What Practitioners Need to Know

. . . About Lognormality

Mark Kritzman multiply large numbers direaly ceive $271.46. And if it is com-
Windbam Capital Manage- than to raise a base to a fractional pounded hourly, we will receive
ment power. In the olden days, how- $271.81 by year-end.
ever, an analyst would use a slide
When reading the financial litera- rule, which is a ruler with a slid- It seems as though the more fre-
ture we often see statements to ing central strip marked with log- pounded, our
quently interest is com-
arithmic scales. the more money we
the effect that a particular result will end up with at the end of the
depends on the assumption that year. No matter how frequentlyit
retums are lognormally distrib- is compounded, though, we will
uted. What exactly is a lognormal In most financial applications, in- never receive more than $271.83.
distribution, and why is it relevant stead of logarithms to the base 10, The limit of the function (1 -f-
to financial analysis? In order to we use logarithms to the base r/n)", when r ecjuals 100%, as n
address this question, let us start 2.71828, which is denoted by the approaches infinity is 2.71828, the
with a review of logarithms. letter e in honor of the famous
Swiss mathematician Euler. These base of the natural log.
Logarithms logarithms, which are called nat- We can use this result to convert
A logarithm is simply the power ln, ural logs and are abbreviated as periodic rates of retum into con-
to which a base must be raised to posehave a special property. Sup- tinuous rates of retum. A periodic
yield a particular value. For exam- ning of invest
we $100 at the begin-
the year at an annual rate of retum is computed as the
ple, the exponent 2 is the loga- interest rate of 10096. At the end percentage change of our invest-
rithm of 16 to the base 4, because ofthe year we will receive $200— ment from the beginning of the
4 squared equals 16. The loga- period to the end of the period,
rithm of 8 to the base 4 equals 1.5, our original principal of 1100 and assuming there are no contribu-
because 4 raised to the power 1.5 another $100 of interest. Now tions or disbursements. A contin-
equals 8. suppose our interest is com- uous rate of retum assumes that
pounded semiannually. Our year- the income and growth are com-
The choice of a base depends on end payment will equal $225. By pounded instantaneously.
the context in which we use log- the middle of the year we will
arithms. For simple mathematical have earned $50 of interest, From our previous example we
procedures, it is common to use which is then reinvested to gen- know that e, the base of the natu-
the base 10, which explains why erate the additional $25. ral log, raised to the power 1 (the
logarithms to the base 10 are continuous rate of retum in our
called common logs. The base 10 In general, we can use the follow- example) yields 1 plus 171.83%
is popular because the logarithms ing formula to compute the year- (the periodic rate of retum in our
of 10, 100, 1000 and so on equal end value of our investment for example). The natural log of the
1, 2, 3 respectively. any interest rate and for any fre- quantity 1 plus the pericJdic rate
quency of compounding: of retum must therefore equal
Why should we care about loga- the corresponding continuous
rithms? In the days prior to E = B • (1 + r/n)" rate of retum. For example, the
i pocket calculators ^ong before natural log of the quantity 1 plus
my time), logarithms were useful where 10% equals 953%. This means
for performing complicated com- that if we invest $100 at a contin-
CC
putations. Financial analysts E = ending value, uously compounded rate of re-
g would multiply large numbers by B = beginning investment, tum of 953%, our investment will
summing their logarithms, and r = annual interest rate and grow to $110. The value 1.10,
they would divide them by sub- n = frequency of compound- which we compute by raising e to
tracting their logarithms. For ex- ing. the power 0.0953, is called the
ample, given a base of 4, we can exponential. These relationships
multiply 16 times 8 by raising the If the 100% rate of interest is are shown below:
u number 4 to the 35 power, which compounded quarterly, our $100
I is the sum of the logarithms 2 and investment will grow to $244.14 ln(1.10) = 0.0953,
cn 1.5. Of course, you might argue by the end of the year. If it is
10 that it would have been easier to compounded daily, we will re- 1.10 = 2.71828°°^'.
Figure A Continuously Compounded curve, as shown in Figure A. The retum. The important distinaion
Retum as a Function of Time logarithms of these values, how- is that we sum the natural logs to
ever, form a straight line when derive the cumulative return,
400 plotted as a function of time, in- whereas we multiply 1 plus the
dicating a constant periodic rate periodic retums to derive the cu-
of grovvth. It is this relation that mulative retum.
gives rise to the logarithmic scale,
300
in which equal percentage The Central Limit Theorem,
changes correspond to equal ver- which is one of the most impor-
tical distances. This scale is shown tant notions of statistical infer-
200 on the left axis in Figure B, with ence, deals with the summation
the logarithms shown along the of random variables. A normal
right axis. distribution, which is symmetric,
can be described by itsfirsttwo
Why Are Returns central moments—its mean and
Lognormally Distributed? its variance.^ Higher moments,
Time Suppose we invest $100 for one such as skewness, have a value of"
year and suppose that the quarterly 0 in a normal distribution. Skew-
retums during this period are 10%, ness is computed by raising a
-5%, 15% and -10%. Although value to an odd-number expo-
We can also compute the contin- the sum of these quarterly retums nent; unlike variance, which is
uous rate of retum within a pe- equals 10%, our $100 investment computed by squaring a value,
riod by subtracting the natural log grows, not to $110, but to $108.16 skewness can take on either a
of the beginning value from the ($100 • 1.10 • 0.95 • 1.15 • 0.90), positive or a negative value. When
natural log of the ending value. yielding a one-year cumulative re- a large number of random varia-
Suppose we start with $100, tum of 8.16%. The sum of the bles are summed, the central mo-
which is invested so that it grows quarterly retums does not mea- ments that are computed with
to $150 after one year, $225 after sure the actual cumulative retum odd-number exponents tend to
two years and $337.50 after three we realize. It is the compounding cancel each other out, leaving
years. The logarithms of these of these quarterly retums that mean and variance as the only
values equal 4.6032, 5.0106,
5.4161 and 5.8216, respectively. yields the cumulative retum. non-zero central moments.
The difference between each log-
arithm and the next one equals We can sum the natural logs of Because we sum logarithms, the
the continuously compounded the quantities 1 plus the quarterly natural logs of the quantities 1
return each year, which is retums to determine the cumula- plus the periodic retums are nor-
40.55%. This continuous retum tive retum. These natural logs mally distributed. Because these
corresponds to a yearly periodic equal 0.0953, -0.0513, 0.1398 natural logs are normally distrib-
retum of 50%. and -0.1054, respeaively, and uted, and because the exponen-
they sum to 0.0784. If we raise e tial of the normal distribution
to the power 0.0784, we get the gives the lognormal distribution,
If we plot these values as a func- exponential 1.0816, which equals the quantities 1 plus the periodic
tion of time, we produce a convex 1 plus the one-year cumulative retums, which are the exponen-
rate of retum. tials of the natural logs, are log-
normally distributed.
Figure B Logarithms of Returns What has all this to do with log- fe
as a Function of Time normality? We have seen that the Normal DistHbution and
sum of the logarithms of the Lognormal Distribution
quantities 1 plus the periodic re- Having said all that, why should
tums equals the logarithm of the we care? Table I shows the esti- 3
»5 8216
400 quantity 1 plus the cumulative mated probabilities of achieving
OE:
300
O

Table I Probability of Achieving Target Retums

200

Target Retums N
One Year
L N
Five Years
L N
10 Years
L
i
u
-5% 80% 79% 99% 97% 100% 100%
100 0% 73 71 96 90 100 96
1 2 3 4 8% 58 57 74 66 93 72
20% 34 37 5 22 0 14
Time 11
 various annualized holding-pe- where
riod retums for an investment
with a geometric mean retum of r = periodic retum in year i and
12% and a standard deviation of n = number ofyears.

It is a better description ofthe true

average retum than is the arithmetic
Given a one-year horizon, it does mean, in the following sense. If we
not make very much difference eamed the geometric retum each
whether we assume a normal dis- year, we would have achieved the
tribution or a lognormal distribu- same terminal wealth that was gen-
tion to estimate probabilities.^ As erated by the actual yearly retums.
we extend our horizon, however, 3. The normal deviate for the normal
the normal distribution overesti- distribution is computed as:
mates the probability of achieving
target retums that are below the 1(1+ T)" - (1 + rf}/(a
expected return. For example,
given afive-yearhorizon, the like- where
lihood of achieving at least an 8%
annualized retum equals 74%, as- T = target retum,
r = geometric retum,
suming a normal distribution, <T = standard deviation and
versus 66% for a lognormal dis- n = number of years.
tribution. Over 10 years, the dif-
ference is even greater. Based on The normal deviate for the lognor-
a normal distribution, there is a mal distribution is computed as:
93% chance of achieving at least
an 8% retum, while under the Mil + TI") - ln(l + r) • nll(j\/n
assumption of a lognormal distri-
bution, the probability falls to
71%. The normal distribution
also Mn<:ferestimates the probabil-
ity of achieving target retums
above the expected retum. Given
afive-yearhorizon and assuming
a normal distribution, the chance
of achieving a retum of 20% or
greater equals 5%, versus 22% for
a lognormal distribution. Over a
10-year horizon, the probabilities
equal 0% and 14% for the normal
and lognormal distributions, re-
spectively.

Based on this evidence, we are

§ well advised to note whether a
normal or lognormal distribution
was assumed in generating the
results of interest to us, especially
if they pertain to a multi-year
horizon.
CC

O
Footnotes
/. For a review ofthe normal distribu-
tion, see M. Kritzman, "What Practi-
tioners Need t6 Know M>out Uncer-
tainty, " Financial Analysts Journal,
March/April 1991.
2. The geometric mean is computed as:
<

+ r) - 1
12