Econ3016: Empirical Finance WEEK 3/4

ECON3016: EMPIRICAL FINANCE
WEEK 3/4
Stock Returns: Annualisations Stock Returns: Continuous Compounding Stock Returns: Stylised Facts Stock Returns: A simple Model
Topics
Annualisations
Continuously Compounded Returns
Stylised Facts
Uncorrelatedness versus Independence
Annualisations
Consider an investment instrument that returned 10%, 12% and 5%

over three periods. The cumulative gross return will be
(1.10)(1.12)(1.05) = 1.29 i.e. if we had invested £1 into the asset,
after three period we would end up with £1.29. Note that this is quite
different from simply adding the three sets of returns and obtaining
27%.
Here the per annum (annualised) performance is
1
((1 + 0.10)(1 + 0.12)(1 + 0.05)) 3 − 1 = 8.96%. Which question is
this quantity answering? I start with £1 and after 3 years I end up with
£1.29. What is the annual interest that leads to this outcome? Check
that 1(1 + 0.0896)3 = 1.29
Annualising Returns
Simple scenario: we have a stream of annual returns, say R1 = 0.05,

R2 = 0.12, R3 = 0.04 and R4 = −0.10. We know that
(1 + 0.05)(1 + 0.12)(1 + 0.04)(1 − 0.10) = 1.10074
i.e. if we placed £1 in a fund that returned the above quantities over four
years we would end up with £1.10074.
Can we come up with a composite return measure, say some yearly return
RA , such that if we received the same RA over 4 years we would also end up
with £1.10074? recall FV = £1(1 + RA )n so that what we want to do here is
to find RA such that (1 + RA )4 = (1 + R1 )(1 + R2 )(1 + R3 )(1 + R4 ). Clearly
RA = [(1 + R1 )(1 + R2 )(1 + R3 )(1 + R4 )]1/4 − 1 = 0.024285
RA is our annualised return (per annum return). Important: We assumed

yearly returns.
Annualising Returns
PT
Arithmetic average return: R = T1 t=1 Rt .
An alternative measure of average return is the cumulative return over
the period from 1 to T, annualised by taking the T th root (raising to
power 1/T). This is called the geometric average return
T
! T1
Y
1 + RG ≡ (1 + Rt )
t=1
Intuition: If we invest $1 in an asset, after one period we have (1 + R1 ),
after two periods we have (1 + R1 )(1 + R2 ) and after T periods we have
QT
t=1 (1 + Rt ). The geometric average RG answers the question:
QT
What
constant return RG leads to a $1 investment turning into t=1 (1 + Rt )
QT
after T periods? It’s the RG that solves (1 + RG )T = t=1 (1 + Rt ).
[similar to the APR!].
Think of RG as the annualised return (p.a).
Annualising Returns
More formally: Let k be the number of compounding periods and let n

be the number of compounding periods in a year, so that there are
N = k/n years of data. The annualised return is then defined as the
geometric average of the returns
 n/k
k−1
Y
Rt (k) =  (1 + Rt−j ) − 1
j=0
Note: With yearly compounding we have n = 1 and k is the number of

years so that the above expression simplifies to
 1/k
k−1
Y
Rt (k) =  (1 + Rt−j ) − 1
j=0
Annualising Returns
Example: Consider an investment instrument that returned 10%, 12%

and 5% over three periods. The cumulative gross return will be
(1.10)(1.12)(1.05) = 1.2936 i.e. if we had invested £1 into the asset,
after three period we would end up with £1.29. The annualised return
1
from this asset is (1.29) 3 − 1 = 8.96%. Note that this different from
the arithmetic average of 9%. Within the previous notation we have
N = 3, n = 1, k = 3.
Meaning of 8.96%: If you start with £1 and the yearly return is 8.96%
you will end up with £1.2936. So the 8.96% figure (annualised return)
has the same interpretation as the APR.
Annualising Returns
Example: Suppose the compounding interval is monthly (n=12), the

monthly return is 1% and there are two years of data (k=24). The
12
annualised return is given by ((1.01)24 ) 24 − 1 = 0.1268 or 12.68%.
Intuition: I am looking for the annual rate, say RA such that
(1 + RA )2 = (1 + 0.12
12 )
12×2
. Solve for RA and obtain RA = 12.68%.
Annualising Returns
Should we use arithmetic averages or geometric averages? Which

average is better? It depends.
In 1999, 2000 and 2001 the SP500 returned R1 = 0.2104,

R2 = −0.0910 and R3 = −0.1189. The two means are
R̄ = (0.2104 − 0.0910 − 0.1189)/3 ≈ 0% and
Q3
RG = [ t=1 (1 + Rt )]1/3 − 1 ≈ −1%.
Geometric means are very useful for highlighting what happens over a
long period. Note that when returns are constant over time then R̄ = RG
otherwise R̄ is always at least as big as RG .
Annualising Returns
Suppose that half the time an investment earns 60% and half the time it
earns -50%. So if you invest for 10 years, then for 5 years the return is
60% and for 5 years -50%. The arithmetic average is
R̄ = (0.6 + 0.6 + 0.6 + 0.6 + 0.6 − 0.5 − 0.5 − 0.5 − 0.5 − 0.5)/10 = 5%.
Not too bad!
Suppose however that I hold the asset for 2 years. In the first year it
returns 60% and in the second -50%. How much money do I have at the
end of the second year? I invest 1 dollar. First it goes 1.60 then to
(1.6)(0.5) i.e 80 cents.
Annualising Returns
Suppose I hold it for 10 years and for five it goes 60% and for 5 it goes
-50%. Then after 10 years my dollar invested is worth
(1.6)5 (0.5)5 = $0.33.
If it goes up for 50 of the next hundred years, and down for the other
50, then in 100 years I have (1.6)50 (0.5)50 ≈ 0!!!!
If half the time the stock goes up by 60% and half the time it goes down
50% then over the long run I get NOTHING, eventhough the arithmetic
average is 5%. Lesson: Arithmetic returns are unreliable for
understanding long run returns
Continuous Compounding
One Period Returns: The continuously compounded return rt is defined as

rt = ln(1 + Rt ) = ln(Pt /Pt−1 ).
Intuition: Take the exponential of both sides, we get ert = 1 + Rt = Pt /Pt−1 .

Rearranging gives Pt = Pt−1 ert . So that rt is the continuously compounded growth
rate in prices between t-1 and t. Notice: rt = ln Pt − ln Pt−1 = pt − pt−1 .
To see further where this comes from go back to page 2 of Week 1 slides. We obtain
1
rt as Pt−1 (1 + rt ∆) ∆ = Pt as ∆ → 0. Note that in page 2 we have the analogy
1
∆ ≡ 1/m. Since lim∆→0 (1 + rt ∆) ∆ = ert we end up with Pt = ert Pt−1 .
Economists often prefer to work with continuously compounded returns. The latter
have important technical advantages. Recall that we could not simply add simple
period returns to obtain multi-period versions. Also, the annualisation process
required us to compute geometric averages.
Disadvantage of ccr’s: When dealing

PN with portfolio averagesP we can only use simple
returns in the sense that Rpt = i=1 wi Rit whereas rpt 6= wi rit .
Given a monthly ccr rt it is easy to solve for the corresponding simple net return as
Rt = ert − 1. Hence nothing is lost by considering continuously compounded returns
instead of simple returns. CCRs are very similar to simple returns as long as the
return is relatively small (this is generally the case for daily or monthly returns). Most
of the time we will be working with CCRs.
Idea: ln(1 + x) ≈ x for x small. Take ln(Pt /Pt−1 ) = ln(1 + Rt ) ≈ Rt when Rt small.
Multi-Period Returns: Consider the two month ccr defined as
rt (2) = ln(1 + Rt (2)) = ln(Pt /Pt−2 ) = pt − pt−2 . Taking exponentials of both sides
shows that Pt = Pt−2 ert (2) so that rt (2) is the continuously compounded growth rate
of prices between months t-2 and t. Using Pt /Pt−2 = (Pt /Pt−1 )(Pt−1 /Pt−2 ) and the
properties of logs it also follows that
rt (2) = ln(Pt /Pt−2 ) = ln(Pt /Pt−1 ) + ln(Pt−1 /Pt−2 ) = rt + rt−1 . So the cc two
month return is simply the sum of the two cc one month returns.
The cc k-month return is defined as

rt (k) = ln(1 + Rt (k)) = ln(Pt /Pt−k ). It is now easy to show that
Pk−1
rt (k) = j=0 rt−j . This additivity of ccr’s to form multiperiod returns
is very convenient for statistical modelling purposes. To “annualise”
simply divide by k.
Adjusting for inflation: The cc one period real return is
rtreal = ln(1 + Rreal
t ). Using our previous analysis we can show that
rtreal = ln((Pt /Pt−1 )(CPIt−1 /CPIt )). This simplifies further to
rtreal = rt − πt where πt = ln(CPIt /CPIt−1 ) (continuously compounded
one period inflation rate).
Annualising Continuously Compounded Returns: If our investment

horizon is one year for instance then the annual ccr is simply the sum of
the twelve monthly ccr’s as in P11
rA = rt (12) = rt + rt−1 + . . . + rt−11 = j=0 rt−j .
Define the average continuously compounded monthly return to be
1
P11
r̄m = 12 j=0 rt−j
P11
Notice that 12r̄m = j=0 rt−j so that we may alternatively express rA as
rA = 12r̄m . That is, the cc annual return is 12 times the average of the
cc monthly returns.
Stylised Facts
How do time series of stock returns behave?

Are there any commonalities across the observed returns around the
world?
Why important? We want to think about simple models for describing
the dynamics of stock returns. Reasoning: Look at how returns behave
and tailor your model in a way to capture the observed stylised facts.
Example: rt = µ + et .
Stylised Facts
Is the normal distribution a reasonably good approximation to the

distribution of stock returns?
Approach: Think of important features of the normal distribution: Tails,
Skewness, Kurtosis etc. Do the data match these features?
Sk[Y] = E[(Y − E[Y])3 /σ 3 ] (Normalised third moment of random

variable Y with mean E[Y] and variance σ 2 )
Ku[Y] = E[(Y − E[Y])4 /σ 4 ] (Normalised fourth moment of random

variable Y with mean E[Y] and variance σ 2 )
PT
Sample Counterparts: Sk = TS1 3 t=1 (yt − ȳ)3 and
PT
Ku = TS1 4 t=1 (yt − ȳ)4 where S2 is the sample variance. Excess
Kurtosis: Ku − 3.
Stylised Facts
The notion of skewness captures any bias (or lack of symmetry) in the
dispersion of the random variable. Measure of asymmetry. There are 3
possibilities:
(i) The distribution of Y exhibits no skewness. In this case the distribution is
symmetric around its mean. For a symmetric distribution the mean,
median and mode are all equal.
(ii) Positive Skewness: Long Tail to the Right (mean > median > mode).
(iii) Negative Skewness: Long Tail to the Left (mean < median < mode).
Stylised Facts
A distribution with positive kurtosis is called leptokurtic. In terms of

shape, a leptokurtic distribution has a more acute ”peak” around the
mean (that is, a higher probability than a normally distributed variable
of values near the mean) and fat tails (that is, a higher probability than a
normally distributed variable of extreme values). Think about market
crashes.
A distribution with negative kurtosis is called platykurtic. In terms of

shape, a platykurtic distribution has a smaller peak around the mean
(that is, a lower probability than a normally distributed variable of
values near the mean) and thin tails (that is, a lower probability than a
normally distributed variable of extreme values).
Stylised Facts
How to formally test for symmetry, tail thickedness, normality?
ˆ
p
Symmetry: use Sk/ 6/T → N(0, 1).
ˆ − 3)/
p
Tail thickedness: use (Ku 24/T → N(0, 1).
Practical Implementation: Construct the test statistics and reject if the

numerical value falls beyond the cutoffs from N(0,1). If you are
conducting the testp at 5% for instance (2-tails)
p use 1.96 as the cutoffs
ˆ
i.e. reject if |Sk/ 6/T|1.96 and |(Kuˆ − 3)/ 24/T| > 1.96
Stylised Facts
Jarque Bera test for normality
ˆ2+
ˆ 2

T (Ku−3)
JB = 6 Sk 4
Under the null of normality JB → χ22 . Reject H0 when JB > χ2 (2)α% .

Stylised Facts
Is it reasonable to assume that simple net returns are normally

distributed? Empirical Evidence: Observed returns have typically much
fatter tails than the Normal. Also some skewness. Common Belief:
Normality is not a suitable assumption. Note: At this stage we are
thinking of Rt = (Pt − Pt−1 )/Pt−1 .
Since many financial assets exhibit limited liability (the largest possible
loss is the total investment) the normal distribution (which assumes that
y varies from minus to plus infinity) is not appropriate. Think that a
lower bound on returns is −1.
If single period returns are assumed to be normal, then multiperiod
returns cannot be normal (since they are the products - not sums - of
single period returns). Example: (1 + Rt (2)) = (1 + Rt )(1 + Rt−1 ). If
daily returns were normal then multiperiod returns would be the product
of normals! Also looking at empirical data suggests that returns show
greater kurtosis (fatter tails) than expected with a normal distribution.
Stylised Facts
Why normality is not appropriate: Let Rt denote the simple annual

return on an asset, and suppose that Rt = N(0.05, (0.50)2 ). Because
asset prices must be non-negative Rt must always be larger than -1
(indeed note that Rt < −1 means PPt−1 t
< 0). However, based on the
assumed normal distribution Pr(Rt < −1) = 0.018. That is, there is a
1.8% chance that Rt is smaller than -1. This implies that there is a 1.8%
chance that the asset price will be negative. This is why the normal
distribution is not appropriate for simple returns.
A Way out? What about considering continuously compounded (log)
returns. Recall we defined the ccr at time t as rt = ln(Pt /Pt−1 ). Instead
of the simple returns what if we assume that rt is normal, say
rt → N(µ, σ 2 ). What does this imply for simple returns and why is this
not a bad idea?
Stylised Facts
Recall rt = ln(1 + Rt ) and suppose rt → N(0.05, (0.50)2 ). Unlike the

simple net return the ccr can take values less than -1. For example
suppose that rt = −2. From −2 = ln(1 + Rt ) we have that
Rt = e−2 − 1 = −0.865. Then
P(rt < −2) = P(Rt < −0.865) = 0.00002.
Assuming that the ccr are normally distributed implies that single
period gross simple returns are distributed as LogNormal random
variables since rt = ln(1 + Rt ).
LogNormal Distribution: The log-normal distribution is the

probability distribution of any random variable whose logarithm is
normally distributed. If X is a random variable with a normal
distribution, then exp(X) has a log-normal distribution; likewise, if Y is
lognormally distributed, then log(Y) is normally distributed.
Stylised Facts
X = N(µX , σX2 ). If I write Y = eX it is clear that the logarithm of Y is

normally distributed so that Y is lognormally distributed. We write
Y = Lognormal(µX , σX2 ) with 0 < Y < ∞. Similarly, if we take the log
of Y we have X which is normally distributed. Due to the exponential
transformation, Y is only defined for nonnegative values. It can be
1 2 2 2
shown that E[Y] = eµX + 2 σX and V[Y] = e2µX +σX (eσX − 1).
Recall rt = ln(1 + Rt ) (ccr). If we assume that rt is Normally

Distributed, it means that our simple return 1 + Rt will follow a
lognormal distribution. If ccr’s are normally distributed then the stock
price is lognormally distributed.
Details: rt = ln(1 + Rt ) = ln(Pt /Pt−1 ). Suppose rt = N(0.05, (0.5)2 ).

That is µr = 0.05 and σr = 0.5. Let Rt = (Pt − Pt−1 )/Pt−1 . Clearly
(1 + Rt ) = ert . Since rt is normally distributed, (1 + Rt ) is lognormally
distributed. Notice that the distribution of (1 + Rt ) is only defined for
positive values of (1 + Rt ). This is appropriate since the smallest value
that Rt can take is -1. Using the above formulae we have
1 2
µ1+R = e0.05+ 2 0.5 = 1.191 and σ1+R 2
= 0.563.
Stylised Facts
Assuming that ccr’s are normally distributed is also handy since the
sum of a finite number of normal random variables is normal, the
conceptual problem with multiperiod returns is eliminated. Still,
empirical data suggest that returns show greater kurtosis (fatter tails)
than expected with a lognormal distribution.
Summary: Stock returns exhibit greater kurtosis than the normal or
lognormal routines would suggest. This means that extreme events
(both positive and negative) are observed more often than predicted by
these distributions. Stock returns also exhibit a certain amount of
skewness. Certainly extreme events are more likely to be crashes than
explosions.
A Simple Model
rt = µ + et with et = N(0, σ 2 ). Equivalently: rt = N(µ, σ 2 ).
We need to impose more structure. The above is only stating that the e0t s
are normally distributed. We are not saying anything about the
dependence structure of the e0t s (or rt0 s).
What if we assume et = IID(0, σ 2 ). Here IID stands for independently

and identically distributed. We could do away with the identically
distributed part (since we assume normality) and write NID (i.e
normally independently distributed) i.e rt = NID(µ, σ 2 ).
Is the model rt = µ + et with et = NID(0, σ 2 ) realistic?

A Simple Model
rt = µ + et with et = NID(0, σ 2 ) realistic?
Issues: Normality? Independence? Constant variance?
Independence: For any two functions m1 (.) and m2 (.) it holds that
Cov(m1 (x), m2 (y)) = 0.
Note that Independence of X and Y implies Cov(X, Y) = 0

(uncorrelatedness) but Cov(X, Y) = 0 does not imply independence
because we could have something like Cov(X 2 , Y 2 ) 6= 0 while
Cov(X, Y) = 0. CAUTION! Two variables may be uncorrelated while
being dependent. The same holds for a single variable at two different
time periods.
A Simple Model
rt = µ + et with et = NID(0, σ 2 )
IID also implies that returns are not predictable
Can we relax independence?
A more realistic specification: rt = µ + et with et a zero mean

uncorrelated random variable.
Note that we havent said anything about the volatility of the e0t s.
Imposing uncorrelatedness leaves the door open for phenomena such as
time varying volatility i.e. σt2 = f (σt−1
2
).
A Simple Model
Why is the above model important?
We will try to relate economic theories to simple econometric

specifications so as to be able to test theories.
A model such as rt = µ + et with et uncorrelated rules out

predictability of the rt0 s with their past values, say rt−1 , rt−2 etc. At the
same time it does not rule out dependence in returns in the sense that rt2
2
may still be related to rt−1 . This is very close to what we often observe
when analysing the time series properties of developed market returns.
Particular versions of the Efficient Markets Hypothesis (EMH) may
translate into the above statement.
Useful Remark: rt = ln(Pt /Pt−1 ) → pt = µ + pt−1 + et with et as
above.

Econ3016: Empirical Finance WEEK 3/4

Uploaded by

Copyright:

Available Formats

Econ3016: Empirical Finance WEEK 3/4

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Econ3016: Empirical Finance WEEK 3/4

Uploaded by

Copyright:

Available Formats

ECON3016: EMPIRICAL FINANCE

Consider an investment instrument that returned 10%, 12% and 5%

Simple scenario: we have a stream of annual returns, say R1 = 0.05,

RA is our annualised return (per annum return). Important: We assumed

More formally: Let k be the number of compounding periods and let n

Note: With yearly compounding we have n = 1 and k is the number of

Example: Consider an investment instrument that returned 10%, 12%

Example: Suppose the compounding interval is monthly (n=12), the

Should we use arithmetic averages or geometric averages? Which

In 1999, 2000 and 2001 the SP500 returned R1 = 0.2104,

One Period Returns: The continuously compounded return rt is defined as

Intuition: Take the exponential of both sides, we get ert = 1 + Rt = Pt /Pt−1 .

Disadvantage of ccr’s: When dealing

The cc k-month return is defined as

Annualising Continuously Compounded Returns: If our investment

How do time series of stock returns behave?

Is the normal distribution a reasonably good approximation to the

Sk[Y] = E[(Y − E[Y])3 /σ 3 ] (Normalised third moment of random

Ku[Y] = E[(Y − E[Y])4 /σ 4 ] (Normalised fourth moment of random

A distribution with positive kurtosis is called leptokurtic. In terms of

A distribution with negative kurtosis is called platykurtic. In terms of

How to formally test for symmetry, tail thickedness, normality?

Practical Implementation: Construct the test statistics and reject if the

Jarque Bera test for normality

Under the null of normality JB → χ22 . Reject H0 when JB > χ2 (2)α% .

Is it reasonable to assume that simple net returns are normally

Why normality is not appropriate: Let Rt denote the simple annual

Recall rt = ln(1 + Rt ) and suppose rt → N(0.05, (0.50)2 ). Unlike the

LogNormal Distribution: The log-normal distribution is the

X = N(µX , σX2 ). If I write Y = eX it is clear that the logarithm of Y is

Recall rt = ln(1 + Rt ) (ccr). If we assume that rt is Normally

Details: rt = ln(1 + Rt ) = ln(Pt /Pt−1 ). Suppose rt = N(0.05, (0.5)2 ).

rt = µ + et with et = N(0, σ 2 ). Equivalently: rt = N(µ, σ 2 ).

What if we assume et = IID(0, σ 2 ). Here IID stands for independently

Is the model rt = µ + et with et = NID(0, σ 2 ) realistic?

rt = µ + et with et = NID(0, σ 2 ) realistic?

Issues: Normality? Independence? Constant variance?

Note that Independence of X and Y implies Cov(X, Y) = 0

IID also implies that returns are not predictable

Can we relax independence?

A more realistic specification: rt = µ + et with et a zero mean

Why is the above model important?

We will try to relate economic theories to simple econometric

A model such as rt = µ + et with et uncorrelated rules out

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.