On The Volatility of Volatility: Electronic Address: Hsu@duende - Uoregon.edu Electronic Address: Bmurray1@uoregon - Edu
On The Volatility of Volatility: Electronic Address: Hsu@duende - Uoregon.edu Electronic Address: Bmurray1@uoregon - Edu
Stephen D. H. Hsu
_
365
30
30
i=1
_
ln
_
S
t+i
S
t+i1
__
2
. (1)
As is common practice, we use a denition which assumes zero mean.
We use the standard denition of the correlation between two time series X
t
and Y
t
consisting
of n points x
i
and y
i
, respectively:
Cor (X
t
, Y
t
) =
n
i=1
(x
i
x)(y
i
y)
(n 1)s
x
s
y
, (2)
where x and y are the usual sample means, and s
x
and s
y
are the usual sample standard deviations.
We nd a signicant correlation between the VIX and the 30-day realized volatility:
Cor (VIX
t
, RVol
t,t+30
) = 0.76. (3)
Scatter plots of combinations of the VIX, the 30-day realized volatility, and the SPX are shown
in Figs. 1, 2, and 3. The correlations between the SPX and both realized and implied volatilities
are not signicant over long timescale of our sample (16 and a half years).
5
5
10
15
20
25
30
35
40
45
50
200 400 600 800 1000 1200 1400 1600
V
I
X
I
n
d
e
x
SPX Index
correlation = 0.28
FIG. 2: The VIX versus the SPX for 2 Jan 1990 to 29 Jun 2006. Over this long timescale, no signicant
correlation between the indexes is observed.
B. Changes in the SPX, the VIX, and realized volatility
In order to examine correlations between changes in the indexes, we make the following deni-
tions:
CSPX
t
=
SPX
t
SPX
t1
SPX
t1
CVIX
t
=
VIX
t
VIX
t1
VIX
t1
. (4)
There is a signicant negative correlation between changes in the VIX and changes in the SPX:
Cor (CVIX
t
, CSPX
t
) = 0.66. (5)
See Fig. 4 for a scatter plot of CVIX
t
versus CSPX
t
. Note that this negative correlation is inter-
esting, as it implies that there is a non-trivial memory in price dynamics which goes beyond the
most naive (i.e., log-normal) models.
In order to examine the correlation between changes in the VIX and changes in the 30-day
realized volatility, we must rst nd a suitable denition for the change in the 30-day realized
6
0
5
10
15
20
25
30
35
40
45
50
200 400 600 800 1000 1200 1400 1600
3
0
-
d
a
y
r
e
a
l
i
z
e
d
v
o
l
a
t
i
l
i
t
y
SPX Index
correlation = 0.37
FIG. 3: The 30-day realized volatility of the SPX versus the SPX for 2 Jan 1990 to 29 Jun 2006. Over this
long timescale, no signicant correlation is observed.
volatility. To this end, we dene:
CRVol
t
=
RVol
t,t+30
RVol
t31,t1
RVol
t31,t1
. (6)
This quantity compares the 30-day realized volatility of the 30-day time period ending at time
to time t 1 with that of the 30-day time period beginning at time t. It is an appropriate way
of measuring the change in the 30-day realized volatility because it compares volatilities of two
independent neighboring time periods. However, it remains true that CRVol
t
and CRVol
t+1
are
not independent. Therefore, in estimating the correlation between the change in the VIX and the
change in the 30-day realized volatility, it would not be appropriate to simply compare the two time
series CRVol
t
and CVIX
t
. Instead, we compare the correlation between CRVol
s+30t
and CVIX
s+30t
for all osets s with 0 < s < 30. All 21 such time series consist of 196 days (there are 21 trading
days for every 30 calendar days). See Fig. 5 for a scatter plot of CRVol
s+30t
versus CVIX
s+30t
for
the case s = 2, i.e., the rst change in the 30-day realized volatility considered is between the 30
days prior to and the 30 days following 5 Feb 1990. For this choice of oset, the correlation is
typical of the values obtained for all values of s, with a value of 0.13.
7
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06
C
h
a
n
g
e
i
n
V
I
X
Change in SPX
correlation = -0.66
FIG. 4: Change in the VIX versus change in the SPX for 2 Jan 1990 to 29 Jun 2006. Changes in the indexes
have a signicant level of negative correlation.
A histogram of the correlations for all possible values of oset s is shown in Fig. 6. Each of
the 21 correlations computed is correlating two time series (CVIX
t
and CRVol
t
) of 196 days each.
With a mean correlation of 0.11 and a standard deviation of the correlations of 0.06, it is clear
that a change in the VIX does not predict a change in the 30-day realized volatility of the SPX.
III. THE BEHAVIOR OF IMPLIED VOLATILITY
In order to try to determine the cause of the negative correlation between changes in the SPX
and changes in the VIX, we obtained closing prices for near-the-money SPX put and call options
on the days immediately before and the days of the four largest VIX increases and four largest
VIX decreases of 2005 and 2006. The data were obtained from Bloomberg.
In Figs. 7 and 8 we plot option price versus the ratio of strike price to index level for both puts
and calls the day before and the day of one of the largest VIX increases and decreases, respectively,
of 2006. As expected, both puts and calls at a given distance from at-the-money become more
expensive when VIX increases and less expensive when VIX decreases.
8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
-0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25
C
h
a
n
g
e
i
n
3
0
-
d
a
y
r
e
a
l
i
z
e
d
v
o
l
a
t
i
l
i
t
y
(
n
o
n
-
o
v
e
r
l
a
p
p
i
n
g
)
Change in VIX
correlation = 0.13
FIG. 5: Change in the 30-day realized volatility versus change in the VIX for non-overlapping intervals from
1 Feb 1990 to 29 Jun 2006. With a correlation of 0.13, this dataset (s = 2, see text) provides a typical
example of the lack of correlation between changes in the VIX and changes in the 30-day realized volatility.
See Fig. 6 for a histogram of the correlations for all possible choices of the oset s.
Similarly, the volatility surface is plotted in Figs. 9 and 10. Although we see that, as expected,
Black-Scholes implied volatility increases or decreases when the VIX does, it is dicult to comment
as to why the VIX is changing, i.e., which particular options are causing the VIX to increase or
decrease. Also, we see roughly linear volatility skews, as have been present in many indexes, since
the 1987 crash.
Without much more data, we cannot say which options cause the VIX to change. Put gouging,
call overwriting or Blacks leverage eect may be at work here, but we cannot say with any
certainty. It would be interesting to further investigate the volatility surface.
IV. TRADING REALIZED AND IMPLIED VOLATILITY
As stated in Sec. I, the VIX has a concrete economic meaning: its square is the variance swap
rate up to corrections due to the fact that there are only SPX options at a nite number of strikes,
as well as the fact that there are occasional jumps in the underlying (the SPX). In Ref. [2], Carr
9
0
1
2
3
4
5
6
-0.05 0 0.05 0.1 0.15 0.2 0.25
N
u
m
b
e
r
Correlation between CVIX
t
and CRVol
t
(non-overlapping intervals)
mean correlation = 0.11
standard deviation of correlations = 0.06
FIG. 6: Correlations between changes in the VIX changes in the 30-day realized volatility computed for 21
dierent non-overlapping intervals, each one consisting of 196 entries.
and Wu use this fact to determine the excess returns that would have been gained from shorting
variance swaps on every day of the sample period, where the excess return is dened as:
ER
t,t+30
= 100
VIX
2
t
RVol
2
t,t+30
VIX
2
t
. (7)
See Figs. 11 and 12 for a time series and a histogram, respectively, of the excess returns.
It is worth emphasizing the fact that the mean excess return gained from continuously shorting
variance swaps is large, at nearly 40 percent. Why is this premium so large? In order to answer this
question, one must take into account that the distribution of excess returns is heavily skewed; there
are a number of occurrences of very large negative excess return. As discussed in Sec. I, parties on
the long side of variance swaps are willing to pay a high premium for the insurance provided against
periods of high realized volatility (relative to the VIX). Whether they are portfolio managers who
are judged against a benchmark, equity funds or others, they are eectively short volatility, and
are therefore willing to pay for the insurance that variance swaps provide. Interestingly, Carr and
Wu [2] argue that the CAPM cannot fully account for the size of the excess return associated with
variance swaps. This perhaps indicates an ineciency or mispricing in this market.
10
10
20
30
40
50
0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02
O
p
t
i
o
n
p
r
i
c
e
(
$
)
Strike/Index
SPX(7/12/06) = $1258.6
VIX(7/12/06) = $14.49
SPX(7/13/06) = $1242.28
VIX(7/13/06) = $17.79
put 7/12/06
put 7/13/06
call 7/12/06
call 7/13/06
FIG. 7: Closing prices for put and call options on the SPX on the day immediately before and the day of a
large VIX increase.
In Sec. II B it was shown that there is no correlation between changes in the VIX and changes in
the 30-day realized volatility of the SPX. This means that the excess returns gained from shorting
30-day variance swaps would increase if the swaps are shorted only on days when there is a large
increase in the VIX as opposed to every day as in Carr and Wu [2]. This is true because a change
in the VIX does not predict a change in the 30-day realized volatility. Therefore, on average the
payo from shorting the swap will be higher when the VIX has recently increased. The opposite
should be true as well: shorting swaps only on days when there is a large decrease in the VIX
should lead to a decrease in the excess returns.
Table I shows the average excess return ER, the standard deviation of the excess returns s
ER
,
and the ratio of the two for various trading strategies. As predicted based on the independence
of changes in the VIX relative to changes in the 30-day realized volatility, we see that strategies
involving the shorting of 30-day variance swaps only on days when there is a large increase in the
VIX slightly outperform a strategy in which the swaps are shorted on every day, while shorting
only on days when the VIX experiences a large decrease does worse.
The eect, however, appears to be small. Presumably, this is the case because the average
excess return is so large for the simplest trading strategy in which the swaps are shorted every
11
10
20
30
40
50
0.98 0.99 1 1.01 1.02 1.03
O
p
t
i
o
n
p
r
i
c
e
(
$
)
Strike/Index
SPX(6/14/06) = $1230.04
VIX(6/14/06) = $21.46
SPX(6/15/06) = $1256.16
VIX(6/15/06) = $15.9
put 6/14/06
put 6/15/06
call 6/14/06
call 6/15/06
FIG. 8: Closing prices for put and call options on the SPX on the day immediately before and the day of a
large VIX decrease.
Number of days swaps are shorted ER s
ER
ER/s
ER
4157 (entire sample) 39.45 36.62 1.08
415 (largest VIX increases) 40.43 35.79 1.13
83 (largest VIX increases) 43.20 35.03 1.23
415 (largest VIX decreases) 35.25 40.81 0.86
83 (largest VIX decreases) 31.10 46.53 0.67
TABLE I: Average excess return ER, the standard deviation of the excess returns s
ER
, and the ratio of the
two for various trading strategies. Strategies involving the shorting of 30-day variance swaps only on days
when there is a large increase in the VIX slightly outperform a strategy in which the swaps are shorted on
every day. Shorting only on days when there is a large decrease in the VIX does worse.
day. Even in the case where one shorts only on the days with the largest one percent of VIX
increases, the relative improvement in the average excess return does not appear to be signicant.
In addition, we considered the possibility of further restricting the trading strategy such that one
is only engaged in one swap contract at a time. Again, such a strategy does not do signicantly
better than the strategy of shorting every day.
12
0.1
0.12
0.14
0.16
0.18
0.2
0.22
0.98 0.985 0.99 0.995 1 1.005 1.01 1.015 1.02
I
m
p
l
i
e
d
v
o
l
a
t
i
l
i
t
y
Strike/Index
SPX(7/12/06) = $1258.6
SPX(7/13/06) = $1242.28
VIX(7/12/06) = $14.49
VIX(7/13/06) = $17.79
(r = 0.03)
put 7/12/06
put 7/13/06
call 7/12/06
call 7/13/06
FIG. 9: Black-Scholes implied volatility for put and call options, the day before and the day of a large VIX
increase. A volatility skew is observed, with implied volatility decreasing with strike price for both puts and
calls, both before and after the big change. An annualized risk-free rate of 0.03 is assumed.
V. CONCLUSION
We investigated a number of features of implied and realized volatility of the SPX index. Sec. II
examines correlations between the SPX, the VIX, and the 30-day realized volatility of the SPX, as
well as between changes in these quantities. We conrmed that the VIX and the 30-day realized
volatility are correlated, and that while changes in the SPX are negatively correlated with changes
in the VIX, the levels of the two indexes are not correlated over the nearly 16 years of data that
we analyzed (although they may be negatively correlated on shorter timescales). Interestingly, as
shown in Fig. 6, we found no signicant correlation between changes in the VIX and changes in
the 30-day realized volatility of the SPX. This means that short term changes in the VIX do not
correctly predict the actual realized volatility, and suggests that at least some options are mispriced
after large moves in the index.
The details of the negative correlation between changes in the SPX and changes in the VIX
were addressed in Sec. III. Without a large dataset of historical options prices, it is dicult to
identify the cause of the negative correlation between changes in the SPX and changes in the VIX.
13
0.12
0.14
0.16
0.18
0.2
0.98 0.99 1 1.01 1.02 1.03
I
m
p
l
i
e
d
v
o
l
a
t
i
l
i
t
y
Strike/Index
SPX(6/14/06) = $1230.04
SPX(6/15/06) = $1256.16
VIX(6/14/06) = $21.46
VIX(6/15/06) = $15.9
(r = 0.03)
put 6/14/06
put 6/15/06
call 6/14/06
call 6/15/06
FIG. 10: Black-Scholes implied volatility for put and call options, the day before and and the day of a large
VIX decrease. A volatility skew is observed, with implied volatility decreasing with strike price for both
puts and calls, both before and after the big move. An annualized risk-free rate of 0.03 is assumed.
To this end, it would be interesting to examine the volatility surface in more detail on days with
large market moves.
We returned to the issue of correlation between realized and implied volatility in Sec. IV. We
began by reproducing the analysis of Carr and Wu [2] regarding the excess return obtained by
continuously shorting variance swaps on every trading day of the sample period. We then analyzed
whether improved returns could be gained by selectively shorting variance swaps using large changes
in the VIX as a signal. In the insurance analogy, this strategy would be similar to an insurance
company carefully selecting when to sell insurance policies based on their expectations about the
excess returns to be had given a particular trigger criterion (e.g., selling hurricane insurance when
demand is high, but the intrinsic probability of a storm has not changed from its historical value).
The lack of correlation identied in Sec. II between changes in the VIX and changes in the 30-
day realized volatility suggests that this strategy would outperform simply continuously shorting
variance swaps. This appears to be the case although statistics are limited. Due to the large
premium (excess returns) already associated with variance swaps, we nd that the additional
advantage is relatively small.
14
-250
-200
-150
-100
-50
0
50
100
2005 2003 2001 1999 1997 1995 1993 1991
E
x
c
e
s
s
r
e
t
u
r
n
s
,
%
Date
FIG. 11: Excess returns from shorting 30-day variance swaps as in Carr and Wu [2].
0
100
200
300
400
500
600
700
-250 -200 -150 -100 -50 0 50 100
N
u
m
b
e
r
Excess returns, %
FIG. 12: Histogram of excess returns from shorting 30-day variance swaps as in Carr and Wu [2]. The mean
excess return is estimated at 39.45 percent, with a standard deviation of 36.62 percent.
15
VI. ACKNOWLEDGEMENTS
We thank Myck Schwetz (PIMCO) for useful comments and some help with historical data,
and Thomas Gould (CSFB) for additional feedback.
[1] Demeter, K., E. Derman, M. Kamal, and J. Zou. More Than You Ever Wanted to Know About
Volatility Swaps. Quantitative Strategies Research Notes, March 1999, Goldman, Sachs & Co.
[2] Carr, P., and L. Wu. A Tale of Two Indices. Working paper, New York University, 2006.
[3] Black, F., and M. Scholes. The Pricing of Options and Corporate Liabilities. Journal of Political
Economy, 81 (1973), pp. 637-654.
[4] Black, F. Studies of Stock Price Volatility Changes. In Proceedings of the 1976 American Statistical
Association, Business and Economical Statistics Section. Alexandria, VA: American Statistical Associ-
ation, 1976, pp. 177-181.