CHAPTER 23

Estimating Volatilities and Correlations

Practice Questions

23.1
Define u i as (Si  Si 1 )  Si 1 , where S i is value of a market variable on day i . In the EWMA
model, the variance rate of the market variable (i.e., the square of its volatility) calculated for
day n is a weighted average of the un2i ’s ( i  1 2 3…). For some constant  ( 0    1 ), the
weight given to un2i 1 is  times the weight given to un2i . The volatility estimated for day
n ,  n , is related to the volatility estimated for day n  1 ,  n1 , by
 n2   n21  (1   )un21
This formula shows that the EWMA model has one very attractive property. To calculate the
volatility estimate for day n , it is sufficient to know the volatility estimate for day n  1 and
un 1 .

23.2
The EWMA model produces a forecast of the daily variance rate for day n which is a
weighted average of (i) the forecast for day n  1 , and (ii) the square of the proportional
change on day n  1 . The GARCH (1,1) model produces a forecast of the daily variance for
day n which is a weighted average of (i) the forecast for day n  1 , (ii) the square of the
proportional change on day n  1 , and (iii) a long run average variance rate. GARCH (1,1)
adapts the EWMA model by giving some weight to a long run average variance rate.
Whereas the EWMA has no mean reversion, GARCH (1,1) is consistent with a mean-
reverting variance rate model.

23.3
In this case,  n1  0015 and un  05  30  001667 , so that equation (23.7) gives
 n2  094  00152  006  0016672  00002281
The volatility estimate on day n is therefore 00002281  0015103 or 1.5103%.

23.4
Reducing  from 0.95 to 0.85 means that more weight is put on recent observations of ui2
and less weight is given to older observations. Volatilities calculated with   085 will react
more quickly to new information and will “bounce around” much more than volatilities
calculated with   095 .

23.5
The volatility per day is 30  252  189% . There is a 99% chance that a normally
distributed variable will be within 2.57 standard deviations. We are therefore 99% confident
that the daily change will be less than 257 189  486% .
23.6
The weight given to the long-run average variance rate is 1     and the long-run average
variance rate is   (1     ) . Increasing  increases the long-run average variance rate;
increasing  increases the weight given to the most recent data item, reduces the weight
given to the long-run average variance rate, and increases the level of the long-run average
variance rate. Increasing  increases the weight given to the previous variance estimate,
reduces the weight given to the long-run average variance rate, and increases the level of the
long-run average variance rate.

23.7
The proportional daily change is 0005  15000  0003333 . The current daily variance
estimate is 00062  0000036 . The new daily variance estimate is
09  0000036  01 00033332  0000033511
The new volatility is the square root of this. It is 0.00579 or 0.579%.

23.8
With the usual notation un–1 = 20/3040 = 0.006579 so that the new variance is

0.000002 + 0.06 × 0.0065792 + 0.92 × 0.012 = 0.00009660

so that n = 0.00983. The new volatility estimate is therefore 0.983% per day.

23.9
(a) The volatilities and correlation imply that the current estimate of the covariance is
025  0016  0025  00001.
(b) If the prices of the assets at close of trading are $20.5 and $40.5, the proportional
changes are 05  20  0025 and 05  40  00125 . The new covariance estimate is
095  00001  005  0025  00125  00001106
The new variance estimate for asset A is
095  00162  005  00252  000027445
so that the new volatility is 0.0166. The new variance estimate for asset B is
095  00252  005  001252  0000601562
so that the new volatility is 0.0245. The new correlation estimate is
00001106
 0272
00166  00245

23.10
The long-run average variance rate is   (1     ) or 0000004  003  00001333 . The
long-run average volatility is 00001333 or 1.155%. The equation describing the way the
variance rate reverts to its long-run average is equation (23.13)
E[ n2k ]  VL  (   )k ( n2  VL )
In this case,
E[ n2k ]  00001333  097k ( n2  00001333)
If the current volatility is 20% per year,  n  02  252  00126 . The expected variance
rate in 20 days is
00001333  09720 (001262  00001333)  00001471
The expected volatility in 20 days is therefore 00001471  00121 or 1.21% per day.

23.11
Using the notation in the text  u n1  001 and  vn1  0012 and the most recent estimate of
the covariance between the asset returns is cov n1  001 0012  050  000006 . The
variable un1  1  30  003333 and the variable vn1  1  50  002 . The new estimate of the
covariance, cov n , is
0000001  004  003333  002  094  000006  00000841
The new estimate of the variance of the first asset,  u n is
2

0000003  004  0033332  094  0012  00001414

so that  u n  00001414  001189 or 1.189%. The new estimate of the variance of the
second asset,  vn is
2

0000003  004  0022  094  00122  00001544

so that  vn  00001544  001242 or 1.242%. The new estimate of the correlation between
the assets is therefore 0.0000841/(0.01189 × 0.01242) = 0.569.

23.12
The FTSE expressed in dollars is XY where X is the FTSE expressed in sterling and Y is
the exchange rate (value of one pound in dollars). Define xi as the proportional change in X
on day i and y i as the proportional change in Y on day i . The proportional change in XY is
approximately xi  yi . The standard deviation of xi is 0.018 and the standard deviation of y i
is 0.009. The correlation between the two is 0.4. The variance of xi  yi is therefore
00182  00092  2  0018  0009  04  00005346
so that the volatility of xi  yi is 0.0231 or 2.31%. This is the volatility of the FTSE
expressed in dollars. Note that it is greater than the volatility of the FTSE expressed in
sterling. This is the impact of the positive correlation. When the FTSE increases, the value of
sterling measured in dollars also tends to increase. This creates an even bigger increase in the
value of FTSE measured in dollars. Similarly, for a decrease in the FTSE.

23.13
Continuing with the notation in Problem 23.12, define zi as the proportional change in the
value of the S&P 500 on day i . The covariance between xi and zi is
07  0018  0016  00002016 . The covariance between y i and zi is
03  0009  0016  00000432 . The covariance between and zi equals the covariance
between xi and zi plus the covariance between xi  yi y i and zi . It is
00002016  00000432  00002448
The correlation between xi  yi and zi is
00002448
 0662
0016  00231
Note that the volatility of the S&P 500 drops out in this calculation.
23.14
 n2  VL   un21   n21
so that
2n  (1    )VL  un21  2n1
2n  2n1  (1    )(VL  2n1 )  (un21  2n1 )
The variable un21 has a mean of  n21 and a variance of
E (un1 )4  [ E (un21 )]2  2 n41
The standard deviation of un21 is 2 n21 .
We can write V   n2   n21 and V   n21 . Substituting for un21 into the equation for
 n2   n21 , we get
V  a(VL  V )  Z

where Z is a variable with mean zero and standard deviation  2V . This equation defines
the change in the variance over one day. It is consistent with the stochastic process
dV  a(VL  V )dt   2Vdz
or
dV  a(VL  V )dt  Vdz
when time is measured in days.

Discretizing the process, we obtain

V  a(VL  V )t  V  t

where  is a random sample from a standard normal distribution.

Note that we are not assuming Z is normally distributed. It is the sum of many small changes
V  t .

When time is measured in years,

V  a(VL  V )252t  V  252 t

and the process for V is
dV  252a(VL  V ) dt  V 252 dz

23.15 See Excel file

The worksheet monitors variances and covariances using EWMA setting initial values equal
to the value calculated in the usual way from data. VaR is $302,459 while ES is $346,516.

23.16
The parameter  is in cell N3 of the EWMA worksheet for the previous problem is changed
to 0.97. VaR increases to $393,300 and ES increases to $450,590.
23.17
The proportional change in the price of gold is 4  600  000667 . Using the EWMA
model, the variance is updated to
094  00132  006  000667 2  000016153
so that the new daily volatility is 000016153  001271 or 1.271% per day. Using GARCH
(1,1), the variance is updated to
0000002  094  00132  004  0006672  000016264
so that the new daily volatility is 000016264  01275 or 1.275% per day.

23.18
The proportional change in the price of silver is zero. Using the EWMA model, the variance
is updated to
094  00152  006  0  00002115
so that the new daily volatility is 00002115  001454 or 1.454% per day. Using GARCH
(1,1), the variance is updated to
0000002  094  00152  004  0  00002135
so that the new daily volatility is 00002135  001461 or 1.461% per day. The initial
covariance is 08  0013  0015  0000156 . Using EWMA, the covariance is updated to
094  0000156  006  0  000014664
so that the new correlation is 000014664  (001454  001271)  07934 . Using GARCH
(1,1), the covariance is updated to
0000002  094  0000156  004  0  000014864
so that the new correlation is 000014864  (001461 001275)  07977 .
For a given  and  , the  parameter defines the long run average value of a variance or a
covariance. There is no reason why we should expect the long run average daily variance for
gold and silver should be the same. There is also no reason why we should expect the long
run average covariance between gold and silver to be the same as the long run average
variance of gold or the long run average variance of silver. In practice, therefore, we are
likely to want to allow  in a GARCH(1,1) model to vary from market variable to market
variable. (Some instructors may want to use this problem as a lead in to multivariate GARCH
models.)

23.19 (Excel file)

In the spreadsheet, the first 25 observations on (vi-)2 are ignored so that the results are not
unduly influenced by the choice of starting values. The best values of for EUR, CAD, GBP
and JPY were found to be 0.947, 0.898, 0.950, and 0.984, respectively. The best values of
for S&P500, NASDAQ, FTSE100, and Nikkei225 were found to be 0.874, 0.901, 0.904,
and 0.953, respectively.

23.20 (Excel file)

As the spreadsheets show, the optimal value of  in the EWMA model is 0.958 and the log
likelihood objective function is 11,806.4767. In the GARCH (1,1) model, the optimal values
of , , and  are 0.0000001330, 0.04447, and0.95343, respectively. The long-run average
daily volatility is 0.7954% and the log likelihood objective function is 11,811.1955.