See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/256021812

Volatility-Decay Risk Premia

Article  in  The Journal of Derivatives · August 2014

DOI: 10.2139/ssrn.2089435

CITATION READS
1 4,727

3 authors:

Dan Galai Haim Kedar-Levy

Hebrew University of Jerusalem Ben-Gurion University of the Negev
103 PUBLICATIONS   4,939 CITATIONS    32 PUBLICATIONS   155 CITATIONS   

SEE PROFILE SEE PROFILE

Ben Z. Schreiber
Bank Of Israel
27 PUBLICATIONS   120 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Risk Management View project

Liquidity and its cost View project

All content following this page was uploaded by Haim Kedar-Levy on 04 August 2018.

The user has requested enhancement of the downloaded file.

 Volatility-Decay Risk Premia*

Dan Galaia

Haim Kedar-Levyb

Ben Z. Schreiberc

The Journal of Derivatives, 2014

*
We are thankful to Gordon Gemmill, the discussant, and participants of the International Risk Management
Conference 2014, Warsaw, Poland, for helpful comments. Special thanks are extended to Steve Figlewski, the
Editor, for insightful suggestions and comments that greatly improved the paper.
a Abe Gray Professor (Emeritus) of Banking and Finance at the School of Business Administration, Hebrew
University, Jerusalem, Israel dan.galai@ huji.ac.il and the center for Academic Studies, Or Yehuda. Partial financial
support of the Zagagi Center is acknowledged. T-(972) 2-5617234.
b Corresponding author. The Faculty of Business and Management, Ben Gurion University of the Negev, Beer
Sheva, and Ono Academic College at Kiryat Ono, Israel. T-(972)8-6472569 hlevy@som.bgu.ac.il
c Head of Methodology and Economics Area, Information and Statistics Department, Bank of Israel, Jerusalem
Israel, and Bar Ilan University. T-(972) 2-6552595. ben.schreiber58@gmail.com
 Volatility-Decay Risk Premia

Abstract

We estimate post-jump volatility-decay risk premia as the predictable difference

between periods of high and low diffusive volatility. By constructing straddle

portfolios after positive and negative jumps occur, we show that the gains that these

hedged options' portfolios yield compensate investors for the uncertain magnitude

and duration of volatility decay, as well as for vega exposure. This paper adds to

the literature by distinguishing between the premia after positive versus negative

jumps, and by exploring premia patterns over time. In particular, we find that

GARCH(1,1) and GJR are inefficient identifiers of jumps, and show that Hampel

[1971] is a superior procedure.

JEL Classification: G10, G13; G14

Keywords: Volatility-decay premia; Post-jump risk; Options strategies; Gamma hedging.

 Introduction

The finance profession has long recognized that models based on stochastic volatility

outperform models based on a constant variance assumption (e.g., Bollerslev [1986]; Engle

[1982]; Fama [1965]). With time varying volatility, the diffusion process of equities has been

used to estimate the compensation risk-averse investors require for bearing market risk, a.k.a.

volatility risk premium (e.g., Bakshi and Kapadia [2003]; Buraschi and Jackwerth [2001]; Coval

and Shumway [2001]). Further research acknowledged that stochastic volatility models do not

capture jump risks, and indeed, joint jump and stochastic volatility models outperform the former

(e.g., Bates [1988]; Pan [2002], Eraker, Johannes, and Polson [2003], Andersen, Bollerslev, and

Diebold [2007], and most recently, Christoffersen, Jacobs, and Ornthanalai [2012]).

If the level of diffusive volatility varies in a predictable pattern, volatility risk premium

must change accordingly, or else arbitrage opportunities may arise. In particular, a large body of

literature documents a sharp increase in diffusive volatility after the realization of a jump,

followed by a gradual decay toward the long-run level (e.g., Barndorff-Nielsen, and Shephard

[2002, 2004], Bates [2000], Bollerslev, Kretschmer, Pigorsch, and Tauchen [2009], Eraker,

[2004]).1 Consequently, investors can establish profitable trading strategies based on the

expectation that the level of post-jump diffusive volatility will gradually decline. However, the

magnitude of post-jump volatility is uncertain, as well as the pace by which it decays toward the

long-run level. Facing both sources of uncertainty, investors may demand compensation for

bearing post-jump volatility-decay risk. Additionally, both the level and pace of volatility decay

may depend on the sign of the realized jump, whether positive or negative. This paper explores,

1 Most diffusion and jump models assume that jump intensity and size, as well as investors’ aversion to those risks,

are independent of the level of diffusive volatility, or the arrival of a particular jump. We adopt those assumptions

here. See a detailed discussion in Bollerslev, Gibson, and Zhou [2010].

1
by jump sign, whether the gap between post-jump and long-run volatility risk premia is explained

by the uncertain magnitude and pace of volatility decay, or whether profitable opportunities still

exist.

Our empirical approach is based on Bakshi and Kapadia [2003], whereby a dynamically

hedged portfolio of a long call and delta-hedged short position of the underlying asset earns the

riskless rate if volatility risk is either zero or not priced. However, if volatility-risk is priced, the

gain that this dynamically hedged portfolio earns reveals the sign and magnitude of volatility risk

premia. We conduct our tests following extreme returns (interchangeably jumps or outliers), as

those discerned by the Hampel [1971] methodology, and check robustness by using GARCH

(1,1). The Appendix describes and compares both methods.

While the Delta-Neutral (DN) portfolio of Bakshi and Kapadia [2003], Low and Zhang,

[2005], and others is robust for time-varying diffusive volatility, it is less robust when jump risks

are involved. Indeed, Bakshi and Kapadia implement the DN strategy also after a jump occurred,

and find zero jump-risk premia. As a result, we measure our post-jump volatility-decay premia in

two different ways: First, we write at the money call option one day after a jump occurs and buy

delta units of the underlying S&P500 index. However, because a jump has been realized, large

swings in moneyness typically follow, rendering the DN strategy an inappropriate hedge. Our

second measure, therefore, accommodates the material changes in moneyness by gamma

hedging: We construct options' portfolios by writing a long-term at-the-money straddle and

concurrently buying a short-term straddle following positive or negative jumps. This portfolio is

rebalanced daily to maintain Delta and Gamma Neutrality (hereafter DGN) until one day before

the short-term options expire. Under DGN the portfolio has a negative vega exposure, and

therefore it should gain on average from volatility decay. Volatility decay will generate gains for

the DGN strategy because the high value received for the written long-term straddle often more
2
than compensates for the change of value of the purchased short-term straddle. We test whether

this gain compensates risk-averse investors for the risks involved in the pace and duration of

volatility decay.

Implementing the DGN or DN strategies during normal volatility periods, i.e., excluding

jump days, and 10 days thereafter, reveals the long-run volatility (diffusive) risk premia net of

post-jump volatility decay risk. We conduct the tests over short- (<25 trading days), medium-

(25-50 days), and long-term (51-250 days). The 20-year sample, from 1/1991 to 12/2011, was

divided into four equal sub-periods.2 The sample captures the year 2000 internet bubble and

crash, as well as the 2008-2009 financial crisis; both had a measurable impact on the S&P500

index and its volatility. Interestingly though, we show that GARCH(1,1) fails to effectively

identify the more risky extreme returns primarily in the less turbulent periods, between 1991-

1996 and between 2002-2006.

Post-jump, volatility-decay risk premia are 2.52% on average for a single DGN straddle

portfolio (relative to the S&P500 index). This premium is primarily generated following negative

jumps (3.03%) rather than positive jumps (1.91%). It is significantly higher than the one implied

by DN strategies, as they yield 1.55% on average, 2.07% post-positive jump, and 1.12% post-

negative jump.

Analyzing the post-jump volatility-decay risk premia over time, we find that DGN yields

the highest premium in the long-term, 5.77% per implemented straddle, while in the medium-

and short-terms it yields 3.11% and 1.53%, respectively. The DN strategy yields lower premia as

2 Starting in March 2004 volatility futures were available for investors to trade on the Volatility Index VIX, and

starting in February 2006 options on VIX were launched as well. As of now, it is unclear whether the introduction of

these derivatives has had an impact on the index (Becker, Clements, and McClellan [2009]; Zhang, Shu, and Brenner

[2010]).

3
compared to DGN, possibly due to its exposure to severe changes in moneyness. Nevertheless, it

also yields increasing returns as the term of the positions lengthens.

Taking a skeptical view on the presence of arbitrage opportunities, we hypothesize that the

longer-term strategies yield increasing premia because of their extended exposure to risk factors.

We therefore regress DGN premia as calculated based on Hampel’s identification of jumps, as

well as on a GARCH(1,1) specification, following Bakshi and Kapadia [2003], Duan [1995],

Heston and Nandi [2000], Ritchken and Trevor [1999], and others. Our explanatory variables are

the squared percentage change in the underlying index, (Δ𝑆𝑃𝑋)2 ; the change in volatility, Δ𝑉𝐼𝑋,

which captures vega effects; and the duration of the straddle portfolio, 𝐷𝑇𝐸. We further control

for jump sign through a dummy variable, and for the three plausible interactions between those

factors, (Δ𝑆𝑃𝑋)2 ∗ 𝐷𝑇𝐸, (Δ𝑆𝑃𝑋)2 ∗ Δ𝑉𝐼𝑋, and 𝐷𝑇𝐸 ∗ Δ𝑉𝐼𝑋. We find that Hampel-based

volatility-decay risk premia are well explained by the three individual factors, as well as their

interactions, all of which are highly significant, except for 𝐷𝑇𝐸, which is marginally significant

(P<0.09). The most economically meaningful variables are (Δ𝑆𝑃𝑋)2 and Δ𝑉𝐼𝑋, and their

interaction. The positive jump dummy variable shows a negative and significant coefficient

(P<0.05), suggesting that volatility decay risk premia after positive jumps are smaller than those

after negative jumps. Regressing the GARCH-based premia on those factors reveals insignificant

coefficients in all, except for marginal significance for 𝐷𝑇𝐸. We therefore conclude that

GARCH(1,1) is a poor identifier of risky extreme returns.

Two key elements distinguish this study from prior art: First, we construct the strategies

only after a tail event occurs; second, our primary focus is on delta- and gamma-hedged

portfolios, rather than delta hedging only. Our rebalancing strategy differs from Bakshi and

4
Kapadia [2003] and Low and Zhang [2005]; while our approach is also model-free, we explore

the premia by holding a two-straddle portfolio.

The rest of the paper presents the hypotheses, methodology, and data in Section I, and

summary statistics in Section II. Volatility-decay risk premia are calculated and discussed in

Section III, Section IV identifies the explanatory factors for the premia, and Section V

summarizes.

I. HYPOTHESES, METHODOLOGY, AND DATA

Our primary goal is to explore whether the risks embedded in the uncertain magnitude and

duration of post-jump volatility-decay premia are priced. We start this section by defining and

measuring the volatility-decay premia, and then proceed to explore the factors that explain our

findings.

We measure the premia by implementing a dynamically hedged portfolio of straddles

immediately after the realization of an extreme return in the S&P500 index. We apply a couple of

outlier identification procedures, which are presented next. High volatility after a jump has been

realized exposes the portfolio to large variations in moneyness, and hence to volatile options'

deltas. Therefore, it is important to control for both delta as well as gamma effects. We first

compare the gain of DGN strategy with the simpler strategy of maintaining delta neutrality alone,

DN. To estimate diffusive risk premia while avoiding post-jump, high diffusive risk, we measure

the profitability of initiating DGN and DN strategies on random trading days that are distant from

jump days. This Random Trade (RT) strategy is expected to capture the stationary, long-term

5
diffusive risk premia.3 As indicated in footnote 1, consistent with similar papers our null

hypothesis is of a zero change in volatility risk premia or jump frequency in all tests.

To conduct the tests one must first discern positive and negative extreme returns from the

body of the distribution. We apply two procedures for this purpose: First, we use the Hampel

[1971] measure of the breakdown point of a distribution4, and second we discern extreme events

by the GARCH (1,1) model. We have examined a number of alternatives to the Hampel measure,

such as ad hoc cut-off rules, as well as the Huber M-Estimator and extreme VIX returns, to verify

that our identification of the particular outliers is consistent with the Hampel identification.

Indeed, we found that the majority of jumps were classified as such across the different

techniques, except for GARCH(1,1). GARCH(1,1) identified a rather different set of outliers,

therefore it is compared here alongside Hampel. To verify our GARCH(1,1) results we further

used the GJR-GARCH model (Glosten, Jagannathan, and Runkle, 1993), which is reportedly

more relevant for stock returns, but found no material differences vs. the Hampel or GARCH(1,1)

models.

Here is a brief description of Hampel's [1971] and GARCH(1,1).

1. Hampel's [1971] method: A common rule for discerning an outlier is to find an observation

outside the limits of   3 . The problem with this rule is that the mean and standard deviations

are sensitive to outliers, therefore outliers may cause severe bias. Hampel's method is a robust

version of this rule as it is based on medians rather than means. It is based on robust estimators

3 The random sampling is uniform not only across months of the year, but also in the first and second halves of each

month.

4 Hampel [1971] defines the breakdown point of a distribution as the smallest percentage of outliers that may cause

an estimator to obtain arbitrary large values. See the Appendix for further details.

6
for the location and the scale of the distribution, i.e. median and MAD (Median Absolute

Deviations) rather than the mean and standard deviation. Hampel's limits are:

Limit = Median( x )  z1− / 2 MAD ( x ) ,

where MAD = 1.4826  Median x − Median( x) , is a scale estimator and z1− / 2 is the 1-α/2

percentile of the z distribution. For a normally distributed sample, MAD is a consistent estimator

of the standard deviation, and α is the type 1 error for detecting an outlier (Olewuezi [2011]). The

moving window to calculate MAD was 225 trading days, and an outlier is detected if MADt is

outside Limit.

2. GARCH(1,1): This model assumes a conditional variance of the form

 t2 =  + t2−1 +  t2−1 ,

where t-1 is the error term from the ARCH equation 𝑟𝑡 = 𝜇𝑡 + 𝜀𝑡 . Usually,  captures persistence

in volatility and + ≤ 1. The day t rate of return is divided by the conditional standard deviation at

t to obtain a volatility adjusted rate of return (𝑣𝑎𝑟𝑜𝑟𝑡 = 𝑟𝑡 ⁄𝜎𝑡 ).Theoretically, varort should have a

standard normal distribution, however the realized distribution is often not normal. Thus, a positive

(negative) outlier is detected if varort> var or + k var or , (varort<  var or − k var or ), where

 var or ,  var or are the mean and the standard deviation of varort ~N(0,1).

Delta-Gamma-Neutral (DGN) Strategy

The DGN strategy is composed of writing a long-term straddle and concurrently buying a

short-term straddle on day t if a jump occurred on day t-1. This strategy is rebalanced daily to

maintain delta and gamma neutrality until one day before the expiration of the short-term options.

7
On that day, the long-term options are sold at prevailing closing prices.5 As Bakshi and Kapadia

[2003] note, it is advantageous to use the Black and Scholes model with constant standard

deviation, because the hedged portfolio will be immunized to all changes except changes in

measures of risk. In our case, this would capture the volatility-decay risk through vega.

If volatility-decay risk is priced, the strategy is expected to yield positive gains. We

examine post-jump risk premia over short- (<25 days), medium- (25-50 days), and long-term (51-

250 days) horizons. Additionally, we condition the tests on the sign of jumps, positive versus

negative. Among others, Bakshi and Kapadia [2003] highlight that volatility risk premia should

compensate investors for bearing the risks of negative extreme returns. However, if diffusive

volatility increases after positive jumps as well, and its level and duration are uncertain, then we

expect to find significant premia in those cases as well.

The underlying asset for both straddles is the S&P500 index, and all options have an

identical striking price. We use at-the-money opening option prices for all transactions. Let

VDGN ,t be the value of a portfolio that makes the DGN strategy on day t,

VDGN ,t = StraddleS ,t − StraddleL,t = QSC CS ,t + QSP PS ,t − QLC,t CL,t − QLP,t PL,t , (1)

where StraddleS ,t and StraddleL,t are the values of the short-position (long-term) and long-

position (short-term) straddles at t, respectively. t = 0, 1, 2,...,T where t=0 is the day after a jump

occurred, and T is one day before the short-term (long-position) straddle expires (T < 250).

Quantities are represented by Q, and C and P are call and put premiums. The straddle held in

long-position is composed of one call and one put options, while quantities of the options making

the short-position straddle are determined daily to comprise the DGN strategy. With respect to

5 We ignore accumulated interest rate profits/losses on both straddles due to their small net amount.

8
(1), we let QL = QL = 1 , thus the target variables that solve for ΔVDGN ,t = ΓVDGN ,t = 0 t
C P

simultaneously are QSC,t and QSP,t ,

 
ΔVDGN ,t = QSC,t N (hS )t + QSP,t (N (hS )t − 1) − 2N (hL )t − 1 = 0 , (2)

where the hedge ratios for the Call and the Put options are ΔtC = N (h)t , and ΔtP = N (h)t − 1 ,

Ln ( St e − d / Ke − r ) + 0.5 t2
respectively. ht = , and N ( h) t is the cumulative standard normal
t 

distribution of short-position (if h = hS ) and long-position (if h = hL ) options. S t is the closing

S&P500 index at t, K is the strike price common to all four options,  is time to expiration in

annual terms, d is the instantaneous dividend yield of the S&P500 and r is the riskless interest

rate, which is measured based on current three month T-bills.  t is the standard deviation of the

underlying asset’s return, where the proxy for  t is the current VIX.6 The second equation solves

for Γ PDGN = 0 ,

ΓVDGN ,t = (QSC,t + QSP,t ) Γ S ,t − Γ L,t = 0 ,

N ' ( hL )t N ' ( hS )t 1 -0 .5ht2

where, Γ L,t  Γ LC,t = Γ LP,t = , Γ S ,t  Γ S ,t = Γ S ,t = and N'(h)t =
C P
e is the
S t t  L S t t  S 2

derivative of N(h)t . Therefore,

N ' (hS ) t N ' (hL ) t

ΓVDGN ,t = (QSC,t + QSP,t ) −2 = 0. (3)
S t t  S S t t  L

Solving (2) and (3) simultaneously yields:

6 VIX is preferable to other measures, such as historical standard deviation as it reflects the market expectation of

future variability. Our qualitative results did not change when the procedure was tested based on historical standard

deviations.

9
 2 N ' ( hL ) t  S
QSC,t = [1 − N ( hS ) t ] + [ 2 N ( hL ) t − 1] , (4)
N ' ( hS ) t  L

2 N ' (hL ) t  S
QSP,t = N (hS ) t − [ 2 N (hL ) t − 1] . (5)
N ' (hS ) t  L

Thus, QSC,t and QSP,t of the short-position straddle are determined each day in order to

maintain a delta and gamma neutral position. The profit or loss of a single straddle portfolio on

date T, a day before the expiration date of the short-position straddle, is

 
T −1
Π DGN ,T = VDGN , 0 − Max ( ST − K , K − ST ) +  (QSC,t − QSC,t −1 )CS ,t + (QSP,t − QSP,t −1 ) PS ,t , (6)
t =1

where the first and the second terms on the RHS of the equation are cumulative cash flows from

the entire portfolio on day 0 and the long-position (short-term) straddle on day T, respectively.

The cash flow on day 0, VDGN,0 is positive, as the long-term (short-position) option premiums are

higher than their respective short-term (long-position) ones. The cash flow on day T depends on

the difference between ST and K of the short-term straddle ( Max(ST -K, K-ST ) ). The third term

represents the rebalancing costs necessary to maintain delta-gamma neutrality over the holding

period. Rebalancing costs depend on the buying or selling amounts of the long-term options

during the holding period, and hence on the changes in volatility and duration of volatility decay.

1 n  DGN ,i ,T
The average yield of DGN is defined by RDGN = 
n i =1 S0
, where n is the number of

strategies we were able to generate from the dataset.

Delta-Neutral (DN) Strategy

To estimate volatility risk premium we construct a DN portfolio one day after a jump

occurs by writing an at-the-money call option and concurrently buying delta of the S&P500

10
index. The resulting portfolio will be instantaneously hedged against small changes in the index,

but not against changes in the level of diffusive risk. As a result, this estimation methodology is

expected to reveal volatility risk premia when implemented on non-jump days, but it might be at

variance with these premia when applied for post-jump risk. Following Bakshi and Kapadia

[2003], the DN strategy at t is

VDN ,t = Ct − N (h)t St , (7)

and the profit or loss at expiration of a single portfolio is

T −1
Π DN ,T = Max( ST − K ,0) − CT −  N (h) t −1 ( S t − S t −1 ) . (8)
t =1

1 n  DN ,i ,T
The average yield is given by RDN = 
n i =1 S0
, and it is calculated across the n

strategies we were able to construct.

We analyze the strategy after positive, negative, and all jumps, as well as for short-,

medium-, and long-term options. By doing so, we expect to shed light on the applicability of DN

to adequately estimate post-jump risk premia.

The premia we estimate after the realization of positive and negative jumps are compared

with the long-term diffusive volatility. To estimate the long-term diffusive risk premia we

compare the DN or DGN immunization strategies on randomly sampled days (jump days and 10

trading days after a jump occurred are excluded), with the relevant post-jump premia (positive,

negative, and total). As noted, we refer to this test as the Random Trade (RT) strategy.

11
 Data

The datasets used in this study are made up of the following time series:

• The S&P500 index (adjusted for dividends and other distributions) and its daily-

adjusted rates of return over 22 years, between January 1, 1990 and December 30,

2011. We use the 1990 data to calibrate the Hampel procedure.

• Daily call and put options prices between 1991 and 2011, written on the S&P500

index, with maturities ranging from 7 to 250 days. While our dataset does not include

bid and ask quotes, none of our profitability measures should be systematically biased.

• Open and Close values of the VIX index.

• Annualized rates of return of T-Bills for three months.

II. SUMMARY STATISTICS

Summary statistics of the S&P500 daily returns are presented in panel A of Exhibit 1. The

rest of Exhibit 1 demonstrates how eliminating extreme returns affect the statistical properties of

the sample: in panel B returns are trimmed by removing extreme returns as discerned by the

Hampel method, while in panel C the removed extreme returns are discerned by GARCH(1,1).

There are 5,546 observations in panel A, out of which 219 were classified as extreme returns by

both the Hampel and GARCH(1,1) methods (3.9%). Hence, 5,327 observations make up the

reminder, Trimmed Data samples.7 Each panel is broken down across the years of our sample

period.

7 An issue may arise if several jumps occur sequentially, causing a number of holding periods to overlap, and hence
their gains might be correlated. To test for independency of nearby jump days we fit an AR model of the form:
DGN_DROR = c + AR(1), where DGN_DROR is the daily mean of the DGN strategy gain under the Hampel

12
 [Exhibit 1]

Daily mean returns increase after excluding extreme returns from 0.023% to 0.040% in

panel B, and to 0.053% in panel C. This means that the Hampel method eliminated less, and/or

smaller, negative returns than GARCH(1,1), or that it has eliminated more of the positive

extreme returns than GARCH(1,1).One can see that both the minimum (maximum) returns of

panel B are higher (lower) in the Hampel method, than the GARCH(1,1) method. i.e., more of the

positive and negative extreme returns remained under GARCH(1,1) but were eliminated by

Hampel (panel C vs. panel A). The reason for this appears to be the rather high conditional

volatility of GARCH(1,1) prior to the arrival of extreme returns. An important indication that the

Hampel method tends to be more efficient than GARCH(1,1) in identifying extreme returns can

be seen through the decline in standard deviations in almost all sub-periods: While the full-

sample standard deviation was 1.186%, it declined to 0.889% after trimming jumps by the

Hampel method (a -25% decline), but when using GARCH(1,1) it declined to 1.010% (only -

15%).

Standard deviation of all daily returns was highest during the financial crisis years; it

declined sharply after removing observations based on the Hampel method (from 1.681% to

1.011%), but moderately changed when outliers were discerned by GARCH (1.681% vs.

1.393%). This finding indicates that the two methods identify outliers differently, possibly with

an overlap. Delving into the details, we found that out of the 219 GARCH outliers and 219

Hampel outliers, only 61 were common to both (about 29%).

A detailed distribution of the count and magnitude of positive and negative outliers along

the years is given in Exhibit 2.

method. Normal days were remarked as NA. We find that c is positive and significantly different from 0 (t-test = 4.4;
p<0.01). Therefore, AR(1) is insignificant, i.e., jump days are not autoregressive.

13
 [Exhibit 2]

The Hampel method discerned more negative outliers than positive ones (123 vs. 96), like

GARCH, which identified 139 negative and 80 positive jumps. The average return of full-period

positive outliers is higher in the Hampel case than in the GARCH(1,1) case (3.87% vs. 2.91%),

and the average of negative jumps was -3.73% for Hampel outliers, but -2.80% for GARCH(1,1)

outliers. These findings show from an additional perspective that GARCH(1,1) fails to identify

the most extreme returns, possibly due to the changing level of conditional volatility. As a result,

it would not be surprising to find that the GARCH(1,1) model underestimates the premia. While

beyond the scope of this paper, it appears that these findings highlight the importance of joint

jump and diffusion conditional models when estimating jump risk premia (Christoffersen, Jacobs,

and Ornthanalai [2012]).

III. VOLATILITY-DECAY RISK PREMIA

This section analyzes average volatility-decay premia of the two trading strategies, DGN

and DN, following Hampel-based and GARCH-based jumps. First, we initiate DGN strategies

one day after a positive or negative jump occurred, and compute their average volatility-decay

premia over short-, medium-, and long-terms. Second, we initiate covered call strategies on the

same days, but maintain DN hedging over the same terms. In each case, we further report the

gains of initiating the relevant strategy on randomly sampled days.

DGN Strategies following Hampel Outliers

The top row in Exhibit 3 shows the count of extreme returns for each period, with a total of

219, and the following row shows the number of straddle portfolios we were able to construct –

513. Average volatility decay premia from implementing DGN strategies after extreme S&P500

returns are presented in Exhibit 3, Panel A. The results are given across the years 1991-2011, and

14
they are broken down by sign; Positive versus Negative, as well as All Jumps. Each of these

categories is further broken down by the terms: short-, medium-, and long-term straddles.

[Exhibit 3[

Panel A shows that 233 of the 513 strategies follow positive jumps and 280 follow negative

jumps, depending on options' availability and jump count. The majority of straddle portfolios,

whether following a positive or a negative jump, are over the short-term, and the minority over

long-term. About 73% of all strategies were profitable, irrespective of the jump sign.

The average volatility-decay risk premium across all terms and both positive and negative

jumps is 2.52% for the average straddle portfolio. It is highest at the long-term, 5.77%, declining

to 3.11% and 1.53% in the medium- and short-terms, respectively. The increasing premium level

over time is consistent with Bakshi and Kapadia [2003].

The average volatility-decay premium after negative jumps is 3.03%. It is highest at the

long-term, 6.69%, declining to 4.19% and 1.72% in the medium- and short-terms, respectively.

The premia after positive jumps are smaller than the premia after negative jumps in almost all

sub-periods, and overall. These economically and statistically significant large differences

capture the higher aversion that investors exhibit toward negative rather than positive jumps.

While joint diffusion-jump models reveal the particular jump-risk-premium, which is

predominantly attributed to negative jumps, our findings highlight the importance of jump sign.

Panel B presents the cost of hedging steady-state volatility risk over the short-, medium-,

and long-terms through DGN straddle portfolios that were initiated on randomly sampled, non-

jump days. The average premia across all holding periods are an insignificant -0.48%, just like

the insignificant individual term premia. Those insignificant gains highlight the absence of

diffusive volatility risk premia in those periods, possibly since the long-term diffusive volatility,

10 days or more after a jump realized, is rather steady.

15
 DN Strategies following Hampel Outliers

Exhibit 4 shows the results of initiating covered call positions and maintaining delta-

neutrality until the option expires at the relevant horizon. This strategy does not effectively hedge

severe changes in moneyness that typically follow extreme returns. It may be more efficient,

however, in estimating volatility-risk premia during non-jump periods.

[Exhibit 4 here[

Our findings indicate that the average volatility-decay premium after all jumps and across

all terms is 1.55%, significantly different from zero (P<0.01). Short-term DN strategies earn

0.59% (P<0.05); however, in the medium- and long-terms the average premia are 2.22% and

4.24%, respectively, and both are significant (P<0.01). The finding whereby the gain increases

with the term of the strategies is attributed to volatility and time decay, as the value of the written

call option declines. These findings are consistent with the same test that Bakshi and Kapadia

[2003] conducted. They found zero gain from initiating DN strategies after jumps, but an annual

average gain of about 3% if initiated randomly. Our findings show how much of the jump risk

premia is lost due to the imperfect hedge of the DN strategy, over time, and contingent on the

jump sign. Because we found an overall 2.52% gain when the DGN was implemented after same

jumps, but 1.55% under DN (highly significant difference, P<0.05), we conclude that about 40%

of the post-jump risk premia is lost by ignoring gamma effects. The differences in volatility

decay premia after positive jumps are not significant, as detailed below.

The average gain from DN strategies that were initiated following negative jumps is 1.12%

(P<0.01). Recalling that the DGN strategy yields an average of 3.03% after negative jumps, we

conclude that the imperfect hedge of the DN strategy implies a loss of more than 60% of the

premia in these states (highly significant difference, P<0.01). Breaking down the post-negative

jumps premia over terms, we find a 0.55% (insignificant) in the short term, 1.66% in the
16
medium-term (P<0.01), and 2.44% in the long-term (insignificant). Volatility decay premia after

positive jumps increase with the terms of the straddles (from 0.63% in the short-term, to 2.84%

and 5.90% in the medium- and long-terms). The medium- and long-term gains are highly

significant, possibly due to a combination of effects, such as reversion of the index to its pre-

jump level, volatility decay, and/or time decay. It should be noted that unlike DGN, under DN

only the written call is affected by volatility and time decay.

DN strategies initiated on random dates yield rather small and insignificant returns,

suggesting that volatility risk is not priced. Recall that the gain from the hedged position should

pay a premium (in excess of the riskless rate, as we measure here), if volatility risk is priced.

However, if it is not priced, either because volatility is not risky or because investors do not

demand premia for it, the strategy should yield zero return. Our findings highlight that if sampled

10 days or longer after jumps occurred, DN-based volatility risk is not priced.

DGN Strategies following GARCH(1,1) Outliers

The DGN premia reported in Exhibit 5, coupled with the count of strategies that

GARCH(1,1) identified across periods, highlight the key weakness of GARCH(1,1) in trading on

volatility decay. Note first that out of the total 927 strategies, 397 were implemented between

1991-1996, and 312 between 2002-2006, a total of 709, or 76.5%. These two sub-periods were

the least volatile in our sample: as shown in Panel A of Exhibit 1, daily standard deviation of the

S&P500 return was lowest between 1990-1996, 0.723%, followed by 1.014% between 2002-

2006 as the second lowest standard deviation.

[Exhibit 5]

Apparently, the low conditional volatility during those periods made many relatively

extreme returns appear as if they were jumps for the GARCH(1,1) procedure, while their absolute

return was rather small. This can be seen in Exhibit 2, where the average return of positive
17
GARCH-based jumps were 1.75% between 1991-1996 and 2.61% between 2002-2006, lower

than the average, 2.91%. A similar finding holds for negative GARCH-based jumps during those

years: -1.86% and -2.18%, against an average of -2.80%. These GARCH-based outliers are

materially smaller than Hampel-based outliers in those two sub-periods: 3.29% and 3.74% in the

positive domain, and -3.14% and -3.12% in the negative domain. Still, during those low-volatility

periods, the average GARCH-based DGN premia was positive, 2.11% between 1991-1996, and

0.17% between 2002-2006. However, during the more volatile periods – 1997-2001 of the

internet bubble and 2007-2011 of the housing bubble and the financial crisis, the strategy was

losing. Overall, then, GARCH-based strategies performed poorer than Hampel-based strategies,

0.89% vs. 2.52%. Technically, it seems like GARCH follows the auto-regressive trend, including

the volatility decay, and therefore cannot reveal the volatility-decay risk premia.

IV. EXPLANATORY FACTORS FOR VOLATILITY DECAY PREMIA

The premia that DGN strategies reveal after a jump occurs would be considered a riskless

arbitrage opportunity if investors could implement them and reap a riskless profit. Alternatively,

risky arbitrage opportunities would prevail if the premia were highly correlated with some factor.

Efficient market arguments rule out riskless premia, therefore if a premium is found in the data it

would imply either that the market is inefficient, or that the premium is risky. We test the latter

alternative on plausible risk factors by using a regression model that accounts for the following

explanatory variables:

(Δ𝑆𝑃𝑋)2 – Squared percentage change of the S&P500 index from its’ period zero

level, S 0 , throughout the span of each DGN strategy. Hence, this variable captures

bidirectional changes in moneyness.

𝐷𝑇𝐸 – Days to Expiration, between the strategy's origination and expiration,

18
 Δ𝑉𝐼𝑋 – Change in volatility, between the first and last days of each DGN strategy,

which proxies for vega effects (a positive gap implies volatility decay).

Positive – obtains 1 if the extreme return was positive or 0 if it was negative.

In addition, we account for possible interactions between the first three variables. The

regression equation is:

𝛱𝐷𝐺𝑁 = 𝛼0 + 𝛼1 𝛥𝑆𝑃𝑋 2 + 𝛼2 𝐷𝑇𝐸 + 𝛼3 𝛥𝑉𝐼𝑋 + 𝛼4 𝛥𝑆𝑃𝑋 2 ∗ 𝐷𝑇𝐸 + 𝛼5 𝛥𝑆𝑃𝑋 2 ∗ 𝛥𝑉𝐼𝑋 +

𝛼6 𝐷𝑇𝐸 ∗ 𝛥𝑉𝐼𝑋 + 𝛼7 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒 + 𝜖 (9)

Note that because the variables are denominated in different terms, e.g., VIX in percentage

points while SPX in points, we present elasticities-at-mean. This technique allows one to

compare the relative impact of each variable (at the mean) on the dependent variable, irrespective

of its nominal level. The sum of elasticities-at-mean coefficients is 1.0. In order to account for

possible overlapping observations, the regression models use the HAC (Heteroskedasticity and

Autocorrelation Consistent Covariance) procedure of Newey and West [1987].

We implemented the regression model (9) three times: First for DGN strategies that were

initiated after Hampel based jumps, and second after GARCH(1,1) jumps. These will differ as

long as the specific method-based jumps differ in their magnitude, duration, or sign. Third, we

regress those explanatory factors on DGN strategies that were initiated on randomly sampled

days away from jump events. In the latter case, the null hypothesis is that no factor be significant,

except for time decay. The regressions results are summarized in Exhibit 6.

[Exhibit 6]

19
 Focusing first on Hampel-based results, we find highly significant coefficients for

(Δ𝑆𝑃𝑋)2 , Δ𝑉𝐼𝑋, and each of the interactions variables, (Δ𝑆𝑃𝑋)2 ∗ 𝐷𝑇𝐸, (Δ𝑆𝑃𝑋)2 ∗ Δ𝑉𝐼𝑋, and

𝐷𝑇𝐸 ∗ Δ𝑉𝐼𝑋. Other than (Δ𝑆𝑃𝑋)2 ∗ Δ𝑉𝐼𝑋, all of these factors have economically large elasticity-

at-mean factor loadings, and an insignificant intercept (P=0.10), highlighting a good explanatory

power for volatility decay premia. This is evident by the relatively high Adjusted 𝑅 2 of this

regression, 33%. Lastly, the dummy coefficient for positive jumps turns out to be negative and

significant, suggesting that the volatility decay premia after positive Hampel-based jumps are

significantly smaller than the premia after negative jumps.

Further exploring elasticity-at-mean values, one can readily notice that VIX (volatility

decay) has the largest positive impact (6.993) on Hampel-based DGN gains, followed by DTE

(4.163). On the other hand, squared changes in moneyness, (Δ𝑆𝑃𝑋)2 , are negatively related to

DGN gains (-2.402). These results are consistent with those of Exhibit 3 regarding the effects that

vega, DTE, and positive versus negative jumps had on the DGN gains.

Second, we turn to explore the volatility decay premia that DGN strategies gained after

GARCH(1,1) extreme returns, in the middle section of Exhibit 6. One can readily notice that

none of the factors is significant, except for marginal significance of the interaction (Δ𝑆𝑃𝑋)2 ∗

Δ𝑉𝐼𝑋, with P=0.09. In particular, the intercept is significant, and unlike the Hampel-based

regression, it is positive instead of being negative. The regression’s Adjusted 𝑅 2 is 10%, far

lower than the Hampel-based regression. We therefore conclude that the GARCH(1,1) model

fails to identify those jumps for which the market pays a volatility-decay premium. As noted

above, we further used the GJR-GARCH model, which allows the conditional variance to

respond differently to past positive and negative innovations, but the qualitative results remain

similar to the GARCH(1,1) model.

20
 The third regression, in the left section of Exhibit 6, presents the explanatory power of the

factors when the DGN strategy was implemented on randomly sampled, non-jump days. As

expected, none of the factors is significant, with the one exception of marginal significance for

Δ𝑉𝐼𝑋 (P=0.07). This regression has the lowest adjusted 𝑅 2 of the three – 9%.

In summary, volatility decay premia turn out to be a risky gain that can be explained by the

variation of the underlying index return, by time to expiration, and by the uncertain changes in

volatility (vega effects), as well as by the interactions between these three variables.

Nevertheless, in order to identify and price this risk one must distinguish effectively between

periods of high and periods of low diffusive volatility. Our findings indicate that the Hampel

method outperforms GARCH(1,1) in identifying jumps, and hence measures more effectively

volatility-decay risk premia, possibly due to the changing level of conditional volatility. Put

differently, given low conditional volatility GARCH(1,1) might identify a rather small extreme

return as a jump, while given high conditional volatility it might overlook more substantial

extreme returns. An alternative explanation is that under GARCH the tails are conditionally

normal (by assumption), and for typical parameter values the conditional volatility adjusts

quickly to new data. Therefore, in GARCH models the effect of historical extreme returns decays

rapidly, while the Hampel procedure has a “longer memory”. Eventually, a big GARCH

innovation is less surprising to investors than a big Hampel innovation.8

V. SUMMARY

Empirical evidence suggests that once a jump event has occurred, diffusive risk is often

higher than its long-run level, and it decays gradually. This gradual decay property suggests that

diffusive risk premia may be time-varying as well: higher immediately after a jump has been

8 We thank Steve Figlewski for highlighting this possibility.

21
realized and gradually declining toward the long-run level. Since investors can form options

positions that would yield positive gains based on the declining level of volatility, and hedge

those positions dynamically, these predictable patterns offer a potentially profitable opportunity.

This paper analyses volatility-decay risk premia by deploying options' portfolios aimed at

gaining from this specific timing, and comparing between the gains of delta hedging alone (DN)

and delta-gamma hedging (DGN) for short-, medium-, and long-term horizons, while controlling

for jump sign. Further, we apply two different techniques for identifying jumps – the Hampel

[1971] method, and the commonly used GARCH(1,1) and GJR methods.

We find that the average gain that DGN strategies yield is higher than DN strategies, and

that Hampel-based jumps are informative for the distinction between high- versus low-post-jump

volatility. The resulting premia are risky, and they are explained by changes in the level of the

volatility index VIX, changes in the S&P500 index, and time-to-expiration of the options

portfolio. We further find that interactions of these variables are significantly and economically

important, and that positive jumps carry a lower volatility-decay risk premia.

An important finding pertains to the failure of GARCH(1,1), as well as the GJR model that

appears to fit best equity returns, to identify the particular extreme returns after which diffusive

volatility increases high enough to warrant a volatility-decay premium.

22
References

Andersen, T.G., T. Bollerslev, and F.X. Diebold. "Roughing It Up: Including Jump Components

in the Measurement, Modeling and Forecasting of Return Volatility." Review of Economics

and Statistics, 89, (2007), pp. 701–720.

Bakshi G., and N. Kapadia. "Delta‐hedged Gains and the Negative Market Volatility Risk

Premium." Review of Financial Studies, 16 (2003), pp. 527–566.

Barndorff-Nielsen, O.E., and N. Shephard. "Estimating Quadratic Variation using Realized

Variance." Journal of Applied Econometrics, 17 (2002), pp. 457–477.

Barndorff-Nielsen, O.E., N. Shephard,. "Power and Bipower Variation with Stochastic Volatility

and Jumps." Journal of Financial Econometrics, 2 (2004), pp. 1–37.

Bates, D.S. "Pricing Options on Jump-diffusion Process." Working Paper. Rodney L. White

Center, Wharton School, 1988.

Bates, D.S. "Post-’87 Crash Fears in the S&P 500 Futures Option Market." Journal of

Econometrics, 94 (2000), pp. 181–238.

Becker, R., A.E. Clements, A. McClellan. "The Jump Component of S&P 500 Volatility and the

VIX Index." Journal of Banking and Finance, 33 (2009), pp. 1033–1038.

Bollerslev, T. "Generalized Autoregressive Conditional Heteroskedasticity," Journal of

Econometrics 31 (1986), pp. 307-327.

Bollerslev, T., M. Gibson, and H. Zhou. "Dynamic Estimation of Volatility Risk Premia and

Investor Risk Aversion from Option-implied and Realized Volatilities." Journal of

Econometrics (2010), doi:10.1016/j.jeconom.2010.03.033

Bollerslev, T., U. Kretschmer, C. Pigorsch, and G. Tauchen. "A Discrete-time Model for Daily

S&P500 Returns and Realized Variations: Jumps and Leverage Effects." Journal of

Econometrics, 150 (2009), pp. 151–166.

23
Christoffersen, P., K. Jacobs, and C. Ornthanalai. "Dynamic Jump Intensities and Risk

Premiums: Evidence from S&P500 Returns and Option.", Journal of Financial Economics,

106 (2012), pp. 447–472.

Coval J., and T. Shumway. "Expected Option Returns." Journal of Finance, 42 (2001), 983–

1009.

Duan, J.C. "The GARCH Option Pricing Model." Mathematical Finance, 5 (1995), pp. 13–32.

Engle, R. F. "Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of

U.K. Inflation," Econometrica 50 (1982), pp. 987-1008.

Eraker, B. "Do Stock Prices and Volatility Jump? Reconciling Evidence from Spot and Option

Prices." Journal of Finance, 59 (2004), pp. 1367–1403.

Eraker, B., M. Johannes, and N. Polson "The Impact of Jumps in Volatility and Returns." The

Journal of Finance, 58 (2003), pp. 1269–1300.

Fama, E. F. "The Behavior of Stock Market Prices," Journal of Business, 38 (1965), pp. 34-105.

Hampel, R.F. "A General Qualitative Definition of Robustness." The Annals of Mathematical

Statistics, 42 (1971), 6, pp. 1887–1896.

Heston, S., and S. Nandi. "A Closed-form GARCH Option Pricing Model." Review of Financial

Studies, 13 (2000), pp. 585–626.

Glosten R.T, R., Jagannathan, and D. Runkle, “On the relation between the expected value and

the volatility of the nominal excess return of stocks”, The Journal of Finance, 48, 1779-

1801.

Low B.S., and S. Zhang. "The Volatility Risk Premium Embedded in Currency Options." Journal

of Financial and Quantitative Analysis, 40 (2005), pp. 803–832.

24
Newey, W.K., and K.D. West. "A Simple, Positive Semi-Definite, Heteroskedasticity and

Autocorrelation Consistent Covariance Matrix." Econometrica, 55 (1987), pp. 703–708.

Olewuezi, N.P. "Note on the Comparison of Some Outlier Labeling Techniques." Journal of

Mathematics and Statistics, 7, (2011): pp. 353–355.

Pan, J. "The Jump-risk Premia Implicit in Options: Evidence from an Integrated Time-series

Study." Journal of Financial Economics, 63 (2002), pp. 3–50.

Ritchken, P., and R. Trevor. "Pricing Options under Generalized GARCH and Stochastic

Volatility Processes." Journal of Finance, 54 (1999), pp. 377–402.

Zhang, J.E., J. Shu, and M. Brenner. "The New Market for Volatility Trading." Journal of

Futures Markets, 30 (2010), pp. 809–833.

25
Exhibit 1
Summary Statistics 1991-2011 S&P500 Returns, and Trimmed Data

(A) Daily rates of return (%)

Period All data
Mean Std. Min Max # Obs.

1991-1996 0.041 0.723 -3.720 3.664 1,770

1997-2001 0.035 1.268 -7.107 4.982 1,257
2002-2006 0.017 1.014 -4.248 5.571 1,259
2007-2011 -0.010 1.681 -9.473 10.954 1,260
Total 0.023 1.186 -9.473 10.954 5,546

(B) Daily rates of return (%) - Hampel Method

Period
Trimmed data - Hampel

1991-1996 0.043 0.704 -3.092 3.129 1,765

1997-2001 0.044 1.035 -2.475 2.633 1,202
2002-2006 0.012 0.853 -2.508 2.576 1,226
2007-2011 0.061 1.011 -2.506 2.660 1,134
Total 0.040 0.889 -3.092 3.129 5,327

(C) Daily rates of return (%) - GARCH(1,1) Method

Trimmed data - GARCH
1991-1996 0.051 0.653 -3.092 3.129 1,717
1997-2001 0.062 1.079 -3.499 3.829 1,206
2002-2006 0.015 0.904 -3.485 4.633 1,214
2007-2011 0.084 1.393 -6.956 6.693 1,190
Total 0.053 1.010 -6.956 6.693 5,327

S&P500 returns are based on closing values. Non-consecutive trading days (other than Friday-Monday) were
excluded from the sample for not being daily returns. "Trimmed data" refers to original data without extreme
observations, as those were identified by the specific method.

26
Exhibit 2
Average Extreme Returns by Type and by Year, 1991-2011

(A) Hampel Method

Positive extreme Negative extreme All extreme

observations observations observations
Average Average Average
Period Count return (%) Count return (%) Count return (%)

1991-1996 2 3.29 3 -3.14 5 -0.57

1997-2001 26 3.53 29 -3.47 55 -0.16
2002-2006 16 3.74 17 -3.12 33 0.20
2007-2011 52 4.11 74 -3.99 126 -0.65
Total extreme
observations 96 3.87 123 -3.73 219 -0.40

(B) GARCH(1,1) Method

Average Average Average
Period Count return (%) Count return (%) Count return (%)

1991-1996 23 1.75 30 -1.86 53 -0.29

1997-2001 20 3.37 31 -3.15 51 -0.60
2002-2006 21 2.61 24 -2.18 45 0.06
2007-2011 16 4.40 54 -3.38 70 -1.61
Total extreme
observations 80 2.91 139 -2.80 219 -0.71

Nonconsecutive trading day returns were excluded for being non-daily returns.

27
Exhibit 3
Delta-Gamma-Neutral (DGN) Strategies after Jumps versus Random Trade
Years 1991-1996 1997-2001 2002-2006 2007-2011 Total
# of extreme obs. 5 55 33 126 219
# of DGN strategies 10 95 52 356 513
Panel A: Delta-Gamma-Neutral Strategies after jumps
Positive jumps Count Yield Count Yield Count Yield Count Yield Count Yield
Long term 0 0.00% 4 2.71% 4 3.49% 16 5.84% 24 4.93%***
Medium term 2 0.53% 20 0.68% 11 1.77% 61 2.37% 94 1.90%**
Short term 3 -1.76% 33 0.31% 10 4.05% 69 1.48% 115 1.28%***
All 5 -0.85% 57 0.61% 25 2.96% 146 2.33% 233 1.91%***
Negative jumps
Long term 2 2.81% 3 9.43% 3 8.44% 14 6.29% 22 6.69%
Medium term 3 1.72% 15 1.70% 8 2.51% 79 4.92% 105 4.19%***
Short term 0 NA 20 2.32% 16 2.66% 117 1.49% 153 1.72%**
All 5 2.15% 38 2.64% 27 3.26% 210 3.10% 280 3.03%***
All jumps
Long term 2 2.81% 7 5.59% 7 5.61% 30 6.05% 46 5.77%***
Medium term 5 1.24% 35 1.12% 19 2.08% 140 3.81% 199 3.11%***
Short term 3 -1.76% 53 1.07% 26 3.19% 186 1.48% 268 1.53%***
All 10 0.65% 95 1.42% 52 3.11% 356 2.78% 513 2.52%***
Panel B: Delta-Gamma-Neutral strategies on Random days

Long term 10 0.47% 5 -2.01% 34 -4.84% 22 -0.84% 71 -2.65%

Medium term 90 0.01% 32 0.90% 94 -0.74% 83 -1.07% 299 -0.43%
Short term 121 0.38% 66 -1.10% 154 0.41% 94 -1.11% 435 -0.15%
All 221 0.23% 103 -0.52% 282 -0.61% 199 -1.06% 805 -0.48%
Yield is calculated by dividing the monetary gain or loss by the S&P500 index for each strategy. We simulate the strategy on Hampel-based outliers (Panel A),
and on random days (Panel B). DGN is preferable to Random Trade along terms and almost in all sub-periods.
Significance level is presented for the entire period only. * = P<0.10; ** = P<0.05; *** = P<0.01

28
Exhibit 4
Delta-Neutral (DN) Strategies: After Jumps versus Random Trade
Years 1991-1996 1997-2001 2002-2006 2007-2011 Total
# of jumps 5 55 33 126 219
# of DN strategies 10 95 52 356 513
Panel A: Delta Neutral Strategies after jumps
Positive jumps Count Yield Count Yield Count Yield Count Yield Count Yield
Long term 0 NA 4 2.03% 4 7.58% 16 6.44% 24 5.90%***
Medium term 2 1.86% 20 1.08% 11 2.11% 61 3.59% 94 2.84%***
Short term 3 -1.34% 33 -0.82% 10 0.16% 69 1.48% 115 0.63%
All 5 -0.06% 57 0.05% 25 2.21% 146 2.90% 233 2.07%***
Negative jumps
Long term 2 -2.05% 3 7.57% 3 1.75% 14 2.12% 22 2.44%
Medium term 3 0.75% 15 -1.20% 8 0.15% 79 2.39% 105 1.66%***
Short term 0 NA 20 -1.40% 16 0.07% 117 0.95% 153 0.55%
All 5 -0.37% 38 -0.61% 27 0.28% 210 1.57% 280 1.12%***
All jumps
Long term 2 NA 7 4.41% 7 5.08% 30 4.43% 46 4.24%***
Medium term 5 1.20% 35 0.10% 19 1.29% 140 2.91% 199 2.22%***
Short term 3 -1.34% 53 -1.04% 26 0.10% 186 1.15% 268 0.59%**
All 10 -0.21% 95 -0.22% 52 1.21% 356 2.12% 513 1.55%***
Panel B: Delta-Gamma-Neutral strategies on Random days
Long term 10 0.63% 5 2.75% 34 -1.10% 22 1.59% 71 0.16%
Medium term 90 0.37% 32 0.43% 94 -0.60% 83 -0.28% 299 -0.09%
Short term 121 -0.11% 66 0.55% 154 -0.08% 94 0.29% 435 0.09%
All 221 0.13% 103 0.60% 282 -0.37% 199 0.19% 805 0.03%
Yield is calculated by dividing the monetary gain or loss by the S&P500 index for each strategy. We simulate the strategy on Hampel-based outliers (Panel A),
and on random days (Panel B). DN loses on average in all sub-periods and across all terms. While DN yields are significantly negative, Random Trade yields are
negative, albeit not significantly. Significance level is presented for the entire period only. * = P<0.10; ** = P<0.05; *** = P<0.01

29
Exhibit 5
Delta-Gamma-Neutral (DGN) Strategies after GARCH(1,1) Jumps
Years 1990-1996 1997-2001 2002-2006 2007-2011 Total
# of extreme obs. 53 51 45 70 219
# of DGN strategies 397 55 312 163 927
Panel A: Delta-Gamma-Neutral Strategies after extreme observations
Positive extreme Coun Coun Coun Coun Coun
obs. t Yield t Yield t Yield t Yield t Yield
Long term 16 15.27% 2 -8.40% 21 -1.13% 9 0.03% 48 4.25%
Medium term 118 0.29% 7 1.94%* 66 -0.42% 12 -1.39% 203 0.01%
Short term 114 1.08%* 14 -5.72% 91 -0.06% 22 -0.26% 241 0.13%
All 248 1.62%** 23 -3.62% 178 -0.32% 43 -0.52% 492 0.48%
Negative extreme obs.
Long term 6 1.10% 4 5.07% 26 2.88% 20 -1.11% 56 1.42%
Medium term 62 6.01%** 9 2.92% 37 0.43% 33 2.13% 141 3.44%**
Short term 81 0.73% 19 1.45% 71 0.28% 67 -1.32%** 238 0.08%
All 149 2.94%** 32 2.32%* 134 0.83% 120 -0.33% 435 1.34%**
All extreme obs.
Long term 22 11.41%* 6 0.58% 47 1.09% 29 -0.75% 104 2.73%
Medium term 180 2.26%** 16 2.49%* 103 -0.12% 45 1.19% 344 1.42%**
Short term 195 0.93%** 33 -1.59% 162 0.09% 89 -1.06%** 479 0.10%
All 397 2.11%** 55 -0.17% 312 0.17% 163 -0.38% 927 0.89%**
Panel B: Delta-Gamma-Neutral strategies on Random days
Long term 14 7.29% 4 -2.49% 33 -5.06% 17 -1.93% 68 -1.58%
Medium term 102 3.02%* 30 1.05% 92 -0.92% 72 -0.79% 296 0.67%
Short term 130 0.79%** 66 -1.10%** 154 0.41%** 93 -0.81% 443 0.04%
All 246 2.08%** 100 -0.51% 279 -0.67% 182 -0.91% 807 0.13%
This table is similar to Exhibit 3, except for the identification of outlying days, which are determined here by GARCH(1,1) rather than Hampel's method (for
details see Exhibit 3).

30
 Exhibit 6
The Factors affecting Volatility Decay Premia
Hampel GARCH(1,1) Control group
Coef. Elasticity Coef. Elasticity Coef. Elasticity
Variable [Prob.] at mean [Prob.] at mean [Prob.] at mean
Intercept -0.83 -4.025 0.64 7.358 -0.06 -2.480
[0.10] [0.03] [0.85]
SPX2 -210.11 -2.402 -54.51 -0.522 -128.64 -4.474
[0.04] [0.81] [0.65]
DTE 0.27 4.163 -0.15 -5.779 0.04 5.169
[0.09] [0.09] [0.69]
VIX 30.80 6.993 5.87 0.016 23.82 -2.819
[0.00] [0.46] [0.07]
SPX2*DTE 73.66 2.694 2.42 0.079 29.22 3.357
[0.03] [0.97] [0.73]
SPX2*VIX -1,184.98 -0.695 -837.12 -0.051 106.43 -0.029
[0.00] [0.09] [0.84]
DTE*VIX -7.01 -5.317 0.01 0.000 -5.53 2.276
[0.00] [0.99] [0.16]
Positive -0.19 -0.411 -0.02 -0.102
[0.05] [0.79]

Adj. R2 0.33 0.10 0.09

D.W. 2.07 2.21 3.04
Elasticity at mean enables the evaluation of relative influence of each explanatory variable on the dependent variable, here DGN gain. Elasticity at means are
point estimates calculated based on coefficients of the explanatory variables, the mean of the variables (not presented here), and the mean of the dependent
variable. Therefore, their sign need not be identical to the sign of the coefficients. The sum of all elasticities at means equal 100%.

31

View publication stats