn k ¨ ¦ ¦F i
k,n 1,l 1 FkN (0) ¸
(0)CiPCA l

© l k 1 i k 1 ¹

The SMM model uses k k . The theoretical volatilities are then

§ n n
·
 kn2 ¨¦ ¦F i
k,n 1,l 1 FkN (0)¸
(0)CiPCA l

© l k 1 i k 1 ¹

RMSE is computed between market and theoretical swaptions volatilities

¦  
m
RMSE Theoritica l
ij   ijMarketl
i,j 1

Afterward, the unconstrained nonlinear optimization method is used to obtain an

optimal solution. For this purpose, we use the Matlab optimization solver called
“fminsearch”, which minimizes unconstrained multidimensional objective function
illustrated above.

The fminsearch optimization solver uses the Nelder-Mead (1965) “simplex” algorithm,
which finds the minimum of a scalar-valued nonlinear function of many variables given
initial values. Suppose we have x fminsearch  fun,x0  , which starts at the initial
point x 0 and finds a local minimum x of the function described in fun. The x 0 can be a
scalar, a vector or a matrix.
 176

Appendix B: Variance-Covariance and Modified-Variance-Covariance Matrices

Table 5. EUR swaption market

VarianceCovariance Matrix
0.0513 0.0517 0.0643 0.0502 0.0128 -0.0509 -0.0807 -0.0605 -0.1020 -0.1742
0.0517 0.0703 0.0644 0.0474 0.0196 -0.0266 -0.0874 -0.1079 -0.1354 -0.1625
0.0643 0.0644 0.0994 0.0806 0.0322 -0.0131 -0.0573 -0.1398 -0.1568 -0.1919
0.0502 0.0474 0.0806 0.1312 0.0916 0.0156 -0.0357 -0.0975 -0.1423 -0.1831
0.0128 0.0196 0.0322 0.0916 0.1631 0.1038 0.0004 -0.0862 -0.1602 -0.2249
-0.0509 -0.0266 -0.0131 0.0156 0.1038 0.2576 0.1646 0.0364 -0.0831 -0.1461
-0.0807 -0.0874 -0.0573 -0.0357 0.0004 0.1646 0.3687 0.2034 0.1031 -0.1621
-0.0605 -0.1079 -0.1398 -0.0975 -0.0862 0.0364 0.2034 0.6058 0.1201 -0.3845
-0.1020 -0.1354 -0.1568 -0.1423 -0.1602 -0.0831 0.1031 0.1201 1.1148 0.1529
-0.1742 -0.1625 -0.1919 -0.1831 -0.2249 -0.1461 -0.1621 -0.3845 0.1529 2.8122

Modified_VarianceCovariance Matrix
0.0529 0.0509 0.0627 0.0507 0.0121 -0.0506 -0.0807 -0.0610 -0.1021 -0.1743
0.0509 0.0708 0.0654 0.0471 0.0200 -0.0268 -0.0874 -0.1076 -0.1353 -0.1624
0.0627 0.0654 0.1011 0.0801 0.0330 -0.0134 -0.0574 -0.1393 -0.1566 -0.1918
0.0507 0.0471 0.0801 0.1313 0.0913 0.0157 -0.0357 -0.0976 -0.1424 -0.1831
0.0121 0.0200 0.0330 0.0913 0.1635 0.1036 0.0004 -0.0860 -0.1601 -0.2249
-0.0506 -0.0268 -0.0134 0.0157 0.1036 0.2576 0.1646 0.0363 -0.0831 -0.1461
-0.0807 -0.0874 -0.0574 -0.0357 0.0004 0.1646 0.3687 0.2034 0.1031 -0.1621
-0.0610 -0.1076 -0.1393 -0.0976 -0.0860 0.0363 0.2034 0.6059 0.1201 -0.3844
-0.1021 -0.1353 -0.1566 -0.1424 -0.1601 -0.0831 0.1031 0.1201 1.1148 0.1529
-0.1743 -0.1624 -0.1918 -0.1831 -0.2249 -0.1461 -0.1621 -0.3844 0.1529 2.8122
 177

Table 7. USD swaption market

VarianceCovariance Matrix
0.0626 0.0993 0.0736 0.0485 0.0297 -0.0637 -0.0728 -0.0641 -0.1280 -0.0762
0.0993 0.0840 0.1034 0.0897 0.0425 0.0350 -0.0548 -0.1255 -0.0757 -0.2384
0.0736 0.1034 0.1254 0.1149 0.0503 -0.0272 -0.0874 -0.1775 -0.3091 -0.2477
0.0485 0.0897 0.1149 0.1817 0.1258 -0.0155 -0.1056 -0.2271 -0.2627 -0.4904
0.0297 0.0425 0.0503 0.1258 0.2441 0.1434 -0.0703 -0.2387 -0.4642 -0.5499
-0.0637 0.0350 -0.0272 -0.0155 0.1434 0.4632 0.2493 -0.1464 -0.4661 -0.8917
-0.0728 -0.0548 -0.0874 -0.1056 -0.0703 0.2493 0.8581 0.4954 -0.1692 -0.6951
-0.0641 -0.1255 -0.1775 - 0.2271 -0.2387 -0.1464 0.4954 1.5091 0.5427 -1.2504
-0.1280 -0.0757 -0.3091 -0.2627 -0.4642 -0.4661 -0.1692 0.5427 3.3371 1.2359
-0.0762 -0.2384 -0.2477 -0.4904 -0.5499 -0.8917 -0.6951 -1.2504 1.2359 7.2584

Modified_VarianceCovariance Matrix
0.0801 0.0793 0.0825 0.0588 0.0293 -0.0525 -0.0738 -0.0601 -0.1261 -0.0741
0.0793 0.1067 0.0934 0.0781 0.0429 0.0222 -0.0536 -0.1301 -0.0778 -0.2408
0.0825 0.0934 0.1299 0.1200 0.0501 -0.0215 -0.0879 -0.1755 -0.3082 -0.2467
0.0588 0.0781 0.1200 0.1877 0.1256 -0.0089 -0.1062 -0.2247 -0.2617 -0.4892
0.0293 0.0429 0.0501 0.1256 0.2441 0.1432 -0.0703 -0.2388 -0.4642 -0.5500
-0.0525 0.0222 -0.0215 -0.0089 0.1432 0.4704 0.2486 -0.1439 -0.4649 -0.8903
-0.0738 -0.0536 -0.0879 -0.1062 -0.0703 0.2486 0.8582 0.4952 -0.1693 -0.6952
-0.0601 -0.1301 -0.1755 -0.2247 -0.2388 -0.1439 0.4952 1.5100 0.5432 -1.2500
-0.1261 -0.0778 -0.3082 -0.2617 -0.4642 -0.4649 -0.1693 0.5432 3.3373 1.2361
-0.0741 -0.2408 -0.2467 -0.4892 -0.5500 -0.8903 -0.6952 -1.2500 1.2361 7.2586

Table 9. GBP swaption market

VarianceCovariance Matrix
0.0308 0.0335 0.0189 0.0165 -0.0021 -0.0303 -0.0573 -0.0252 -0.0937 -0.1220
0.0335 0.0467 0.0419 0.0174 -0.0051 -0.0182 -0.0514 -0.0993 -0.0913 -0.1627
0.0189 0.0419 0.0781 0.0605 0.0029 -0.0573 -0.0848 -0.0994 -0.1461 -0.0977
0.0165 0.0174 0.0605 0.1129 0.0700 -0.0264 -0.0865 -0.1014 -0.1537 -0.1732
-0.0021 -0.0051 0.0029 0.0700 0.1562 0.1031 -0.0069 -0.1019 -0.1561 -0.2080
-0.0303 -0.0182 -0.0573 -0.0264 0.1031 0.2569 0.1772 0.0071 -0.1248 -0.2490
-0.0573 -0.0514 -0.0848 -0.0865 -0.0069 0.1772 0.4046 0.2393 -0.0607 -0.3296
-0.0252 -0.0993 -0.0994 -0.1014 -0.1019 0.0071 0.2393 0.7023 0.2328 -0.6249
-0.0937 -0.0913 -0.1461 -0.1537 -0.1561 -0.1248 -0.0607 0.2328 1.5846 0.0080
-0.1220 -0.1627 -0.0977 -0.1732 -0.2080 -0.2490 -0.3296 -0.6249 0.0080 4.6734

Modified_VarianceCovariance Matrix
0.0335 0.0305 0.0203 0.0160 -0.0024 -0.0300 -0.0571 -0.0256 -0.0935 -0.1221
0.0305 0.0500 0.0404 0.0179 -0.0049 -0.0185 -0.0517 -0.0988 -0.0915 -0.1626
0.0203 0.0404 0.0788 0.0602 0.0028 -0.0571 -0.0847 -0.0997 -0.1460 -0.0978
0.0160 0.0179 0.0602 0.1130 0.0701 -0.0265 -0.0866 -0.1013 -0.1537 -0.1732
-0.0024 -0.0049 0.0028 0.0701 0.1562 0.1031 -0.0070 -0.1019 -0.1561 -0.2080
-0.0300 -0.0185 -0.0571 -0.0265 0.1031 0.2569 0.1772 0.0071 -0.1247 -0.2490
-0.0571 -0.0517 -0.0847 -0.0866 -0.0070 0.1772 0.4046 0.2393 -0.0607 -0.3296
-0.0256 -0.0988 -0.0997 -0.1013 -0.1019 0.0071 0.2393 0.7023 0.2327 -0.6249
-0.0935 -0.0915 -0.1460 -0.1537 -0.1561 -0.1247 -0.0607 0.2327 1.5846 0.0080
-0.1221 -0.1626 -0.0978 -0.1732 -0.2080 -0.2490 -0.3296 -0.6249 0.0080 4.6734
 EKONOMI OCH SAMHÄLLE
Skrifter utgivna vid Svenska handelshögskolan

ECONOMICS AND SOCIETY

Publications of the Hanken School of Economics

185. MARIA SUOKANNAS: Den anonyma seniorkonsumenten identifieras. Om

identitetsskapande processer i en marknadsföringskontext. Helsingfors 2008.

186. RIIKKA SARALA: The Impact of Cultural Factors on Post-Acquisition Integration.

Domestic and Foreign Acquisitions of Finnish Companies in 1993-2004. Helsinki
2008.

187. INGMAR BJÖRKMAN et al. (Eds.): Innovation, Leadership, and Entrepreneurship.

A Festschrift in Honour of Professor Martin Lindell on his 60th Birthday. Helsinki
2008.

188. JOACIM TÅG: Essays on Platforms. Business Strategies, Regulation and Policy in
Telecommunications, Media and Technology Industries. Helsinki 2008.

189. HENRIK TÖTTERMAN: From Creative Ideas to New Emerging Ventures.

Entrepreneurial Processes Among Finnish Design Entrepreneurs. Helsinki 2008.

190. ANNIKA RAVALD: Hur uppkommer värde för kunden? Helsingfors 2008.

191. TOM LAHTI: Angel Investing in Finland: An Analysis Based on Agency Theory and
the Incomplete Contracting Theory. Helsinki 2008.

192. SYED MUJAHID HUSSAIN: Intraday Dynamics of International Equity Markets.

Helsinki 2009.

193. TEEMU TALLBERG: The Gendered Social Organisation of Defence. Two Ethno-
graphic Case Studies in the Finnish Defence Forces. Helsinki 2009.

194. JONAS HOLMQVIST: Language Influence in Services. Perceived Importance of

Native Language Use in Service Encounters. Helsinki 2009.

195. ENSIO ERÄ-ESKO: Beskattningsrätt och skattskyldighet för kyrkan i Finland.

Steuerrecht und Versteuerung der Kirche in Finnland. Mit einer deutschen
Zusammenfassung. Helsingfors 2009.

196. PIA BJÖRKWALL: Nyttighetsmodeller - ett ändamålsenligt innovationsskydd?

Helsingfors 2009.

197. ARTO THURLIN: Essays on Market Microstructure. Price Discovery and Informed
Trading. Helsinki 2009.

198. PETER NYBERG: Essays on Risk and Return. Helsinki 2009.

199. YANQING JIANG: Growth and Convergence: The Case of China. Helsinki 2009.

200. HANNA WESTMAN: Corporate Governance in European Banks. Essays on Bank

Ownership. Helsinki 2009.
201. CATHARINA von KOSKULL: Use of Customer Information. An Ethnography in
Service Development. Helsinki 2009.

202. RITVA HÖYKINPURO: Service Firms’ Action upon Negative Incidents in High
Touch Services: A Narrative Study. Helsinki 2009.

203. SUVI NENONEN: Customer Asset Management in Action. Using Customer

Portfolios for Allocating Resources Across Business-to-Business Relationships for
Improved Shareholder Value. Helsinki 2009.

204. CAMILLA STEINBY: Multidimensionality of Actors in Business Networks. The

Influence of Social Action in Pharmacy Networks in Finland. Helsinki 2009.

205. JENNIE SUMELIUS: Developing and Integrating HRM Practices in MNC

Subsidiaries in China. Helsinki 2009.

206. SHERAZ AHMED: Essays on Corporate Governance and the Quality of Disclosed
Earnings – Across Transitional Europe. Helsinki 2009.

207. ANNE HOLMA: Adaptation in Triadic Business Relationship Settings. A Study in

Corporate Travel Management. Helsinki 2009.

208. MICHAL KEMPA: Monetary Policy Implementation in the Interbank Market.

Helsinki 2009.

209. SUSANNA SLOTTE-KOCK: Multiple Perspectives on Networks. Conceptual

Development, Application and Integration in an Entrepreneurial Context. Helsinki
2009.

210. ANNA TALASMÄKI: The Evolving Roles of the Human Resource Function.
Understanding Role Changes in the Context of Large-Scale Mergers. Helsinki 2009.

211. MIKAEL M. VAINIONPÄÄ: Tiering Effects in Third Party Logistics: A First-Tier

Buyer Perspective. Helsinki 2010.

212. ABDIRASHID A. ISMAIL: Somali State Failure. Players, Incentives and

Institutions. Helsinki 2010.

213. ANU HELKKULA: Service Experience in an Innovation Context. Helsinki 2010.

214. OLLE SAMUELSON: IT-innovationer i svenska bygg- och fastighetssektorn. En

studie av förekomst och utveckling av IT under ett decennium. Helsingfors 2010.

215. JOANNA BETH SINCLAIR: A Story about a Message That was a Story. Message
Form and Its Implications to Knowledge Flow. Helsinki 2010.

216. TANJA VILÉN: Being in Between. An Ethnographic Study of Opera and Dialogical
Identity Construction. Helsinki 2010.

217. ASHIM KUMAR KAR: Sustainability and Mission Drift in Microfinance. Empirical
Studies on Mutual Exclusion of Double Bottom Lines. Helsinki 2010.

218. HERTTA NIEMI: Managing in the “Golden Cage”. An Ethnographic Study of Work,
Management and Gender in Parliamentary Administration. Helsinki 2010.

219. LINDA GERKMAN: Topics in Spatial Econometrics. With Applications to House

Prices. Helsinki 2010.