What Practitieners Need te
Knew... by Mark Kritzman
Windham Capital Management
. . . About Estimating Volatility
Parti
Volatility is important to financial analysts for several Table I Standard Deviations of Return
reasons. Perhaps most obvious, estimates of volatil-
ity, together with information about central ten- 1 2 3
dency, allow us to assess the likelihood of experienc- Return Return - Squared
Day (%) Average (%) Difference (%)
ing a particular outcome. For example, we may be
interested in the likelihood of achieving a certain level 1 1.00 0.72 0.0052
2 1.50 1.22 0.0149
of wealth by a particular date, depending on our 3 2.10 1.82 0.0332
choice of alternative investment strategies. In order to 4 -0.40 -0.68 0.0046
assess the likelihood of achieving such an objective, 5 1.00 0.72 0.0052
we must estimate the volatility of returns for each of 6 -1.40 -1.68 0.0281
the alternative investment strategies. 7 0.45 0.17 0.0003
8 -0.75 -1.03 0.0106
Financial analysts are often faced with the task of 9 1.00 0.72 0.0052
combining various risky assets to form efficient port- 10 1.40 1.12 0.0126
folios—portfolios that offer the highest expected re- 11 -2.00 -2.28 0.0519
turn at a particular level of risk.' Again, it is necessary 12 1.00 0.72 0.0052
13 -1.50 -1.78 0.0316
to estimate the volatility of the component assets. 14 0.35 0.07 0.0001
Also, the valuation of an option requires us to esti- 15 -0.30 -0.58 0.0033
mate the volatility of the underlying asset. These are 16 1.00 0.72 0.0052
but a few examples of how volatility estimates are 17 0.00 -0.28 0.0008
used in financial analysis. 18 -0.60 -0.88 0.0077
19 -1.20 -1.48 0.0218
20 2.90 2.62 0.0688
Historical Volatility Average 0.28 0.0167
The most commonly used measure of volatility in Square 1.2904
financial analysis is standard deviation. Standard Root
deviation is computed by measuring the difference
between the value of each observation in a sample
and the sample's mean, squaring each difference,
taking the average of the squares and then determin- In this example, the standard deviation measures
ing the square root of this average. the volatility of daily returns. It is typical in financial
analysis to annualize the standard deviation. Unlike
Suppose, for example, that during a particular rates of return, which increase proportionately with
month we observe the daily returns shown in column hme, standard deviations increase with the square
1 in Table I. The average of the returns in column 1 root of hme. We can take two approaches to convert-
equals 0.28 per cent. Column 2 shows the difference ing a daily standard deviation into an annual stan-
between each observed return and this average re- dard deviation.
turn. Column 3 shows the squared values of these
differences. The average of the squared differences— We can reconvert the standard deviation back into
0.0167 per cent—equals the variance of the returns. a variance by squaring it. Then we can multiply the
(In computing the variance, we divide by the number variance by 260 (the number of trading days in a year)
of observations less one, because we used up one and take the square root of this value to get the
degree of freedom to compute the average of the annualized standard deviation:
returns.) The square root of the variance—1.2904 per
0.012904^ = 0.000167
cent—equals the standard deviation of the daily re-
turns.
0.000167 • 260 = 0.0433
1. Footnotes appear at end of article.
FINANCIAL ANALYSTS JOURNAL / JULY-AUGUST 1991 D 22
Table II Standard Deviations of Logarithmic Returns sult may not be the best estimate if volatility is
unstable through time. In the fall of 1979, for exam-
I 2 3 ple, the Federal Reserve changed its operating policy
Lo^(3 + log - Squared with respect to its management of the money supply
Day Return) (%) Average (%) Difference (%) and interest rates. Over the 10 years ending in 1978,
I 1.00 0.73 0.0053 the annualized standard deviation for long-term cor-
2 1.49 1.22 0.0149 porate bonds was a little less than 8 per cent. In the
3 2.08 1.81 0.0327 subsequent 10-year period, from 1979 through 1988,
4 -0.40 -0.67 0.0045
0.73 0.0053
the annualized standard deviation for long-term cor-
5 1.00
-1.41 -1.68 0.0282 porate bonds rose to more than 13 per cent. Clearly,
7 0.45 0.18 0.0003 historical precedent made a poor guide for estimating
3 -0.75 -1.02 0.0104 bond market volatility in the 1980s.
9 1.00 0.73 0.0053
1.12 0.0126 As an alternative to historical data, we can infer the
10 1.39
\\ -2.02 -2.29 0.0524 investment community's consensus outlook for the
12 1.00 0.73 0.0053 volatilities of many assets by examining the prices at
13 -1.51 -1.78 0.0317 which options on these assets trade. These implied
14 0.35 0.08 0.0001 volatilities presumably reflect all current information
15 -0.30 -0.57 0.0032
0.73 0.0053 that impinges on an asset's volatility.
16 1.00
17 0.00 -0.27 0.0007 The value of an option depends on five factors—
18 -0.60 -0.87 0.0076 the current price of the underlying asset; the strike
19 -1.21 -1.48 0.0218 price, or price at which the option can be exercised;
20 2.86 2.59 0.0671 the time remaining until expiration; the riskless rate
Average (%) 0.27 0.0166
i.ZoO/
of interest; and the volatility of the underlying asset.
Square Root (%) We know the strike price and the time remaining
Annualized 20.75
Standard until expiration from the terms of the option conti-act.
Deviation (%) The price of the underlying asset and the riskless rate
of interest can be determined from a variety of market
quote services. The only factor that we do not know
with certainty is the volatility of the underiying asset.
Alternatively, we can multiply the daily standard In order to determine volatility, we can substitute
deviation by the square root of 260 to determine the various values into the option pricing formula until
annualized value. This is algebraically equivalent to the solution to this formula equals the price at which
the first approach: the option is trading.
Consider a call option with 90 days to expiration
260 = 16.1245 and a strike price of $295.00, written on an underly-
ing asset currently priced at $300.00. Suppose the
0.012904- 16.1245 = 0.2081. annualized riskless rate of interest is 8 per cent and
the option trades for $15.00. In order to determine the
The approach we have just described for estimating
standard deviation, or implied volatility, consistent
the standard deviation from historical returns yields
with the price of this option, we start by assuming
an estimate that is approximately correct, but not
some value for volatility and use this value in the
exact. The inexactitude arises because the dispersion
following Black-Scholes option-pricing formula:
of investment returns conforms to a lognormal distri-
bution, rather than a normal distribution, owing to
the effect of compounding. We can attain a more C = S • N(D) - • N(D - (1)
precise estimate of standard deviation by calculating
the logarithms of one plus the returns, squaring the where
differences of these values h-om their average, and S = price of underlying asset,
taking the square root of the average of the squared X = strike price,
differences. Table II shows these calculations. T = time remaining until expiration,
Comparing the standard deviation from Table I r = instantaneous riskless rate of interest.
with the standard deviation from Table 11, we see that
it does not make much difference which approach we D = (ln(S/X) + {r+ a^l2) • T)/(cr •
use given this short measurement interval. As the ln( ) = natural log,
measurement interval lengthens, however, the dis- a = standard deviation (volatility) of underly-
tinction becomes more important. ing asset and ^
N( ) = cumulative normal distribution function.
Implied Volatility Suppose we use the historical volatility over the
As we have seen, estimating volatility from histoncal previous 90 days—say, 20 per cent. By substituting
data is fairly straightforward. Unfortunately, the re-
FINANCIAL ANALYSTS JOURNAL / JULY-AUGUST 1991 D 23
this value and the values we assigned earlier into The volatility of 20 per cent resulted in an option
Equation
E t i (1)
(1), we ffind
i d tthat
h t D equals
l 0401
0.401. Thus
Th C,C the
h value of $17.69. We compute the Newton-Raphson
option value, is calculated as: volatility estimate as follows:
C = 300 • N(0.401) - 295 e - 0.077(90/365)
• N(0.401 N-R Estimate = 0.20 - (17.69 - 15)/54.64
- 0.2 • 90/365) = 0.1507.
A volatility estimate of 15.07 per cent results in an
C = 17.69.
option value of $15.06. One more iteration yields a
This estimated value is greater than the price at volatility estimate of 14.96 per cent, for an option
which the option currently trades. We must therefore price of $15.00.
lower our estimate for volatility. Suppose we next try
a value of 12 per cent. Based on this volatility as- Method of Bisection
sumption, the option value equals $13.50; this is less The efficiency of the Newton-Raphson Method de-
than the actual price of $15.00, implying that we pends to a certain extent on the choice of the initial
should raise our estimate. volatility estimate. An alternative search procedure,
If we continue substituting various volatility values which tends to be less sensitive to the initial volatility
into the Black-Scholes formula, we will eventually estimate, is called the Method of Bisection.* This
discover that a volatility estimate of 14.96 per cent is approach is more intuitive. We start by choosing a
consistent with an option value of $15.00—the price low estimate for volatility corresponding to a low
at which the option is currently trading. This 14.96 option value and a high estimate for volatility corre-
per cent is the implied volatility, given the current sponding to a high option value.
values for the underlying asset, the option and the For example, suppose we start with a low estimate
riskless rate and given the terms of the option con- of 10 per cent and a high estimate of 30 per cent.
tract. Given the assumptions from our previous example,
the corresponding option values are $12.56 and
Newton-Raphson Method $23.27. Our next estimate for volatility is found by
Of course, as we are solving for the implied volatility, interpolation, as shown below:
the prices of the underlying asset and the option may New Estimate = (C-
be changing. We need a reasonably quick way to
arrive at the implied volatility. The Newton-Raphson (CH - CJ, (3)
Method is one way.
According to the Newton-Raphson Method, we 0.1456 = 0.1 + (15 - 12.56) • (0.3 - 0.1)/
start with some reasonable estimate for volatility and
evaluate the option using this estimate. Unless we are (23.27 - 12.56),
unusually lucky, however, we will not arrive at the
correct value for implied volatility on our first try. We 0.1494 = 0.1456 + (15 - 14.79) • (0.3
therefore revise our initial volatility estimate by sub- -0.1456)/(23.27-14.79),
tracting an amount equal to the estimated option
value minus the option's actual price, divided by the 0.1496 = 0.1494 + (15 - 14.99) • (0.3
derivative of the option formula with respect to
volatility evaluated at our estimate for volatility. This -0.1494)/(23.27-14.99).
derivative is shown below:
If the option value corresponding to our interpo-
lated estimate for volatility is below the actual option
=S• \/f (2)
price, we replace our low volatility estimate with the
where interpolated estimate and repeat the calculation. If
the estimated option value is above the actual option
n = 3.1416 and price, we replace the high volatility estimate with the
e = 2.7183. interpolated estimate and proceed accordingly. When
the option value corresponding to our volatility esti-
C, S, T and D are as defined earlier. mate equals the actual price of the option, we have
Our earlier example used an assumed volatility of arrived at the implied volatility for that option.
20 per cent. Using this assumption, the derivative for
the Newton-Raphson Method is: Historical vs. Implied Volatility
Is it better to estimate volatility from historical obser-
dC/i)a = 300 • , - 0.401^/2
vations or to infer it from the prices at which options
trade? The answer, of course, depends on the quality
SO da = 54.64. of the inputs. If volatility is stationary through time.
FINANCIAL ANALYSTS JOURNAL / JULY-AUGUST 1991 D 2 4
and we have reliable prices from which to estimate even though it has a lower R-squared than implied
returns, then historical volatility is a reasonably good volatility.'
indicator of subsequent volatility. Unfortunately, and An alternative method of comparison would be to
especially over short measurement intervals, nonre- examine the tracking errors associated with historical
curring events or conditions often cause volatility to and implied volatilities. Tracking error is computed
shift up or down temporarily, so that historical vola- by squaring the differences between actual values
tility will over or underestimate subsequent volatility. and historical or implied values, taking the average of
To the extent that the investment community recog- these differences, and then calculating the square
nizes the transitory nature of these nonrecurring root of the average.
events, implied volatility may provide a superior A relatively recent innovation for estimating vola-
estimate of subsequent volatility. In estimating im- tility uses a technique known as ARCH, an acronym
plied volatility, however, we must use contempora- for Autoregressive Conditional Heteroscedasticity.*'
neous observations for the inputs to the Black- Essentially, ARCH and related models incorporate
Scholes formula. the time-series dynamics of past volatility to forecast
The following procedure provides an intuitively subsequent volatility. We will introduce ARCH mod-
appealing test of whether historical or implied vola- els in Part 2 of "What Practitioners Need to Know
tility is a superior forecaster of actual volatility. First, About Estimating Volatility."
estimate historical volatility in periods one through n.
Then estimate implied volatility as of the end of Footnotes
periods one through n. Next estimate actual volatility 1. See, for example, M. Kritzman, "What Practition-
in periods two through n plus one. Two simple linear ers Need to Know: The Nobel Prize," Financial
regressions can then be performed—one in which the Analysts Journal, January/February 1991.
independent variable is the historical volatility and 2. To be precise, we should use the continuously
the dependent variable is the actual volatility in the compounded riskless rate of interest. Thus, if the
subsequent period, and the other in which the inde- rate is quoted as simple interest, we should use
pendent variable is the implied volatility and the the natural log of one plus the interest rate.
dependent variable is the actual volatility in the 3. For a discussion of the cumulative normal distri-
subsequent period. bution function, see M. Kritzman, "What Practi-
It may be tempting to conclude that the regression tioners Need to Know About Uncertainty/' Finan-
equation with the higher R-squared reveals the better cial Analysts journal, March/April 1991.
predictor, but this conclusion may very well be 4. For an excellent review of this approach, see S.
wrong. The regression equation with implied volatil- Brown, "Estimating Volatility," in Figlewski, Sil-
ity as the independent variable may have a higher ber and Subrahmanyam, eds.. Financial Options:
R-squared, but the slope of the regression line may be From Theory to Practice (Homewood, IL: Business
significantly greater or less than one, or the intercept One Irwin, 1990).
may be significantly greater or less than zero. In these 5. For further elucidation, see L. Canina and S.
cases, the forecasts would be biased, although the Figlewski, "The Information Content of Implied
R-squared alone would not reveal this. The regres- Volatility" (Working paper number S-91-20, Sa-
sion equation with historical volatility as the indepen- lomon Center, New York University).
dent variable might have the weaker R-squared but a 6. See R. Engle, "Autoregressive Conditional Het-
slope and intercept closer to one and zero, respec- eroscedasticity with Estimates of the Variance of
tively. In this contrived example, historical volatility United Kingdom Inflation," Econometrica 17, pp.
may be the better predictor of subsequent volatility. 5-26.
From the Board concluded from page 12.
from about 10 to 18—is unlikely to be the key factor contrast, P/Es at the extremes—over 20 and under
determining the attractiveness of the stock market. In 8—are much more difficult to ignore. Very low or
this area, contrarian logic has demonstrated littie very high P/Es—although they do not guarantee the
value, and factors other than P/Es are likely to pro- outcomes shown in Figure C—suggest that the con-
vide the clue to the outlook for stock prices. In trarian argument be accorded considerable weight.
FINANCIAL ANALYSTS JOURNAL / JULY-AUGUST 1991 D 2 5
