The Performance of Implied Volatility in

Forecasting Future Volatility: an Analysis of

Three Major Equity Indices from 2004 to 2010
By
MASSACHUSETTS INSTITUTE
OF TECHNOLOGY
Vladimir M. lonesco

B.Sc. in Applied Mathematics and Computer Science

JUN 1 5 2011
University of Paris 1 Pantheon Sorbonne, 2008
LIBRARIES
Submitted to the MIT Sloan School of Management
In partial fulfillment of the Requirements for the Degree of ARCHiNES
MASTER OF SCIENCE IN MANAGEMENT STUDIES

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

JUNE 2011

© 2011 Vladimir M. Ionesco. All Rights Reserved.

The author hereby grants to MIT permission to reproduce and to

distribute publicly paper and electronic copies of this thesis document in
whole or in part in any medium now known or hereafter created.

Signature of Author:

MIT Sloan School of Management

May 6,2011

Certified By:

Eric Jacquier
Visiting Associate Professor of Finance
Thesis Advisor
MIT Sloan School of Management

Accepted By:

USMR Distinguished Professor

Michael A. Cusumano
of Management
gram Director, M.S. in Management Studies Program
MIT Sloan School of Management
 The performance of implied volatility in forecasting
future volatility: an analysis of three major equity
indices from 2004 to 2010
by

Vladimir M. lonesco

Submitted to the MIT Sloan School of Management on May 6, 2011 in partial fulfillment
of the requirements for the degree of Master of Science in Management Studies.

Abstract
In this thesis, we investigate whether implied volatility is an efficient es-
timator of future one-month volatility from an informational perspective and
whether it outperforms historical volatility in this regard.

We first compare the predictive powers of implied volatility, simple histori-

cal volatility, and exponential historical volatility, using monthly observations
of the S&P 500, FTSE 100, and DAX equity and option markets from 2004
to 2010.

Then, we introduce a GARCH(1,1) model and compare in-sample GARCH-

fitted volatility and implied volatility from 2004 to 2010, as well as out-of-
sample GARCH-forecasted volatility and implied volatility from 2005 to 2010,
using data on the S&P 500.

We find that implied volatility is not only an efficient estimator of future

volatility, but also that its information content is at least as good, if not much
better, than that of historical volatility. Our results also suggest that im-
plied volatility systematically subsumes the information included in historical
volatility, even when a GJR-GARCH model is utilized.

Thesis Advisor: Eric Jacquier

Title: Visiting Associate Professor of Finance
Contents
1 Introduction 4

2 Methodology
2.1 Data specifications .......
2.2 Sampling procedure . . . . . . .
2.3 Variable definitions . . . . . . .
2.3.1 Implied volatility . . . .
2.3.2 Historical volatility . . . . . . . . . . . . . . . .
2.3.3 Ex-post future volatility . . . . . . . . . . . . .
2.4 Regressions..... . . . .. . . . . . . . . . . . . . .

3 Descriptive statistics 10

4 Empirical Results 16
4.1 S&P 500 . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 FTSE 100 . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 DAX ..... ........................... 18
4.4 Residual diagnostics . . . . . . . . . . . . . . . . . . 20
4.5 Intermediary conclusion . . . . . . . . . . . . ... 22

5 Introducing GARCH 22
5.1 The GJR-GARCH(1,1) model . . . . . . . . . . . . . 22
5.2 In-sample GARCH-fitted values v. implied volatility. 23
5.2.1 Methodology . . . . . . . . . . . . . . . . . . 23
5.2.2 Empirical results . . . . . . . . . . . . . . . . 24
5.3 Out-of-sample GARCH forecasts v. implied volatility 25
5.3.1 Methodology . . . . . . . . . . . . . . . . . . 25
5.3.2 Empirical results . . . . . . . . . . . . ...
. 26

6 General conclusion
List of Tables
1 S&P 500: Descriptive statistics . . . . . . . . . . . . . . . . . 12
2 FTSE 100: Descriptive statistics . . . . . . . . . . . . . . . . . 13
3 DAX: Descriptive statistics . . . . . . . . . . . . . . . . . . . . 14
4 S&P500: OLS estimates . . . . . . . . . . . . . . . . . . . . . 17
5 FTSE 100: OLS estimates . . . . . . . . . . . . . . . . . . . . 19
6 DAX: OLS estimates . . . . . . . . . . . . . . . . . . . . . . . 21
7 GJR-GARCH(1,1) parameter estimates . . . . . . . . . . . . . 23
8 GARCH-fitted values v. implied volatility: OLS estimates . . 25
9 GARCH forecasts v. implied volatility: OLS estimates . . . . 27

List of Figures
1 S&P 500: simple and exponential historical volatility . . . . . 11
2 S&P, FTSE, and DAX: simple historical volatility and implied
volatility........ . . . . . . . . . . . . . . . . . . . . . . 15
3 GARCH-fitted volatility . . . . . . . . . . . . . . . . . . . . . . 24
4 GARCH-forecasted volatility . . . . . . . . . . . . . . . . . . . . 26
1 Introduction
Volatility is a most critical concept in both the theory and practice of
finance. It is a measure of the uncertainty about a financial instrument's
probability distribution of returns and is commonly defined as the standard
deviation of the returns of the said instrument within a set time span. Volatil-
ity is thus intimately linked to the fundamental concept of risk, "the central
element that influences financial behavior" as Nobel Laureate Robert C. Mer-
ton once put it. Consequently, many investors seek to forecast volatility, for
risk management, portfolio selection, or valuation purposes, or for designing
trading strategies, such as volatility arbitrage. To do so, they mainly rely
on two estimators: historical volatility, also known as past realized volatility,
and implied volatility, a concept formally born in 1973 with the publication
of the Black-Scholes option pricing model and the creation of the Chicago
Board Options Exchange (CBOE), the first market of its kind in the United
States. While historical volatility is directly computable from past market
data, implied volatility is extracted from option prices, using models such as
Black-Scholes.

The rationale behind the use of historical volatility to forecast future volatility
lies in the idea that the past tends to repeat itself, which leads to common
financial principles such as mean-reversion. By definition, simple historical
volatility is an unconditional predictor that ignores the most recent publicly
available data. Moreover, when computed using standard statistical tech-
niques, historical volatility fails to reflect the possible predictability of "true"
volatility. In contrast, implied volatility is widely seen as the market's esti-
mate of future volatility, and if markets are efficient, it should thus reflect
all the information available at a given time, including that contained in his-
torical volatility. Bodie and Merton (1995) use the period that precedes the
Persian Gulf War of 1991 to illustrate the superiority of implied volatility over
historical volatility when estimating future volatility.

Nevertheless, despite the widely-shared belief among finance practitioners that

implied volatility is, indeed, a much better estimator than historical volatility,
research has produced intriguingly divergent results about whether implied
volatility actually estimates future volatility or whether it does so efficiently.

Early academic articles support the superiority of the forecasting characteris-

tics of implied volatility over those of historical volatility. In the first study on
this kind, Latane and Rendleman (1976) test the relationship between an aver-
age of implied volatilities and subsequent realized volatility, using the closing
call option and stock prices of twenty four companies traded on the Chicago
Board Options Exchange. They conclude that weighted implied volatility is
generally a better predictor of future volatility than historical volatility. Chi-
ras and Manaster (1978) reach the same conclusion, after analyzing twenty
three monthly observations between 1973 and 1975. Although the authors
discover no significant difference between the forecasting characteristics of
historical volatility and implied volatility in the first nine months covered by
their study, they find that implied volatility becomes a much better predictor
of future volatility in the following fourteen months. The findings of Beck-
ers (1981) also suggest that implied volatility incorporates and outperforms
the predictive information of historical volatility. However, in the absence of
large time-series data, the above-mentioned research resorts to a static cross-
sectional regression approach.

Using time-series data in a dynamic context, several later works reach opposite
conclusions after studying the actively-traded OEX options on the S&P 100
index. Day and Lewis (1992) analyze these options between 1983 and 1989
and conclude that, although implied volatility may contain some information
about subsequent volatility, it is still outperformed by time-series models of
conditional volatility, such as GARCH and EGARCH. In their study of in-
dividual stock options from 1982 to 1984, Lamoureux and Lastrapes (1993)
also find that the information contained in historical volatility is superior to
that contained in implied volatility. Canina and Figlewski (1993) reach a more
radical conclusion, arguing that "implied volatility has virtually no correlation
with future volatility" and that "it does not incorporate the information con-
tained in recent observed volatility".

However, recent research, notably since Christensen and Prabhala (1998),

tends to support the idea that implied volatility does not only contain sub-
stantial information about future volatility, but that it is also more predictive
than historical volatility. Christensen and Prabhala (1998) study OEX options
with longer time series and non-overlapping data covering the period between
November 1983 and May 1995. They show that implied volatility is an unbi-
ased and efficient estimator of subsequent volatility and that, in some cases,
implied volatility includes the information contained in past realized volatil-
ity. Hansen (1999) analyzes the Danish option and equity markets and also
concludes that implied volatility is a good forecaster of subsequent realized
volatility, that its bias is negligible, and that it incorporates the information
of historical volatility. Christensen and Hansen (2002) confirm the results of
Christensen and Prabhala (1998) in their study of a more recent period with
checks of robustness. In addition, they extend their analysis to put options
and find that put implied volatility is also predictive, though not as much
as call implied volatility. A subsequent study of the S&P 500 index and its
options by Shu and Zhang (2003) also support the idea that implied volatility
is a superior predictor of future volatility. In a different context, Szakmary et
al. (2003) study thirty five options markets and find that, for a large majority
of commodities, implied volatility outperforms historical volatility in forecast-
ing the volatility of the underlying prices. More recently, after analyzing the
S&P/ASX 200 index options traded on the Australian Stock Exchange, Li and
Yang (2008) also conclude that the implied volatilities of both calls and puts
are better than historical volatility at predicting subsequent volatility. More-
over, they find that the volatility implied in call options is a nearly unbiased
estimator of future volatility.

Contrasting with past findings, recent research thus gives credit to the widely-
shared belief that implied volatility does contain some information about fu-
ture volatility, and that it is superior in this respect to historical volatility.

By analyzing the time-series data of three major equity indices - S&P500,

FTSE 100, and DAX - and their options from 2004 to 2010, we aim to inves-
tigate whether the forecasting power of implied volatility for future volatility,
as demonstrated by Christensen and Prabhla (1998) and confirmed by subse-
quent research, is still verified in the most recent years and in both American
and foreign major option and stock markets. In this respect, the aftermath of
the financial crisis of 2008 deserves a close examination.

2 Methodology
2.1 Data specifications
Our empirical analysis focuses on the S&P 500 (SPX), FTSE 100 (ESX),
and DAX equity indices and their options. The data covers the period from
January 2004 to December 2010 for the S&P 500 and the DAX, and the period
from November 2004 to December 2010 for the FTSE 100. The option data
used in the present study has been provided by IVolatility.com.

S&P 500 options are European-style options traded on the Chicago Board
Option Exchange (CBOE). The options are quoted in index points, with a
multiplier of US$100. The average daily volume in January-November 2010
was 711,231. The last trading day is the last business day before the third
Friday of the expiration month. The expiration date is the Saturday following
the third Friday of the expiration month. The nearest twelve calendar months
are available for trading. SPX options are cash-settled.

FTSE 100 options are European-style options traded on the NYSE Liffe Lon-
don exchange. The options are quoted in index points, with a multiplier of
10. The last trading day is the third Friday of the expiration month. Trad-
ing stops after 10:15 am, London time. In the event of the third Friday not
being a business day, the last trading day is the last business day prior such
Friday. The settlement day is the business day following the last trading day.
The expiration months are the nearest eight of March, June, September, and
December, as well as the nearest four calendar months. FTSE100 options are
cash-settled.

DAX options are European-style options traded on the Frankfurt Stock Ex-
change. The options are quoted in index points, with a multiplier of EUR25.
The last trading day is the third Friday of the expiration month. In the event
of the third Friday not being a business day, the last trading day is the last
business day prior such Friday. The expiration day is the business day fol-
lowing the last trading day. The next, the second, and the third quarter-end
months (March, June, September, December) are available for trading. DAX
options are cash-settled.
2.2 Sampling procedure
The sampling procedure partly follows that used by Christensen and Prab-
hala (1998). Our data consists of monthly observations with no overlaps, so as
to avoid drawbacks such as high residual auto-correlation, as well as imprecise
or biased regression estimates. In order to limit any excess volatility resulting
from the opening of new contracts and obtain consistent and accurate data,
we record the closing characteristics of both call and put options on the S&P
500, FTSE 100, and DAX indices between 2004 and 2010 that are:
1. the nearest-to-the-money, the least deep out-of-the-money, and the sec-
ond least deep out-of-the-money;
2. traded on the Wednesday following the last trading day of a given month,
or on the following business day when such Wednesday happens to not
be a trading day;
3. traded with significant volume;
4. expiring the coming month.
This procedure gives 82 monthly observations for the S&P 500, 73 monthly
observations for the FTSE 100, and 82 monthly observations for the DAX. The
corresponding historical volatility and ex-post future volatility over the re-
maining life of the options are computed according to the procedures detailed
below. The data is subsequently divided into two distinct samples - 2004-2007
and 2008-2010 - in order to examine the relationship between implied volatil-
ity and future volatility with and without the potentially-distorting impact of
the financial crisis of 2008. We thus end with 47 observations in 2004-2007
and 35 in 2008-2010 for both the S&P 500 and the DAX indices, and with 38
in 2004-2007 and 35 in 2008-2010 for the FTSE 100 index.

2.3 Variable definitions

2.3.1 Implied volatility
For any call option price Ct observed at time t, the implied volatility 0cri,t is
computed by numerically solving the Black-Scholes call option pricing formula,
assuming no dividends:

Ct = StN(di,t) - Kte~-rtN(d2,t) (1)

where
w ln(S/K) + (r + lo,t) (2
di,t = 2 e(2)

d2,t= di, - oic,tV (3)

For any close put option price Pt observed at time t, the implied volatility
aip,t is computed by numerically solving the Black-Scholes put option pricing
formula, assuming no dividends:

Pt = Kte-rftN(-d 2,t) - StN(-di,t) (4)

where
= ln(S/K) + (rf,t + jo;,t (5)

d2,t = di,t - utr~ t (6)

In the above equations, Kt denotes the strike price of the option, St the spot
price of the underlying index, -rt the time to expiration, rf,t the risk-free in-
terest rate, and N(.) stands for the standard normal cumulative distribution
function. The risk-free rate is the London inter-bank borrowing rate in the
Eurodollar market (LIBOR). The time to expiration is generally around 24
calendar days. The data has been provided by IVolatility.

On each observation date t, the final implied volatility uj,t is computed as

a weighted average of the three call and three put option implied volatilities
recorded:

ait = 0.2(o-ci,t + uipl,t) + 0. 2 (cTj

0 2,t + Cip2,t) + 0.1(cic3,t + uip3,t) (7)

where ajci,t, 0 ipi,t, aic2,t, gip2,t, ucac,t, and uip3,t denote, respectively, the implied
volatility at time t of the nearest-to-the-money call, the nearest-to-the-money
put, the least deep out-of-the-money call, the least deep out-of-the-money put,
the second least deep out-of-the-money call, and the second least deep out-
of-the-money put. Such a weighted average is both simple to implement and
more likely than a single implied volatility to attenuate the noise that might
impact an observation.

2.3.2 Historical volatility

Merton (1980) has shown that the accuracy of an estimate of volatility us-
ing past volatility increases with the sampling frequency within a given overall
observation period. We thus choose to use daily data. Furthermore, we de-
cide to compute both a simple average and an exponentially-weighted (EW)
average of historical volatility. Assuming that volatility varies with time, the
EW version compensates to some extent for one of the shortcomings of simple
historical volatility, namely the inevitable lag in information incorporation, by
giving greater weight to the most recent daily observations, and thus to more
recent information known to the market. Usually, EW historical volatility is
used to forecast one or a few days ahead, using high frequency data. How-
ever, we still would like to assess how it behaves as an estimator of one-month
volatility.
Simple average

For each option price observation at time t, we measure the historical volatility
over the past rt trading days, including the day of the observation, where rt is
the number of days until the expiration of the option, by the sample standard
deviation of the daily index returns during that period. We decide to omit the
usual estimator of the mean to avoid excessive noise, and use the following
formula:
O'h,t =
252
252 t 72) (8)
7-titt

where R, denotes the log-return on day i. Let Si be the index level on the
same day i, we have:

R, = ln(Si/Si_1) (9)
where in denotes the natural logarithm.

Exponentially-weighted average

For each option price observation at time t, we also measure the historical
volatility using an exponentially-weigthed average of past daily volatility, in-
cluding the day of the observation, with a decay factor of 0.94. This method
of computation is inspired by that developed by Riskmetrics. We use the
formula:
t
Che,t = 252 * 0.06 * 00.94"(Ri) 2 (10)

where 0.94 is the decay factor, 0.06 the sum of the weights, and where R-
denotes the log-return on day i according to (9).

2.3.3 Ex-post future volatility

For each time-t option price observation, we measure the ex-post future
volatility by the sample standard deviation of the daily index returns over the
remaining life -rt of the option. Again, we deliberately omit the estimator of
the mean, which would have been too noise-sensitive. We use the following
formula:
252 t+Tt
=,t -
2
E (R) (ii)
i=t+1

where R, denotes the log-return on day i according to (9).

Finally, to improve the normality of our variables, we apply a logarithmic

transformation to our observations of implied volatility, historical volatility,
and ex-post future volatility. Namely, for a given observation of volatility
with value o-, we use the value in(a) instead, where in denotes the natural
logarithm. The ordinary least squares (OLS) regressions are run with these
logvalues.

2.4 Regressions
To examine the information content of implied volatility and historical volatil-
ity before and after the financial financial crisis of 2008, we run the following
regressions for each index, first in 2004-2007 and then in 2008-2010:

up = a + #hcOh,t + Et (12)

Of,t = a + #heUhe,t + Et (13)

of,t = a + #ivoivt + Et (14)
3
Of,t= a + hrh,t + ivOiv,t + Et (15)
0 3
f,t = a + heahe,t + #+ivt - Et (16)

These regressions allow us to test three relevant hypotheses (Christensen

and Prabhala (1998), Christensen and Hansen (2002)). First, if implied volatil-
ity contains some information about future realized volatility, then the coeffi-
cient of implied volatility #i, should be nonzero in equations (14), (15), and
(16). Second, if implied volatility is an unbiased estimator, 3, should be equal
to 1 and the intercept a to 0 in equations (14),(15), and (16).

Finally, if implied volatility is an informationally-efficient predictor of future

realized, it should subsume any information contained in historical volatility,
and thus the coefficients #h in equation (15) and #he (16) should both be equal
to 0 and the adjusted R-squared of regressions (15) and (16) should not be
higher than that of regression (14). Based on a number of results and previous
research, we expect historical volatility to contain some information about fu-
ture volatility, and thus #h and #he to be nonzero and statistically significant
in equations (12) and (13). We will also check whether exponentially-weighted
estimates of standard deviation are better than simple estimates.

3 Descriptive statistics
Tables 1, 2, and 3 display relevant statistics describing the data on ex-post
future volatility, simple historical volatility, exponentially-weighted volatility,
and implied volatility.

First of all, we observe that the magnitude of the means and standard devia-
tions of ex-post future, simple, exponentially-weighted, and implied volatilities
across the three markets are somewhat equivalent, and since volatility is not
rising or decreasing uniformly throughout the sample, we would not expect
otherwise. For instance, in 2004-2007, the means of ex-post future volatility
are 0.127 for the S&P 500, 0.126 for the FTSE 100, and slightly higher for the
DAX, with a value of 0.147, while the standard deviations are 0.044, 0.065,
and 0.053 respectively. Besides, as demonstrated by figure 2, we can see that
simple and EW historical volatility differ only slightly.

Second, we notice that between 2004-2007 and 2008-2010, both the means
and standard deviations of the four measured volatilities rise sharply. For ex-
ample, the mean of the simple historical volatility of the S&P 500 increases
from 0.112 to 0.259, that of the FTSE 100 from 0.116 to 0.241, and that of
the DAX from 0.142 to 0.268, while the standard deviations rise by an even
greater factor: from 0.037 to 0.259, from 0.053 to 0.128, and from 0.042 to
0.145 respectively.

Third, in the cases of the S&P 500 and the DAX, the cross-correlations between
historical and implied volatility increase substantially in 2008-2010, virtually
reaching a perfect correlation of 1. For instance, the correlation between sim-
ple historical and implied volatility rises from 0.78 to 0.95 in the case of the
SP, and from 0.70 to 0.92 in the case of the DAX, while it only rises from 0.86
to 0.88 in the case of the FTSE.

Fourth, while implied volatility is slightly greater than the three other volatil-
ities across the three markets in 2004-2007, it stops being so in 2008-2010.
However, in both periods, its standard deviations are always lower than those
of ex-post future volatility, simple historical volatility, and exponential histor-
ical volatility.

Finally, the Jarque-Bera statistics demonstrate the effectiveness of applying a

logarithmic scale to normalize a set of data. As an example, the Jarque-Bera
statistic of the DAX ex-post future volatility in 2004-2007 falls from a huge
140.33 to an acceptable level of 5.38.

0.9
0.8
0.7
0.6
0.5
04
0.3
0.2

0
Jan-04 May-OS Oct-06 Feb-08 Jul-09 Nov40

- Simple historical volatility - EW historical volatility

Figure 1: S&P 500: simple and exponential historical volatility from 2004 to 2010
Table 1: Data based on 47 observations of the S&P 500 index and its options from January
2004 to December 2007 and 35 observations from January 2008 to December 2010. Values have
been annualized.

2004-2007
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.127 0.112 0.113 0.125
StDev 0.044 0.037 0.035 0.035
Skewness 1.50 1.29 1.33 1.62
Kurtosis 4.79 4.52 4.69 6.44
Jarque-Bera 23.85 17.58 19.50 43.64

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -2.207 -2.239 -2.225 -2.111
StDev 0.332 0.305 0.284 0.248
Skewness 0.85 0.55 0.59 0.86
Kurtosis 3.03 2.98 3.19 3.60
Jarque-Bera 5.61 2.27 2.78 6.52

2008-2010
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.256 0.259 0.265 0.256
StDev 0.164 0.162 0.160 0.112
Skewness 1.94 1.79 1.76 1.63
Kurtosis 6.93 6.11 5.76 5.76
Jarque-Bera 44.46 32.78 29.23 26.63

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -1.514 -1.496 -1.463 -1.437
StDev 0.536 0.527 0.499 0.377
Skewness 0.59 0.61 0.83 0.80
Kurtosis 3.23 2.97 3.02 3.07
Jarque-Bera 2.09 2.18 3.98 3.78

Correlation matrix (log of o)

2004-2007 2008-2010
O'h,t O'he,t O'io,t O'h,t O'he,t Oliv,t
Oh,t 1 0.95 0.78 1 0.98 0.95
ahe,t 0.95 1 0.82 0.98 1 0.97
O'iv,t 0.78 0.82 1 0.95 0.97 1
Table 2: Data based on 38 observations of the FTSE 100 index and its options from November
2004 to December 2007 and 35 observations from January 2008 to December 2010. Values have
been annualized.

2004-2007
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.126 0.116 0.120 0.125
StDev 0.065 0.053 0.054 0.047
Skewness 1.76 1.68 1.60 1.85
Kurtosis 5.72 5.27 5.04 6.63
Jarque-Bera 31.41 26.07 22.80 42.46

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -2.173 -2.233 -2.201 -2.130
StDev 0.423 0.382 0.380 0.314
Skewness 1.06 1.06 1.04 1.14
Kurtosis 3.18 3.28 3.00 3.74
Jarque-Bera 7.18 7.22 6.83 9.07

2008-2010
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.240 0.241 0.249 0.249
StDev 0.132 0.128 0.130 0.101
Skewness 2.25 2.06 2.02 1.68
Kurtosis 9.34 8.10 7.68 5.60
Jarque-Bera 88.23 62.81 55.63 26.43

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -1.532 -1.527 -1.492 -1.454
StDev 0.441 0.440 0.433 0.344
Skewness 0.87 0.77 0.81 0.99
Kurtosis 3.73 3.53 3.47 3.34
Jarque-Bera 5.23 3.84 4.20 5.93

Correlation matrix (log of a)

2004-2007 2008-2010
O'h,t O'he,t Olio,t O'h,t ahe,t Oliv,t
'h,t 1 0.97 0.86 1 0.98 0.88
O'he,t 0.97 1 0.92 0.98 1 0.93
civ,t 0.86 0.92 1 0.88 0.93 1
Table 3: Data based on 47 observations of the DAX index and its options from January 2004
to December 2007 and 35 observations from 2008 to 2010. Values have been annualized.

2004-2007
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.147 0.142 0.143 0.157
StDev 0.053 0.042 0.036 0.036
Skewness 2.10 0.81 0.67 0.83
Kurtosis 10.34 3.07 2.65 3.30
Jarque-Bera 140.33 5.18 3.80 5.62

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -1.967 -1.995 -1.978 -1.872
StDev 0.315 0.288 0.249 0.218
Skewness 0.68 0.23 0.22 0.37
Kurtosis 3.98 2.47 2.33 2.55
Jarque-Bera 5.38 0.98 1.28 1.45

2008-2010
Statistics Realized Historical Historical Implied
volatility volatility (SA) volatility (EWA) volatility
Mean 0.262 0.268 0.273 0.264
StDev 0.140 0.145 0.142 0.098
Skewness 2.07 1.64 1.76 1.44
Kurtosis 7.77 5.59 6.14 4.41
Jarque-Bera 58.15 25.54 32.38 14.97

Statistics Log realized Log historical Log historical Log implied

volatility volatility (SA) volatility (EWA) volatility
Mean -1.442 -1.429 -1.402 -1.388
StDev 0.433 0.465 0.439 0.325
Skewness 0.92 0.66 0.76 0.88
Kurtosis 3.60 2.80 3.05 2.98
Jarque-Bera 5.49 2.59 3.38 4.53

Correlation matrix (log of o)

2004-2007 2008-2010
0
Oh,t he,t civ,t Oh,t Uhe,t Oiv,t
Oh,t7
1 0.95 0.70 1 0.99 0.92
Ohe,t 0.95 1 0.80 0.99 1 0.94
uiv,t 0.70 0.80 1 0.92 0.94 1
 0.8
0.7
S&P 500
0.6
0.5
0.4
0.3
0.2

0.1
0
Jan-04 May-05 Oct-06 Feb-08 Jul-09 Nov-10

0.8
0.7
FTSE 100
0.6

0.5
0.4

0.3
0.2
0.1

Nov-04 Mar-06 Aug-07 Dec-08 May-10

0.8
0.7
DAX
0.6
0.5
0.4

0.3
0.2

0.1

Jan-04 May-OS Oct-06 Feb-08 Jul-09 Nov-10

implied volatility - Historical volatility

Figure 2: S&P, FTSE, and DAX: simple historical volatility and implied volatility from 2004
to 2010
4 Empirical Results
4.1 S&P 500
The ordinary least squares (OLS) coefficient estimates for the S&P 500 are
displayed in Table 4.

Let us first focus on 2004-2007. In our first regression, we observe that

the coefficient estimate for simple historical volatility has a value of 0.567,
which, with a t-stat of 4.10, is statistically significant. The adjusted R-
squared indicates that, when considered alone, simple volatility accounts for
25.57% of the variance of future volatility. Our second regression shows that
exponentially-weighted historical volatility performs negligibly better. If the
equally-significant coefficient estimate has a higher value of 0.633, we obtain
an R-squared of 0.2765, which is virtually equal to that of our previous re-
gression. However, we observe in our third regression that implied volatility
outperforms both simple and exponentially-weighted volatility. Not only is
the coefficient estimate greater, with a value of 0.943, and more significant
statistically with at t-stat of 6.67, but, above all, the R-squared indicates that
implied volatility explains nearly 50% of the variance of future volatility. What
is more, in contrast with our two first regressions, we get an intercept that is
not statistically significant. These results suggest that while the information
content of simple and EW historical volatility are relatively equivalent, that
of implied volatility is substantially greater.

Equations (15) and (16) enable us to assess whether the information con-
tent of implied volatility subsumes those of our two other estimators. The
first striking element is that the R-squared of both regressions do not exceed
that of regression (14). Secondly, the coefficient estimates of both simple and
EW historical volatility are both not statistically significant at the commonly-
accepted level. On the other hand, the coefficient estimates of implied volatil-
ity are very close to one and statistically significant, with t-stats greater than 4.

Our analysis of the period 2004-2007 suggests that implied volatility does
contain some significant amount information about future volatility, outper-
forms both simple and EW historical volatility in that regard, and subsumes
the information contained in the other two estimators. Besides, with coeffi-
cient estimates very close to 1, it appears to be a nearly unbiased estimator
of future volatility.

If we move on to the period 2008-2010, we obtain very different results. First of

all, the R-Squared of all six regressions, with values within the 0.51-0.56 range,
are virtually equal to each other and are also higher than any single R-squared
obtained in our analysis of 2004-2007. Of the three estimators, it seems that
none outperforms the others. In regressions (12), (13), and (14), the coeffi-
cient estimates of simple volatility, EW volatility, and implied volatility are
all statistically significant, with values of 0.739, 0.796, and 1.070 respectively,
consistent with the hierarchy found in 2004-2007. However, in our last two
regressions, none of the coefficient estimates manages to have a t-stat greater
than 1.96. A plausible explanation is that the estimators "neutralized" each
other because of their higher correlation, as seen in Table 1.

It appears that in 2008-2010, the three estimators show equivalent perfor-

mance in predicting future volatility, that this performance is greater than in
2004-2007, and that there is no added value in combining these estimators in
terms of forecasting power.

Table 4: S&P500: conventional OLS coefficient estimates

2004-2007
Intercept Oh,t Ohe,t Giv,t Adj. R 2 JB White Autocorr.
-0.938 0.567 - - 0.2557 2.36 2.89 0.05
(-3.00) (4.10)
-0.799 - 0.633 - 0.2765 2.64 0.74 0.06
(-2.43) (4.31)
-0.216 - - 0.943 0.4859 14.14 0.58 0.07
(-0.72) (6.67)
-0.233 -0.080 - 1.020 0.4764 15.19 1.24 0.10
(-0.76) (-0.43) (4.47)
-0.251 - -0.145 1.079 0.4794 13.94 2.44 0.11
(-0.81) (-0.66) (4.30)

2008-2010
Intercept Oh,t ahe,t o*i,t Adj. R 2 JB White Autocorr.
-0.408 0.739 - - 0.5149 5.94 1.08 0.17
(-2.12) (6.09)
-0.350 - 0.796 - 0.5356 7.78 0.94 0.23
(-1.81) (6.34)
0.024 - - 1.070 0.5547 21.20 3.01 0.26
(0.10) (6.58)
-0.026 0.123 - 0.908 0.5424 19.35 5.49 0.24
(-0.09) (0.33) (1.73)
-0.054 - 0.210 0.802 0.5435 18.55 5.27 0.25
(-0.18) (0.43) (1.25)

4.2 FTSE 100

Table 5 displays the OLS coefficient estimates for the FTSE 100.

We start our analysis by studying 2004-2007. The coefficient estimates of

simple and exponentially-weighted historical volatility are both statistically
significant, with values of 0.688 (t-stat of 4.75) and 0.691 (t-stat of 4.73) re-
spectively. The R-squared of the first regression is 0.3684 and that of the
second is 0.3661. It thus appears that there is no palpable difference in the
predictive power of simple and EW volatility. On the other hand, in our
third regression, we obtain a statistically-significant (t-stat of 5.96) coefficient
estimate of 0.950 for implied volatility, an insignificant intercept, and an R-
squared of 0.4828. Our results tend again to support the idea that implied
volatility is a better estimator than historical volatility.

Let us now consider regressions (15) and (16). First and foremost, we ob-
serve that their R-squareds (0.4689 and 0.4751 respectively) are equivalent
to that of regression (14), suggesting that combining historical and implied
volatility does not produce better results in estimating future volatility than
implied volatility alone. Secondly, the only coefficient estimates that have sta-
tistical significance are those of implied volatility, with a value of 0.883 and a
t-stat of 2.79 in regression (15), and a value of 1.215 and a t-stat of 2.91 in
regression (16).

Our analysis of 2004-2007 shows that implied volatility is a better estimator

than historical volatility, and that it incorporates the information content of
the other two estimators. Furthermore, combining implied volatility with his-
torical volatility does not outperform the predictive power of implied volatility
taken alone.

Moving on to 2008-2010, we observe that all five regressions account for the
variance of future volatility to an equivalent extent, with R-squared of around
0.465. Only simple historical volatility slightly lags behind, with an R-squared
of 0.4150 for regression (12). When considering each estimator separately, we
get statistically-significant coefficient estimates of 0.658 for simple volatility,
0.707 for EW volatility, and 0.892 for implied volatility, with t-stats at the 5%
significance level. Out of these three regressions, only regression (14) gives an
insignificant intercept. In regressions (15) and (16), no coefficient estimates,
except (barely) for that of simple historical volatility (t-stat of 1.97) in regres-
sion (15), is statistically significant.

Our results for 2008-2010 suggest that the performances of the three esti-
mators are relatively equivalent, and that combining historical and implied
volatility does not improve the overall predictive power.

4.3 DAX
The OLS coefficient estimates for the DAX are displayed in Table 6.

Again, let us first consider 2004-2007. Surprisingly, and contrary to the re-
sults of our studies of the S&P 500 and the FTSE 100, we find not only that
the R-squared of all five regressions are equivalent, but that they are also of
relatively low level (around 10%). For instance, while implied volatility alone
accounts for as much as 48.59% of the variance of future volatility in the case
of the S&P 500 and for 48.28% in that of the FTSE 100, here, it only ac-
counts for 10.15%. In our first three regressions, the coefficient estimates of
simple, exponentially-weighted, and implied volatility, with values of 0.391,
0.444, and 0.502 respectively, are also much lower. Moreover, regressions (15)
 Table 5: FTSE 100: conventional OLS coefficient estimates

2004-2007
Intercept OUh,t oihe,t O:v,t Adj. R 2 JB White Autocorr.
-0.637 0.688 - - 0.3684 7.48 6.19 0.07
(-1.94) (4.75)
-0.653 - 0.691 - 0.3661 10.44 3.27 0.08
(-2.00) (4.73)
-0.150 - - 0.950 0.4828 17.88 1.50 -0.06
(-0.44) (5.96)
-0.150 0.064 - 0.883 0.4689 17.09 2.51 -0.06
(-0.43) (0.25) (2.79)
-0.109 - -0.237 1.215 0.4751 17.88 1.54 -0.08
(-0.31) (-0.69) (2.91)

2008-2010
Intercept Och,t Ohe,t av,t Adj. R 2 JB White Autocorr.
-0.527 0.658 - - 0.4150 4.49 0.17 0.13
(-2.52) (5.01)
-0.476 - 0.707 - 0.4670 7.24 0.01 0.18
(-2.40) (5.55)
-0.235 - - 0.892 0.4690 13.73 0.33 0.30
(-0.98) (5.57)
-0.256 0.201 - 0.666 0.4621 11.23 3.97 0.22
(-1.06) (1.97) (0.76)
-0.309 - 0.357 0.474 0.4702 11.10 2.91 0.23
(-1.24) (1.09) (1.04)
and (16) give a statistically-significant coefficient estimate for none of the three
estimators. We also note that the intercept is significant in all five regressions.

Intriguingly enough, our analysis of 2004-2007 indicates that no estimator out-

performs the other two, that a combination of historical and implied volatility
does not generate greater performance, and that the overall predictive power
is comparatively poor.

Now, if we move to 2008-2010, our results are consistent with what we find for
the other markets. First of all, the adjusted R-squareds of all five regressions
are equivalent, with a level slightly higher than 0.50. In regressions (12), (13),
and (14), the coefficient estimates of simple, exponential, and implied volatil-
ity are statistically significant, with a t-stat of approximately 6, and their
magnitude, respectively 0.674, 0.729, and 0.966, are in line with our findings
in the other two markets for 2008-2010. Again, our two last regressions do
not result in any significant coefficient estimate for either historical or implied
volatility.

In 2008-2010, we find that the three estimators show equivalently-high pre-

dictive power, and that combining historical and implied volatility does not
entail better performance in that regard.

4.4 Residual diagnostics

Tables 4, 5 and 6 display the Jarque-Bera statistic (JB), the White statistic
(White), and the first-order residual autocorrelation (Autocorr.) for each re-
gression.

The Jarque-Bera test is a test of normality, based on kurtosis and skewness,

that we conduct on the residuals of each regression. The test statistic has an
asymptotic chi-square distribution with two degrees of freedom. This means
that with a level of significance of 5%, we can reject the null hypothesis that
the residuals are normally distributed if the JB statistic is above 5.99. As we
can see, in most regressions, the hypothesis is rejected by a substantial margin.

The White test is a test of homoskedasticity, i.e. whether the variance of

the residuals is constant. The statistic has a chi-square distribution with k - 1
degrees of freedom, where k is equal to the number of regressors excluding
the intercept. In the case of the number of regressors being equal to one,
the statistic is computed so as to still have a chi-square distribution with one
degree of freedom. At a level of significance set a 5%, we can the reject the
null hypothesis of homoskedasticity when the White statistic is above 3.84.
We can see that the hypothesis of hornoskedasticity of the residuals is rejected
in only a few instances: in regressions (15) and (16) in 2008-2010 in the case
of the S&P 500, and in regressions (12) in 2004-2007 and (15) in 2008-2010 in
the case of the FTSE 100.

We also computed the first order autocorrelation of the residuals of each re-
 Table 6: DAX: conventional OLS coefficient estimates

2004-2007
Intercept Ost Ohe,t uiv,t Adj. R 2 JB White Autocorr.
-1.187 0.391 - - 0.1088 7.20 1.41 0.08
(-3.87) (2.57)
-1.090 - 0.444 - 0.1038 9.09 1.38 0.09
(-3.10) (2.51)
-1.028 - - 0.502 0.1015 18.19 2.40 0.10
(-2.71) (2.49)
-0.963 0.245 - 0.275 0.1079 13.06 3.53 0.04
(-2.51) (0.97) (1.15)
-0.960 - 0.255 0.268 0.0964 14.10 3.62 0.06
(-2.47) (0.86) (0.79)

2008-2010
Intercept OYh,t Ohe,t uiv,t Adj. R 2 JB White Autocorr.
-0.479 0.674 - - 0.5092 19.93 0.10 0.20
(-2.85) (6.02)
-0.420 - 0.729 - 0.5330 23.77 0.39 0.27
(-2.48) (6.31)
-0.101 - - 0.966 0.5133 26.66 0.68 0.23
(-0.44) (6.07)
-0.231 0.337 - 0.525 0.5205 29.52 2.00 0.20
(-0.92) (1.22) (1.33)
-0.231 - 0.337 0.525 0.5205 28.63 0.93 0.24
(-0.93) (1.22) (1.33)
gression. Across the three markets, we observe that autocorrelation tends to
be very low in 2004-2007, with a range of [-0.08;0.11], but rises sharply in
2008-2010, with a range of [0.13;0.30], while remaining at a tolerable level.

4.5 Intermediary conclusion

As regards which of simple, exponentially-weighted and implied volatility con-
stitutes the best estimator of future volatility, the results of our analysis seem
quite compelling: implied volatility is always at least as good as the other
two, and often better. In 2004-2007, the superior informational content of
implied volatility is unequivocal. Indeed, not only does implied volatility
partly explain the variance of future volatility, but it does so better than past
historical volatility and incorporates the information content of the latter.
Concerning 2008-2010, it is reasonable to hypothesize that our results were
impacted by the violent turbulences of the financial crisis. In that scenario,
the three estimators appear somewhat redundant. Nevertheless, these results
also demonstrate that the superiority, in a large sense, of implied volatility is
robust, for implied was at least as good as the other estimators. Our results
suggest that implied volatility is a robust and powerful estimator of future
volatility, notably when compared with other usual measures such as simple
and exponentially-weighted historical volatility.

5 Introducing GARCH
5.1 The GJR-GARCH(1,1) model
The Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model
extends the ARCH model developed by Nobel Prize laureate R. Engle in 1982
as a method of analyzing time-varying volatility. The model considers that
future variances depend upon past variances and squared returns; variance
is both conditional and autoregressive. It is widely used among practitioners
and academic alike to model and forecast volatility. Developed by Glosten,
Jagannathan, and Runkle in 1993, GJR-GARCH is an extension that takes
into account asymmetries in the response of the conditional variance to an
innovation. Focusing on the S&P 500, we decide to use a GJR-GARCH(1,1)
model and compare its predictive power to that of implied volatility using two
different methods. The model is the following:

o K + out-1 + ac2_1 + #q- 1 :t_1 (17)

where
It_1 = 0 if Et > 0, (18)
It- 1 = 1 if Et < 0 , (19)
Ct = ZtUt , (20)
zt is a sequence of independent and identically distributed variables, and 0 g,t
denotes the GJR-GARCH conditional volatility at time t.

In our first method, we use the data on the S&P 500 from 2004 to 2010
to estimate the parameters of our model by maximum likelihood, before doing
an in-sample comparison of GARCH-fitted volatility and implied volatility. In
our second method, on each date t when we observe an option from 2005 to
2010, we use the daily volatility data from 2004 up to t to calibrate the GJR-
GARCH(1,1) model and forecast the market volatility over the remaining life
of the option. We then compare the predictive power of these forecasts with
that of implied volatility in 2005-2007 and 2008-2010.

5.2 In-sample GARCH-fitted values v. implied volatil-

ity
5.2.1 Methodology
We collect monthly observations of the return of the S&P 500 from 2004 to
2010 on the same dates t we observed historical and implied volatility (see
2.2). At time t, the return is:

Rt = ln(St/St__) (21)

where St denotes the level of the S&P 500 at time t and rt the days remaining
before the expiration of the option whose implied volatility was observed at
time t.

Then, we use our collection of 82 returns to estimate the parameters of our

GJR-GARCH(1,1) model through maximum likelihood and generate 82 fit-
ted values of conditional volatility. The parameter estimates are presented in
Table 7, while the GARCH-fitted values of volatility are plotted in in Figure
3.

Table 7: GJR-GARCH(1,1) parameter estimates

Parameter Value T-stat

K 0.0012972 3.7158
6 0 0
a 0.23906 0.7259
# 1 2.1967

We run the following regressions over 2004-2010 to assess the respective in-
formation contents of GJR-GARCH modeled volatility and implied volatility:

Of = a + #goUt + Et (22)

Of't = a + #iveivt + Et (23)

Up = a + #gog,t + #ivO'it + Et (24)

 1.2

0.8

0.6

0.4

0.2

0
Jan-04 May-05 Oct-06 Feb-08 Jul-09 Nov-10

- GARCH-fitted volatility

Figure 3: GARCH-fitted volatility

Since the parameters of our GJR-GARCH(1,1) model are estimated using

the very same sample data, there is definitely a look-ahead bias in favor of
GJR-GARCH volatility. However, considering our results, we deem the short-
coming acceptable.

5.2.2 Empirical results

The results of the regressions are displayed in Table 7. Despite the look-
ahead bias entailed by the the in-sample comparison, we observe that the
does not perform very well, certainly because of the too-small amount of data
used for calibration. In regression 21, we obtain a statistically-significant (t-
stat of 4.11) coefficient of 0.599 for fitted volatility, as well as a statistically-
significant intercept. The R-squared indicates that the volatility inferred from
our GARCH model only accounts for 16.38% of the variance of future volatility,
while implied volatility alone accounts for over 71.86%. Moreover, when con-
sidering fitted and implied volatility together in regression 23, the R-squared
of 71.60% shows that fitted volatility does not add any information content
to that already included in implied volatility. We also observe that the co-
efficient estimate of implied volatility is virtually equal to 1, with a t-stat of
12.51, while both that of fitted volatility and the intercept are very close to 0
and not statistically significant.

Our results suggest that, using this method, volatility fitted with a GJR-
GARCH(1,1) model constitutes a relatively poor estimator, even with a look-
ahead bias, and that its information content is subsumed by implied volatility,
which appears to be a nearly unbiased estimator.
Table 8: S&P500 2004-2010: OLS coefficient estimates of GARCH-fitted volatility and implied
volatility

Intercept Og,t Oiv,t Adj. R 2 JB White Autocorr.

-0.823 0.599 - 16.38% 17.89 1.12 0.60
(-3.04) (4.11)
-0.041 - 1.025 71.86% 50.71 0.93 0.17
(-0.31) (14.42)
0.014 0.049 1.006 71.60% 54.80 0.91 0.18
(0.08) (0.52) (12.51)

5.3 Out-of-sample GARCH forecasts v. implied volatil-

ity
5.3.1 Methodology
Generally, GARCH models rely on large amounts of high-frequency historical
data to forecast a few periods ahead. With our second method, we endeavour
to exploit the predictive power of GJR-GARCH(1,1) in a much more efficient
way.

From 2006 to 2010, on each date t we collect an observation of implied volatil-

ity, use the daily return data from January 2004 up to t to estimate the param-
eters of our GJR-GARCH(1,1) model, and then forecast the daily volatility of
the next -rt days, where -r denotes the number of days before the expiration of
the option whose implied volatility we collected on day t. We then compute
and annualize the average daily volatility over these rt days:

gt Ugd,i (25)
Tt i=t+1

By doing so, we are able to avoid the flaws that come with a GJR-GARCH
model fed with only a small amount of data.

We subsequently run the following regressions, first on 2005-2007 and then

on 2008-2010, to compare the predictive powers of simple historical volatility,
GARCH forecasted volatility, and implied volatility:

oft = a + fgog,t + ft (26)

3
Uf,t = + #iv li,t + et (27)

of,t = a + #h Jh,t + #9Ug,t + Et (28)

aft =a + #go9g,t + #iVOiV't + Ct (29)
 0.8
0.7
0.6
0.5
04
0.3
0.2
0.1
0
Jan-05 May-06 Oct-07 Feb-09 Jul.10

- GARCH-forecasted volatility

Figure 4: GARCH-forecasted volatility

5.3.2 Empirical results

The results, displayed in Table 8, confirm to some extent what we found be-
fore introducing GARCH. We can see that, with a large amount of daily data,
the model perform quite well when considered alone. The coefficient estimate
of GARCH-forecasted volatility has a statistically-significant value of 1.486 in
2005-2007 and of 0.893 in 2008-2010, while the t-stats of the intercepts fail to
reach 1.96. With an R-squared of 40.60 in the first period and of 58.99% in
the second, GARCH explains a substantial portion of the variance of future
volatility. Furthermore, when combining GARCH-forecasted volatility and
simple historical volatility, not only does the R-squared fail to increase above
those when GARCH is taken alone, but also the coefficient estimates of his-
torical volatility are insignificant from a statistical perspective. On the other,
regression (26) gives us coefficient estimates of 1.607, with a t-stat of 2.57,
in 2005-2007 and of 1.044, with a t-stat of 2.45, in 2008-2010 for GARCH-
forecasted volatility.

However, in 2005-2007, we observe that regression (26) gives us an R-squared

which is higher than that from regression (25), indicating that implied volatil-
ity accounts for a greater part of future volatility than GARCH-forecasted
volatility. In addition, when combining GARCH and implied volatility, we
observe that the R-squared is equivalent to that of regression (26), and that
only implied volatility has a coefficient estimate that is statistically significant.
In 2008-2010, as we found the fourth part of the present thesis, all measure
volatilities seem to perform equally well and to be redundant from an infor-
mational perspective.

The results of our second method suggest that, although GJR-GARCH out-
performs historical volatility, it is still inferior to implied volatility in terms
of forecasting power and does not seem to contain supplementary information.
Table 9: S&P 500: OLS estimates of GARCH forecasted volatility and implied volatility for
2005-2007 and 2008-2010

2005-2007
Intercept O'h,t cog,e olivt Adj. R 2 JB White Autocorr.
1.108 - 1.486 - 40.60% 4.29 0.31 -0.02
(1.67) (4.99)
0.142 - - 1.093 58.77% 15.96 0.55 -0.07
(0.43) (7.13)
1.228 -0.066 1.607 - 38.88% 4.76 7.02 0.00
(1.42) (-0.22) (2.57)
-0.185 - -0.308 1.260 57.95% 13.64 3.72 -0.04
(-0.28) (-0.59) (3.88)

2008-2010
Intercept OUh,t g,t oiv't Adj. R 2 JB White Autocorr.
-0.102 - 0.893 - 58.99% 9.11 0.48 0.19
(-0.49) (7.06)
0.024 - - 1.070 55.48% 21.20 0.91 0.26
(0.10) (6.59)
-0.0727 -0.139 1.044 57.89% 9.12 1.01 0.20
(-0.32) (-0.37) (2.45)
-0.096 - 0.869 0.031 57.71% 9.47 17.73 0.19
(-0.39) (1.66) (0.05)

6 General conclusion
Our analysis indicates that implied volatility is an informationally-efficient and
largely-unbiased estimator of one-month-ahead future volatility. It also shows
that implied volatility is superior to simple historical volatility, exponential
historical volatility, and GJR-GARCH(1,1)-fitted or -forecasted volatility. In
2004-2007, implied volatility unequivocally outperformed its competitors in
forecasting the one-month volatility of the S&P 500 and the DAX, while in
2004-2007 for the DAX and in 2008-2010 across all three markets, it performed
equally well. Above all, in every single instance, combining implied volatility
with any of the other candidate estimators did not yield a higher R-squared
than when implied volatility was used as a lone forecaster, suggesting that im-
plied volatility subsumes the information content of all the other estimators.
As in Christensen and Prabahala (1998), we can only conclude that there is
undeniable evidence of the superior efficiency of implied volatility as a predic-
tor of future volatility one month ahead. We have also shown that the result
has been robust across different markets since 2004.

What is more, in light of our results for the 2008-2010 period, we dare general-
ize our findings, by formulating the hypothesis that, in times of turbulence and
exceptional market volatility, all major estimators share the same information
content and forecasting performance regarding one-month future volatility.
We leave this statement open for further investigation.

Of course, if the present study confirms the results of recent research, such
as Christensen and Prahbala (1998), it also paves the way for more studies.
First of all, it would be interesting to check the robustness of the conclusion
in a wide range of other markets, particularly with different levels of liquidity.
Secondly, it would be interesting to decrease the length of the forecast and see
if implied volatility keeps its advantage, notably over GARCH models, when
estimating volatility over shorter periods of time. Finally, comparing implied
volatility with more sophisticated historical volatility models would perhaps
yield different results.
References
[1] Beckers, S.: Standard deviation implied in option prices as predictors of
future stock price variability, Journal of Banking and Finance (1981)
[2] Black, F. & Scholes M.: The valuation of options and corporateliabilities,
Journal of Political Economy (1973)
[3] Bodie, Z. & Merton, R. C.: The informational role of asset prices, The
Global Financial System (1995)
[4] Bollerslev, T.: Generalized Autoregressive ConditionalHeteroskedasticity,
Journal of Econometrics (1986)
[5] Canina, L. & Figlewski, S.: The informational content of implied volatil-
ity, Review of Financial Studies (1993)
[6] Chiras, D. P. & Manaster, S.: The information content of prices and a
test of market efficiency, Journal of Financial Economics (1978)
[7] Christensen, B. J. & Prabhala, N. R.: The relation between implied and
realized volatility, Journal of Financial Economics (1998)
[8] Christensen, B. J. & Hansen, C. S.: New evidence on the implied-realized
volatility relation, The European Journal of Finance (2002)
[9] Day, T. E. & Lewis, C. M.: Stock market volatility and the information
content of stock index options, Journal of Financial Economics (1992)
[10] Engle, R.: Autoregressive conditional heteroskedasticitywith estimates of
the variance of UK inflation, Econometrica (1982)
[11] Glosten, L. R., Jagannathan, R. & Runkle, D. E.: On the relation between
the expected value and the volatility of the nominal excess return on stocks,
Journal of Finance (1993)
[12] Hansen, C. S.: The Relationship between implied and realized volatility in
the Danish option and equity markets (1999)
[13] Lamoureux, C. G. & Lastrapes, W.: Forecasting stock return variance:
understanding stochastic implied volatility, Review of Financial Studies
(1993)
[14] Latane, H. & Rendleman, R.: Standard deviation of stock price rations
implied in option prices, Journal of Finance (1976)
[15] Li, S. & Yang, Q.: The Relationship between implied and realized volatil-
ity: evidence from the Australian stock index option market, Review of
Quantitative Finance and Accounting (2009)
[16] McDonald, R. L.: Derivatives markets (2006)
[17] Merton, R. C.: Theory of rational option pricing, The Bell Journal of
Economics and Management Science (1973)
[18] Pan, J.: Time-series patterns in equity, Part 1 & 2, Teaching slides for
15.433 at MIT Sloan School of Management (2011)
[19] Szakmary, A. et al.: The predictive power of implied volatility: evidence
from 35 futures markets, Journal of Banking and Finance (2003)