Questions Short Concept Questions

The following questions are intended to help candidates understand the material.
They are not actual FRM exam questions.
QUESTIONS
Short Concept Questions

1.1 Can independent events be mutually exclusive? 1.3 W hat are the sam ple spaces, events, and event spaces if
1.2 Can Bayes' rule be helpful if A and B are independent? y ° u are interested in measuring the probability of a corpo-
W hat if A and B are perfectly dependent so that B is a rate takeover and w hether * was hostile friendly?
subset of A?
Practice Questions
1.4 Based on the probabilities in the diagram below, what are a. Pr(Ac)
the values of the following? b. P r (D |A (J B (J C )
c. P r(A |A )
d. Pr(B| A)
e. Pr(C| A)
f. Pr(D| A)
g. P r((A U D )c)
h. Pr(A cn D c)
i. Are any of the four events pairwise independent?
1.6 Continue the application of Bayes' rule to com pute the

probability that a m anager is a star after observing two
years of "high" returns.
a. Pr(A)
b. P r(A |B ) 1.7 There are two com panies in an econom y, Com pany A and
c. Pr(S| A) Com pany B. The default rate for Com pany A is 10%, and
d. P r(A n e n c ) the default rate for Com pany B is 20% . Assum e defaults
e. P r (B |A D Q for the two com panies occur independently.
f. P r (A D B |Q a. W hat is the probability that both com panies default?
g. P r ( A U B |Q b. W hat is the probability that either Com pany A or
h. Considering the three distinct pairs of events, are any Com pany B defaults?
of these pairs independent?
1.8 C red it card com panies rapidly assess transactions
1.5 Based on the probabilities in the plot below, what are the for fraud. In each day, a large card issuer assesses
values of the following? 10,000,000 transactions. O f these, 0.001% are fraudulent.
If their algorithm identifies 90% of all fraudulent transac
tions but also 0.0001% of legitim ate transactions, what
is the probability that a transaction is fraudulent if it has
been flagg ed ?
1.9 If fund managers beat their benchmark at random (meaning

a 50% chance that they beat their benchm ark), and annual
returns are independent, what is the chance that a fund
m anager beats her benchm ark for ten years in a row?
8 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

The following questions are intended to help candidates understand the material. They are not actual FRM exam questions.
ANSW ERS

1.1 No. The probability of observing both must be the prod probabilities, Pr(A)/Pr(B). Here the probability also only
uct of the individual probabilities, and mutually exclusive depends on unconditional probabilities and so never
events have probability 0 of sim ultaneously occurring. changes.
Pr(B| A)Pr(A) 1.3 Using the notation No Takeover (NT), Hostile Takeover
1.2 Bayes' rule says that Pr(A| B) = . If these
Pr(B) (HT) and Friendly Takeover (FT), the sam ple space is {N T,
events are independent, then Pr(B| A) = Pr(B) so that HT, H T}. The events are all distinct com binations of the
Pr(A| B) = Pr(A). B has no information about A and so sam ple space including the em pty set, </>, NT, HT, FT,
updating with Bayes' rule never changes the conditional {N T, H T}, {N T, F T }, {H T, F T }, {N T, HT, F T }. The event
probability. If these events are perfectly dependent, space is the set of all events.
then Pr(B | A) = 1, so that Pr(A | B) is just the ratio of the
Solved Problems
1.4 a. Pr(A) = 12% + 11% + 15% + 10% = 48% c. This is trivially 100%.
b. Pr(A| B) = P r(A rW P r(B ) = (11% + 10%)/ d. P r(B H A ) = 9%. The conditional probability is
(11% + 10% + 16% + 18%) = 38.2%
c. Pr(B| A) = P r(A rW P r(A ) = (11% + 10%)/48%
e. There is no overlap and so P r(C n A ) = 0.
= 43.8%
f. P r(D H A ) = 9%. The conditional probability is 30%.
d. P r (A D B n C ) = 10%
g. This is the total probability not in A or D. It is
e. Pr(B|ADC) = Pr(BnADC)/Pr(AnC)
1 - P r(A U D ) = 1 - (Pr(A) + Pr(D) - P r(A D D ))
= 10% /(15% + 10%) = 40%
= 100% - (30% + 36% - 9%) = 43%.
f. Pr(ADB|C) = Pr(AnenC)/Pr(C)
h. This area is the intersection of the space not in A with
= 10%/(15% + 10% + 18 + 10%) = 18.9%
the space not in D. This area is the same as the area
g. P r (A U B |C ) = P r((A (JB )n C )/P r(C ) that is not in A or D, P r((A (JD )c) and so 43% .
= (15% + 10% + 18%)/53% = 81.1%
i. The four regions have probabilities A = 30%,
h. We can use the rule that events are indepen B = 30% , C = 28% and D = 36% . The only region
dent if their joint probability is the product of the that satisfied the requirem ent that the joint probability
probability of each. Pr(A) = 48% , Pr(B) = 55%, is the product of the individual probabilities is A and B
Pr(C) = 53%, P r(A D B ) = 21% # (48% X 55%), because P r(A D B ) = 9% = Pr(A)Pr(B) = 30% X 30%.
P r(A D C ) = 25% ^ (48% X 53%),
1.6 Consider the three scenarios: (High, High), (High, Low)
P r(B D C ) = 28% ^ (55% X 53%). None of these
and (Low, Low).
events are pairwise independent.
We are interested in Pr (Star\High, High) using Bayes'
1 .5 a. 1 - Pr(A) = 100% - 30% = 70%
P r(High, High] Star)Pr(Star)
rule, this is equal t o -------- ^
P
-'
b. This value is P r(D n (A U B U C ))/P r(A (JB (JC ).
The total probability in the three areas A , B, and Stars produce high returns in 20% of years, and so
C is 73% . The overlap of D with these three is Pr (High, H igh\Star) = 20% X 20% Pr (Star) is still 10%.
9% + 8% + 7% = 24% , and so the conditional prob Finally, we need to com pute Pr (High, High), which is
24% Pr (High, H ig h jS ta r) Pr (Star) + Pr(H/gh, High\ Normal)
ability is 33%.
73%
Chapter 1 Fundamentals of Probability ■ 9

Pr(Normal). This value is 20% X 20% X 10% + 5% X 5% 1.8 We are interested in Pr(Fraud|Flag). This value is
= 90% = 0.625% . Com bing these values, Pr(Fraud D FIag )/Pr(Flag ). The probability that a trans
20% X 20% X 10% action is flagged is .001% X 90% + 99.999% X 0.0001%
.625% = .000999% . The P r(F ra u d n F la g ) = .001% X 90%
This is a large increase from the 30% chance after one year. .0009
= .0009. Com bining these values, = 90% . This
.000999
1.7 a. 0.10*0.20 = 0.02 —»• 2% indicates that 10% of the flagged transactions are not
b. Calculate this in two ways: actually fraudulent.
i. Using Equation 1.1: 1.9 If each year was independent, then the probability of
P r(A U B ) = Pr(A) + Pr(B) - P r(A D B ) beating the benchm ark ten years in a row is the product
P r(A U B ) = Pr(A) + Pr(B) - P r(A D B ) of the probabilities of beating the benchm ark each year,
= 10% + 20% - 2% = 28%.
1Y ° = 1 . 1% .
2 1,024
ii. Using the identity: P r(A U A c) = Pr(A) + Pr(Ac) = 1
Let C be the event of no defaults. Then
Pr(C) = (1 - (1 - Pr(A))*(1 - Pr(B)) 最新cfafrm加微信286982279
= (1 - 10%)(1 - 20%) = 0.9*0.8 = 72%.
Then the com plem ent of C is the event of any
defaults Pr(C c) = 1 - Pr(C) = 28%.

QUESTIONS

2.1 W hat is the key difference between a discrete and a con 2 .5 W hat is excess kurtosis?
tinuous random variable?
2 .6 How are quantile functions related to the C D F? W hen X is
2 .2 W hat do a PM F and a C D F m easure? How are these a continuous random variable, when does this relationship
related? not hold?
2 .3 W hat two properties must a PM F satisfy? 2 .7 W hat are the median and interquartile range? W hen is the
2 .4 How are the mean, standard deviation, and variance of Y median equal to the mean?
related to these values of X when Y = a + bX ?
Practice Questions
2 .8 W hat is E[XE[X]]? Hint: recall that E[X] is a constant. 2 .1 0 Suppose the return on an asset has the following
2 .9 Consider the following data on a discrete random distribution:
variable X:
Return Probability
X -4 % 6%
1 - 2 .4 5 6 -3 % 9%
2 - 3 .3 8 8 -2 % 11%
3 - 6 .8 1 6 -1 % 12%
4 1.531 0% 14%
5 1.737 1% 16%
6 - 1 .2 5 4 2% 15%
7 - 1 .1 6 4 3% 8%
8 1.532 4% 5%
9 2.550 5% 4%
10 0.296 a. Com pute the mean, variance, and standard deviation.
11 - 0 .9 7 9 b. Verify your result in (a) by computing E[X2] directly and

using the alternative expression for the variance.
12 - 4 .2 5 9
c. Is this distribution skew ed?
13 2.810 d. Does this distribution have excess kurtosis?
14 - 1 .6 0 8 e. W hat is the median of this distribution?
15 - 0 .5 7 5
a. Calculate the mean and variance of X.

b. Standardize X and check that the mean and variance
of X are 0 and 1 respectively.
c. Calculate the skew and kurtosis of X.
Chapter 2 Random Variables ■ 23

ANSW ERS

2.1 Discrete random variables have support on only a finite PDF
(or possibly countably infinite) set of points. These points
are all distinct in the sense that it there is always some
number between any two points in the support. A con
tinuous random variable has support over a continuous
range of values. This range can be finite, as in a uniform
random variable; or infinite, as in a standard normal ran
dom variable.
2.2 The PM F m easures the probability of a single value.

The C D F m easures the total probability that the ran
dom variable is less than or equal to a sp ecific value—
that is, the cum ulative probability that X < x for som e
value x. The C D F is the sum of the PM F across all val
ues less than x.
2.3 A PM F must be non-negative and sum across the support

of the random variable to 1.
2.4 E[Y] = E[a + bX]

= E[a] + E[bX] Expectation of sum is the
sum of expectations
= a + bE[X] Expectation of a constant is constant.
The variance of V is V[a + bX] = V[bX] = b2V[X]

because the variance of a constant is 0 and the variance
of b X is b2 tim es the variance of X.
The standard deviation of V is simply the square root of

the variance given above, V b ^ v jx j = |b| v V [ X ] .
2.5 Excess kurtosis is the standard kurtosis measure minus 3. 2.7 The median is the point where 50% of the probability is
This reference value com es from the normal distribution, located on either side. It may not be unique in discrete
and so excess kurtosis m easures the kurtosis above that distributions, although it is common to choose the sm all
of the normal distribution. est value where this condition is satisfied. The IQR is the
2.6 The quantile function is the inverse of the C D F function. difference between the 25% and 75% quantiles. These
That is, if u = Fx(x) returns the C D F value of X, then are the points where 25% and 75% of the probability of
q = F x \u ) is the quantile function. The C D F function is the random variable lies to the left. This difference is a
not invertible if there are regions where there is no prob measure of spread that is like the standard deviation.
ability. This corresponds to a ju m p in the CDF, as shown The median is equal to the mean in any sym m etric distri
in the figures that follow. bution if the first moment exists (is finite).

Solved Problems
2.8 E[XE[X\] = 1 E[X \X Pr(X = x) = E [X ]2 X Pr(X = x) X X Standardized
= E[X] X E[X] = (E[X])2 because E[X] is a constant and
8 1.532 0.898
does not depend on X. In other words, the expected
value of a random variable d o e s not depend on the ran 9 2.550 1.289
dom variable. 10 0.296 0.423
2.9 a. The mean is —0.803 and the variance is 6.762. 11 - 0 .9 7 9 - 0 .0 6 8
b.
12 - 4 .2 5 9 - 1 .3 2 9
X X Standardized 13 2.810 1.389
1 - 2 .4 5 6 - 0 .6 3 6 14 - 1 .6 0 8 - 0 .3 1 0
2 - 3 .3 8 8 - 0 .9 9 4 15 - 0 .5 7 5 0.088
3 - 6 .8 1 6 - 2 .3 1 2 Mean - 0 .8 0 3 0.000
4 1.531 0.897 Variance 6.762 1.000
5 1.737 0.977 Standard 2.600 1.000

Deviation
6 - 1 .2 5 4 - 0 .1 7 3
7 - 1 .1 6 4 - 0 .1 3 9
(X Standardized)3 (X Standardized)4
X X Standardized (Skew) (Kurtosis)
1 - 2 .4 5 6 - 0 .6 3 6 - 0 .2 5 7 0.163
2 - 3 .3 8 8 - 0 .9 9 4 - 0 .9 8 2 0.977
3 - 6 .8 1 6 - 2 .3 1 2 - 1 2 .3 6 4 28.590
4 1.531 0.897 0.723 0.649
5 1.737 0.977 0.932 0.910
6 - 1 .2 5 4 - 0 .1 7 3 - 0 .0 0 5 0.001
7 - 1 .1 6 4 - 0 .1 3 9 - 0 .0 0 3 0.000
8 1.532 0.898 0.724 0.650
9 2.550 1.289 2.143 2.764
10 0.296 0.423 0.075 0.032
11 - 0 .9 7 9 - 0 .0 6 8 0.000 0.000
12 - 4 .2 5 9 - 1 .3 2 9 - 2 .3 4 8 3.120
13 2.810 1.389 2.682 3.726
14 - 1 .6 0 8 - 0 .3 1 0 - 0 .0 3 0 0.009
15 - 0 .5 7 5 0.088 0.001 0.000
Mean - 0 .8 0 3 0.000 -0 .5 8 1 2.773
Variance 6.762 1.000
Standard Deviation 2.600 1.000
Hence, the skew of X is —0.581 and the kurtosis of X is 2.773.
Chapter 2 Random Variables ■ 25

2.10 a. The mean is E[X] = Pr(X = x) = 0.25% . d. The kurtosis requires computing
The variance is Var[X] = £ (x — E[X])2 Pr(X = x)

E[(X - E[X])4
= 0.000555. kurtosis(X)
The standard deviation is \ / Var[X] = 2.355% .

Pr(X = x).
b. E[X2] = 2 x 2 Pr(X = x) = .000561 and so
E[X2] - (E[X])2 = 0.000561 - (.0025)2 = .000555, Thus the kurtosis is 2.24. The excess kurtosis is then
which is the same. 2.24 - 3 = - 0 .7 6 . This distribution does not have
c. The skewness requires computing excess kurtosis.
e. The median is the value where at least 50% probabil

X - ix
skew(X) = I M 1 ! = E ity lies to the left, and at least 50% probability lies to
a
(T
the right. Cum ulating the probabilities into a CDF, this
X — fJL^
occurs at the return value of 0%.
2
3
Pr(X = x).
= a
Thus the skewness is 0.021, and the distribution has a

mild positive skew.

The following questions are intended to help candidates understand the material. They are n o t actual FRM exam questions.
QUESTIONS

3.1 W hat is the name of the distribution that describes any 3.5 How much probability is within ± 2 a of the mean of a nor
random binary event? mal random variable?
3.2 Under what conditions can a binomial and a Poisson be 3.6 How many independent standard normal random vari
approxim ated by a normal random variable? ables are required to construct a
3.3 W hat type of events are Poisson random variables used to 3.7 If a Beta-distributed random variable is used to describe
describe? some data, what requirem ent must the data satisfy?
3.4 If X is a standard uniform, what is Pr(0.2 < Z < 0.4)? W hat 3.8 How are the mean and variance of a mixture of normal
about Pr(/ < X < u), where 0 < / < u < 1? random variables related to the mean and variance of the
com ponents of the m ixture?
Practice Questions
3.9 Either using a Z table or the Excel function N O RM .S.D IST, a. The fund has access to a USD 10 million line of credit
com pute that does not count as part of its portfolio. W hat is the
chance that the firm's loss in a month exceeds this line
a. P r ( - 1.5 < Z < 0), where Z ~ N(0, 1)
b. Pr(Z < - 1 .5 ), where Z ~ N(0, 1) of credit?
b. W hat would the line of credit need to be to ensure
c. Pr(—1.5 < X < 0), where X ~ N(1, 2)
d. Pr(X > 2), where X ~ N(1, 2) that the firm's loss was less than the line of credit in
99.9% of months (or equivalently, larger than the LO C
e. Pr(W > 12), where W ~ N(3, 9)
in 0.1% of months)?
3.10 Either using a Z table or the Excel function N O RM .S.IN V,
com pute
3.13 If the kurtosis of some returns on a small-cap stock portfo
lio was 6, what would the degrees of freedom param eter
a. z so that Pr(z < Z) = .95 when Z ~ N(0, 1) be if they were generated by a generalized Student's t„?
b. z so that Pr(z > Z) = .95 when Z ~ N(0, 1)
W hat if the kurtosis was 9?
c. z so that Pr(—z < Z < z) = .75 when Z ~ N(0, 1)
d. a and b so that Pr(a < X < b) = .75 and
3.14 An analyst is using the following exponential function to
Pr(a < X) = 0.125 when X ~ N(2, 4) model corporate default rates:
3.11 If the return on a stock, R, is normally distributed with a M y) = % exp ( - % ) , y 2: 0,
daily mean of 8%/252 and a daily variance of (20%)2/252, where y is the number of years.
find the values where
a. W hat is the cum ulative probability of default within the
a. Pr(R < r) = .001 first five years?
b. Pr(R < r) = .01 b. W hat is the cum ulative probability of default within the
c. Pr(R < r) = .05 first ten years given that the com pany has survived for
five years?
3.12 The monthly return on a hedge fund portfolio with USD 1
billion in assets is N[.02, .0003). W hat is the distribution of
the gain in a month?

A N SW ERS

3.1 Any binary random variable can be described as a Ber them , so Pr(0.2 < Z < 0.4) = 0.4 - 0.2 = 20% and
noulli by mapping the outcom es to 0 or 1. For exam ple, Pr(/ < X < u) = u - I.
a corporate default (1 for default) or the direction of the
3.5 Form ally this quantity is 95.45% , but this area is com
m arker (1 for positive return). monly approxim ated as 95%.
3.2 A binomial can be well approxim ated by a normal when 3.6 A ^ is defined as the sum of v in dependent standard
np and n(1 — p) are both greater than or equal to 10.
normal random variables, and so v.
A Poisson can be accurately approxim ated by a normal
3.7 Beta random variables are continuous values on the
when the param eter A is large. The approxim ation will
interval [0,1], and so the values must lie in this range.
becom e more accurate as A increases. The standard
recom m endation for approxim ating the Poisson with a 3.8 The mean is the weighted average of the means of the
normal requires A > 1000. com ponents. The variance is more com plicated. The
second non-central moment, E[X 2] of the mixture is the
3.3 Poisson random variables are used for counts— the num
weighted average of the second non-central moments
ber of events in a fixed unit for tim e. For exam ple, the
of the com ponents. The variance is then E[X2] — E[X]2,
number of m ortgage defaults over a month.
which depends the first and second moments of
3.4 In a standard uniform, the probability between the mixture.
two points only depends on the distance between
Solved Problems
3.9 a. 43.3% . In Excel, the command to com pute this value answer to the previous problem by re-centering on
is N O R M .S.D IST(0,TR U E) - N O R M .S .D IS T (-1 .5 ,T R U E ). the mean and scaling by the standard deviation, so
that a = 2 X —1. 15 + 2 and b = 2 X 1.15 + 2. Note
b. 6.7% . In Excel, the command to com pute this value is
N O R M .S .D IS T (—1.5,TR U E). that the formula is a = a X q + / j l , where q is the
quantile value.
c. 20.1% . In Excel, the command to com pute this
value is N O RM .S.D IST((0 - 1)/SQ RT(2),TRUE) 3.11 a. The mean is 0.031% per day and the variance is
- N O R M .S .D IS T ((-1 .5 -1 )/S Q R T (2 ),T R U E ). 1.58 per day (so that the standard deviation is
1.26% per day). To find these values, we trans
d. 24.0% . In Excel, the command to com pute this value
form the variable to be standard normal, so that
is 1 - N O RM .S.D IST((2 - 1)/SQ RT(2),TRU E).
r - /x
Pr(R < r) = .001 = Pr Z < = .001
e. 0.13% . In Excel, the command to com pute this value \ (T /
is 1 - N O RM .S.D IST((12 - 3)/3,TRU E). The value for the standard normal is —3.09
3.10 a. 1.645. In Excel, the command to com pute this value is (NORM.S.INV(O.OOI) in Excel) so that
N O RM .S.IN V(.95). —3.09 X a + /jl = —3.86% .
b. The same idea an be used here where z = —2.32

b. —1.645. In Excel, the command to com pute this value
so that Pr(Z < z) = .01. Transforming this value,
is N O RM .S.IN V(.05).
r = —2.32 X a + /jl = —2.89% .
c. 1.15. Here the tail to the left should have 12.5%
c. Here the value of z is - 1 .6 4 5 so that r = —1.645 X
and the tail to the right should also have 12.5%.
a + /jl = —2.04% . These are all common VaR quan
In Excel, the command to com pute this value is
tiles and suggest that there is a 5% chance that the
-N O R M .S.IN V (.125).
return would be less than 2.04% on any given day, a
d. —0.3 and 4.3. The area of the left and right should
1% change that it would be less than 2.89% , and a
each have 12.5% . These can be constructed using the
Chapter 3 Common Univariate Random Variables ■ 45

one in 1,000 chance that the return would be less than that k [v — 4) = 3{v — 2) so that k v — 4 k = 3v — 6 and
3.86% , if returns were normally distributed. 4 k —6
— 3v = 4 — 6. Finally, solving for v, v = • Plug-
—3
k v k
3.12 a. The monthly return is 2%, and the monthly standard k
2 4 -6
deviation is 1.73% . In USD, the monthly change in gmg in v 18 = 6andi^ = = 5. The
6 -3 3 9 -3
portfolio value has a mean of 2% X USD 1 billion =
kurtosis falls rapidly as v grows. For exam ple, if v = 12
USD 20 million and a standard deviation of 1.73% X
then k = 3.75, which is only slightly higher than the kur
USD 1 billion = USD 17.3 million. The probability that
tosis of a normal (3).
the portfolio loses more than USD 10 million is than
(working in millions) 3.14 a. Fy(5) = 1 — exp
V - 20 b. W e need to divide the marginal probability of default
Pr(V < - 1 0 ) = Pr <
17.3 between years five and ten:
= Pr (Z < - 1 .7 3 )
Fy(10) - Fy(5)
Using the normal table, Pr(Z < - 1 .7 3 ) = 4.18% .
By the SURVIVAL probability through year five
b. Here we work in the other direction. First, we find (1 - result from part a).
the quantile where Pr(Z < z) = 99.9% , which gives
z = - 3 .0 9 . This is then scaled to the distribution of Fy(10) - Fy(5) = exp - exp ( - 1% )
the change in the value of the portfolio by m ultiply
ing by the standard deviation and adding the mean, — exp
Fy( 10) ~ Fy(5)
17.3 X —3.09 + 20 = —33.46. The fund would need
1 - Fy(5)
a line of credit of USD 33.46 million to have a 99.9% exp
change of having a change above this level.
3.13 In a Student's t, the kurtosis depends only on the degree
of freedom and is k = 3 ----—. This can be solved so

v —4

QUESTIONS

4.1 How are the marginal distributions related when two ran 4.6 Under what conditions are conditional expectations and
dom variables are independent? marginal expectations equal?
4.2 W hen are conditional distributions equal to marginal 4.7 If X-| and X 2 are independent continuous random variables
distributions? with PDFs fXl(x 1 ) and fx2 (x 2 )» respectively, what is the joint
4.3 If X-| and X 2 are independent, what is the correlation PD F of the two random variables?
between X-| and X 2? 4.8 If X-| and X 2 both have univariate normal distributions, is
the joint distribution of X-, and X 2 a bivariate normal?
4.4 If X-| and X 2 are uncorrelated (Corr[X-|, X 2] = 0), are they
independent? 4.9 In the previous question, what if X-, and X 2 are
independent?
4.5 W hat is the effect on the covariance between X-| and X 2 of
rescaling X-| by w and X 2 by (1 — w), where w is between 4.10 Are sums of iid normal random variables normally distributed?
0 and 1 ?
Practice Questions
4.11 Suppose that the annual profit of two firm s, one an of the profit of Big Firm when Small Firm either has no
incum bent (Big Firm , X-|) and the other a startup (Small profit or loses money (X 2 < 0)?
Firm , X 2), can be described with the following probability 4.15 The return distributions for the S&P 500 and Nikkei for the
m atrix:
next year are given as:
Small Firm (X2) Return - 1 0 % 0 % + 1 0 %

S&P 500
-U SD 1M USD 0 USD 2M USD 4M Probability 25% 50% 25%
>< -U SD 50M 1.97% 3.90% 0 .8 % 0 .1 % Return -5 % 0 % 8 %
£ Nikkei
•mb USD 0 3.93% 23.5% 1 2 .6 % 2.99% Probability 2 0 % 60% 2 0 %
LL
U) USD 10M .8 % 12.7% 14.2% .6 8 %
5 a. W hat is the joint distribution m atrix if the two return
0 6
USD 100M 0 % 3.09% 6.58% 6.16% series are unrelated?

b. If the actual m atrix looks as follows:
a. W hat are the marginal distributions of the profits of
each firm? S&P 500
b. Are the returns on the two firms independent?
-10% 0% +10%
c. W hat is the conditional distribution of Small Firm's
profits if Big Firm has a $100M profit year? -5 % 15% 1%
4.12 Using the probability m atrix for Big Firm and Small Firm, Nikkei 0% 5% 40%
what are the covariance and correlations between the 8% 6% 9%
profits of these two firm s?
Fill in the missing cells to match the original given
4.13 If an investor owned 20% of Big Firm and 80% of Small marginal distributions.
Firm, what are the expected profits of the investor and the
c. For the m atrix in b, what is the conditional distribution
standard deviation of the investor's profits?
of the Nikkei given a 10% return on the S&P?
4.14 In the Big Firm -Sm all Firm exam ple, what are the condi
tional expected profit and conditional standard deviation

ANSW ERS

4.1 The joint PM F is the product of the two marginals. This is 4.5 Use the properties of rescaled covariance,
a key property of independence. C o v[w X 1f (1 — w)X2] = w(1 — w)Cov[X-|, X 2].
4.2 W hen the random variables are independent. If the condi 4.6 If the two components X-] and X 2 are independent, then the
tional distribution of X i given X 2 equals the marginal, then conditional expectation of any function of X i given X 2 = x 2
X 2 has no information about the probability of different is always the same as the marginal expectation of the same
values in X-|. function. That is, E[g(X-|)|X 2 = x 2] = E[g(X-|)] for any value x 2.
4.3 Independent random variables always have a correlation of 0. 4.7 The joint PD F is fx1 (x 2 (x i/ x 2 ) = fx 1 (x i)^x2 (x 2 ) because when
This statement is only technically true if the variance of both independent the joint is the product of the marginals.
is well defined so that the correlation is also well defined. 4.8 Not necessarily. It is possible that the joint is not a bivari
4.4 No. There are many ways that two random variables can ate normal if the dependence structure between X-| and
be dependent but not correlated. For exam ple, if there is X 2 is different from what is possible with a normal. The
no linear relationship between the variables but they both distribution that generated the data points plotted below
tend to be large in m agnitude (of either sign) at the same has normal marginals, but this pattern of dependence is
tim e. In this case, there may be tail dependence but not not possible in a normal that always appear elliptical.
linear dependence (correlation). The image below plots
realizations from a dependent bivariate random variable
with 0 correlation. The dependence in this distribution
arises because if X-| is close to 0 then X 2 is also close to 0.
- 2 - 1 0 1 2
Xi
4.9 It would be if these are independent, because this is a

bivariate normal with 0 correlation.
4.10 Yes because iid normal random variables are jointly nor
- 2 - 1 0 1 2
Xi mally distributed with 0 correlation, the sum of iid normal

random variable is also normal.
Solved Problems
4.11 a. The marginal distributions are com puted by summing b. They are independent if the joint is the product of the
across rows for Big Firm and down columns for Small marginals. Looking at the upper left cell, 6.77% X 6.70%
Firm. They are is not equal to 1.97%, and so they are not.
c. The conditional distribution is just the row correspond
Big Firm Small Firm ing to USD 100M normalized to sum to 1. The condi
tional is different from the marginal, which is another
- U S D 50M 6.77 - U S D 1M 6.70
dem onstration that the profits are not independent.
USD 0 43.02 USD 0 43.19
- U S D 1M USD 0 USD 2M USD 4M
USD 10M 34.38 USD 2M 34.18
USD 4M % 19.5% 41.6% 38.9%

USD 100M 15.83 15.93
0
Chapter 4 Multivariate Random Variables ■ 59

4.12 First, the two means and variances can be com puted from Finally, the conditional expectation is E[X-||X 2 < 0] =
the marginal distributions in the earlier problem . Then 2x-|Pr(Xi = x-||X2 < 0) = USD 3.01 M. The conditional
the covariance can be com puted using the alternative expectation squared is E [X f|X 2 < 0] = 940.31, and so
form , which is E[X-[X 2\ — E[X-|]E[X2]. The means can be the conditional variance is \/[X 1 ] = E[Xf] — E [X - |]2 =
com puted using E[Xj] = I x j Pr(Xy = xj) for j = 1,2. The 940.31 - 3.01 2 = 931.25 and the conditional standard
variance can be com puted using E [X f] - (E[Xy])2, which deviation is USD 30.52M .
requires computing E [X 2] using X x 2 Pr(X/ = xj). 4.15 a. If the two variables are unrelated, then the joint distri
For Big Firm , these values are E[X-|] = USD 15.88M , bution is just given by the products
E[X?j = 1786.63 and V[X-,] = 1534.36.
fx1(x2(*-w *2) =
For Small Firm , these values are E [X 2] = USD 1.25M , For exam ple, the probability of seeing —10% on the
E [X 2] = 3.98 and V[X2] = 2.41.
S&P and —5% on the Nikkei is 25% *20% = 5%:
The expected value of the cross product is
E [X 1 X 2] = S S x 1 x 2 Pr(X 1 = x-,, X 2 = x 2) = 43.22. S&P 500
The covariance is then 43.22 — 1.25 X 15.88 = 23.37 -1 0 % 0% + 10%

and the correlation is 2 3 .3 7 /(V 2 .4 1 X 1534.36 = 0.384 -5 % 5% 10% 5%
4.13 Here we can com pute these from the com ponents Nikkei 0% 15% 30% 15%
in the previous step. The expected profits are then
8% 5% 10% 5%
E[P] = 20% X E [X n] + 80% X E [X 2] = 4.18. The variance
of the portfolio is then
S&P 500
(20%) 2 Var[X.,] + (80%)2Var[X2] + 2(20% )(80% )Cov[X1, X 2]
-1 0 % 0% + 10%
This value is 0.04 X 1534.36 + 0.64 X 2.41 + 2 X 0.16 X
23.37 = 70.39, and so the standard deviation is USD 8.39M. -5 % 15% 1%
4.14 We need to com pute the conditional distribution given Nikkei 0% 5% 40%
X 2 < 0. The relevant rows of the probability m atrix are 8% 6% 9%
- U S D 1M USD 0 The sum of the rows needs to match the S&P 500 mar
X^Xz
ginal distribution:
- U S D 50M 1.97% 3.90%
USD 0 3.93% 23.5% Return - 1 0 % 0 % + 1 0 %

S&P 500
USD 10M 0 .8 % 12.7% Probability 25% 50% 25%
USD 100M 0 % 3.09% In other words, the x in this column:
The conditional distribution can be constructed by

-1 0 %
summing across rows and then normalizing to sum to unity.
-5 % 15%
The non-normalized sum and the normalized version are
0% 5%
X,|X2 < 0 Non-Normalized Normalized
8% x%
- U S D 50M 5.87% 11.77%
needs to be such that the sum of 15 + 5 + x = 25.
USD 0 27.43% 54.98%
So, x = 5%.
USD 10M 13.5% 27.06%
USD 100M 3.09% 6.19%

Continuing yields: Dividing each entry by the total mass (25% in this case)
S&P 500 S&P 500 Conditional

Distribution
-1 0 % 0% + 10% + 10%
-5 % 15% 4% 1% -5 % 1% 4%
Nikkei 0% 5% 40% 15% Nikkei 0% 15% 60%
8% 5% 6% 9% 8% 9% 36%
The same process works going across the columns, 25% 100%
c. Working with the last column:
S&P 500
+ 10%
-5 % 1%
Nikkei 0% 15%
8% 9%
25%
Chapter 4 Multivariate Random Variables

QUESTIONS

5.1 W hat is the difference between a sam ple mean and the 5.5 W hy is annualization useful, and how are means and stan
population mean? dard deviations annualized?
5.2 W hat is the difference between the standard deviation of 5.6 W hat do skewness and kurtosis measure?
the sample data and the standard error of the sample mean?
5.7 W hat is the skewness and kurtosis of a normal random
5.3 Is an unbiased estim ator always better than a biased variable?
estim ator?
5.8 How are quantiles estim ated?
5.4 W hat is the sam ple analog to the expectation operator, 5.9 W hat advantages do quantiles have compared to moments?
and how is it used?
5.10 W hen dealing with two random variables, how is coskew
ness interpreted?
Practice Questions
5.11 Suppose four independent random variables X h X 2, X 3, and It is hypothesized that the data comes from a uniform dis
X 4 all have mean /z = 1 and variances o f 1 2, 1 2 <2, and 2, tribution, U(0 , b).
1
respectively. What is the expectation and variance of — a. Calculate the sam ple mean and variance.
5.12 Using the same information from question 1, com pute the b. W hat are the unbiased estim ators of the mean and
1 1 variance?
expectation and variance of 2X-| + 2 X 2 + —X 3 + —X 4.
c. Calculate the b in 1/(0, b) using the formula for the
5.13 Using the same information from question 1, com pute the mean of a uniform distribution and the value of the
2 2 1 1
expectation and variance of —X i + —X? + — Xa + — X 4? unbiased sam ple mean found in part b.
^ 5 1 5 ^ 10 3 10 4
d. Calculate the b in 1/(0, b) using the formula for the
5.14 W hat is the effect on the sam ple covariance between X
variance of a uniform distribution and the value of the
and Y if X is scaled by a constant a? W hat is the effect on
unbiased sam ple variance found in part b.
the sam ple correlation?
5.15 An experim ent yields the following data: 5.16 For the following data:
Observation Value
Trial Number Value
1 0.38
1 0 .0 0
2 0.28
2 0.07
3 0.27
3 0.13
4 0.99
4 0.13
5 0.26
5 0 .2 0
0.43
0.23
6
6
7 0.25 a. W hat is the median?

8 0.27 b. W hat is the IQR?
9 0.34
1 0 0.41
1 1 0.60
1 2 0 .6 6
13 0.76
14 0.77
15 0.96
Chapter 5 Sample Moments ■ 79

ANSW ERS

5.1 The sam ple mean is a numerical average of observed common scale factors are 252, 52, 12, and 4 to convert
data. The population mean is the unknown (true) value from daily, w eekly, monthly, and quarterly. These scale
that the sam ple mean is used to approxim ate. factors are applied to the high-frequency mean or variance
to produce annualized means and variances. The standard
When data are iid, the standard error of the mean is ——,
Vn deviation is scaled by the square root of the scale.
where a is the standard deviation of the data. The standard
5.6 Skewness measures the tendency to see larger values in
error measures the estimation error in the sample mean. As
one direction or the other relative to the mean. For exam
the sample size grows the standard error converges to 0. The ple, a distribution with negative skew will tend to produce
(population) standard deviation measures the dispersion in
larger in m agnitude values less than the mean. Kurtosis is
the data. It is a fixed value independent of the sample size.
a m easurem ent of how frequently large observations of
5.3 Not necessarily. W hile an unbiased estim ator is always either sign are m easured.
right, on average, it may be very inaccurate, which would 5.7 The skewness of a normal is 0, and the kurtosis is 3.
be reflected by a large variance. It may be better to use
5.8 Q uantiles are estim ated by sorting the data from sm allest
an estim ator with a small bias but with lower variance. This
to largest and then choosing the values with the sorted
tradeoff may be especially valuable if the bias depends on
index na to measure the quantile.
the sam ple size, decreasing as the sam ple size increases,
and the sam ple size is large. 5.9 There are two potential advantages. First, quantiles are
always well defined (exist) even when random variables
5.4 The sam ple analog replaces the expectation opera
are very heavy-tailed. Second, quantiles are less sensitive
tor with the average so that E[(g(X))] is approxim ated
to outliers than moments.
by n~ 1 ^ g (x ,). The sam ple analog is a universally used
approach to estim ating moments of any order. 5.10 Coskew ness measures the likelihood of one variable tak
ing a large directional value w henever the other variable
5.5 Annualization transform s moments measured using
is large in m agnitude. W hen there is no sensitivity to the
data sam pled at any frequency— daily, w eekly, monthly,
direction of one variable to the m agnitude of the other,
quarterly— to a value that is equivalent under an assum p
the coskewness measures are .
tion to a measure com puted from annual data. The
0
Solved Problems
5.13 The expectation is com puted using the same steps:
5.11 The expectation is E —(Xi + X 2 + X 3 + X 4 ) = 2 2 1 1
-X-, + - x 2 + — x 3 + — x 4
1 5 1 5 1 10 3 10 4
+ E [X 2] + E [X 3] + E [X 4]) — — (/x + /x + /x + /x) — /x — 1
= | E [X , ] + | e [X2] + ^ £ [ X 3] + ^ E [ X 4]
The variance is Var — (X-i + X 2 + X 3 + X 4) = T ( V a r [ X , ]
2 2 1 1
5/x + 5 /x + 1 0 jJL + — /JL = /X = 1.
1 0 ‘
+ Var[X2] + Var[X3] + Var[X4]) — — ( — + 2 + ^ + / ~ ^\6
This estim ator is unbiased.
5.12 The expectation is E 2 Xy + 2 X 2 + —X 3 + —X 4 = 2\± The variance of the sum is the sum of the variances
because the random variables are independent.
1 1
+ 2/x + —/jl + = 5/x = 5. The variance is 2Xi + 2X 2 2 2 1 1
Var 77X3 + —-x4
— f x V a r
5 X l + 5 Xa + *10 6 1 0 4
1 1 1 1
+ X 3 + ~X 4 = 4Var[X-,] + 4V ar[X 2] + ~Var[X3]
— Trz’VarfX'i ] + 7rir\/ar[X2] + Var[X3] + - 7 —Var[X4]
77
2 6 2 4
25 1J 25 2 100 1 0 0
1 4 4 2 2 4 1 4 1 1 1
+ - V a r[X 4] = - + - + - + - = 5 .
4 4J 2 2 4 4 — X - + — x - + x 2+ x 2=
25 2 25 2 1 0 0 1 0 0 25

1 — From here, we anchor the first observation at 0% and

5.14 The sample covariance is defined — ..(X,- — X)(Y,- — Y ).
n ' the last at 1 0 0 %, equally spacing the in-between val
Multiplying each X,- by a constant a will rescale the
ues. The divisions here will be 10 0 /n —1 = 20%:
1 — —
sam ple mean by a so that —^ ;= 1 (aX(- — aX)(Y, — Y) =
Ranked
- 2 ( = i a<x i - - V) = a - - The Position Value Rank
final covariance is a tim es the original covariance. The 1 0.26 0 %
(JX Y
correlation is then p = /V /V so that scaling X,- by a will 2 0.27 %
OXCTy
2 0
3 0.28 40%
produce p = „ , and the sam ple correlation is
dC Ty(Ty CTy(Ty 4 0.38 60%
invariant to rescaling the data.
5 0.43 80%
5.15 a. Use the standard formula to get the sam ple variance
6 0.99 1 0 0 %
(here, n = 15):
n The median (50%) lies exactly half way between the
jl = n“ 1 2 */ = 0-39, third and fourth ranked observations. Therefore:
n
0.28 + 0.38
a 2 = n“ 1 2 ( X / - A ) 2 = 0.08. Median = ---------------= 0.33
/=1
b. The sam ple mean is already unbiased. b. This requires calculating the 25% and 75% levels
For the variance: The 25% level is 5/20 = 25% of the way between
ranked observations 2 and 3. Therefore:
n
s2 = ct 2 = l 5 0.080 = 0.086.
n - 1 14 q25% = 0-75 * 0.27 + 0.25 * 0.28 = 0.2725
c. The mean for a U(a,b) distribution is given as: Sim ilarly, the 75% level is 75% of the way between
a + b observations 4 and 5:
q75% = 0.25 * 0.38 + 0.75 * 0.43 = 0.4175

+ b
Thus, IQR = (0.2725, 0.4175).
0
0.385 = — ------->b = 0.77.
d. The variance for a U(a,b) distribution is given as
2 _ (b - a ) 2
fJ “ 12
b2
0.086 = — ^ b = 1.016
12
5.16 a. The first step is to rank order the observations:
Ranked
Position Value
1 0.26
2 0.27
3 0.28
4 0.38
5 0.43
6 0.99
Chapter 5 Sample Moments ■ 81

Hypothesis Testing
Learning Objectives
A fter com pleting this reading you should be able to:
Construct an appropriate null hypothesis and alternative • Understand how a hypothesis test and a confidence
hypothesis and distinguish between the two. interval are related.
Construct and apply confidence intervals for one Explain what the p-value of a hypothesis test measures.
sided and two-sided hypothesis tests, and interpret
the results of hypothesis tests with a specific level of Interpret the results of hypothesis tests with a specific
confidence. level of confidence.
Differentiate between a one-sided and a two-sided test Identify the steps to test a hypothesis about the difference
and identify when to use each test. between two population means.
Explain the difference between Type I and Type II errors Explain the problem of multiple testing and how it can
and how these relate to the size and power of a test. bias results.
最新cfafrm加微信286982279
83
QUESTIONS

6.1 W hat role does the alternative hypothesis play when test 6.6 W hat is the critical value in a hypothesis test?
ing a hypothesis?
6.7 W hat does the p-value of a hypothesis m easure?
6.2 W hen is a one-sided alternative useful? 6.8 How can a confidence interval be used when testing a
6.3 W hat is the size of a hypothesis test? hypothesis?
6.4 W hat are the trade-offs when choosing the size of a test? 6.9 W hat does the VaR of a portfolio measure?
6.5 W hat is the power of a hypothesis test? 6.10 W hat are three m ethods to evaluate a VaR model of a
portfolio?
Practice Questions
6.11 Suppose you wish to test w hether the default rate of 6.15 You collect 50 years of annual data on equity and bond
bonds that are rated as investm ent grade by S&P is returns. The estim ated mean equity return is 7.3% per
the same as the default rate on bonds rated as invest year, and the sam ple mean bond return is 2.7% per year.
ment grade by Fitch. W hat are the null and alternative The sam ple standard deviations are 18.4% and 5.3% ,
hypotheses? W hat data would you need to test the null respectively. The correlation between the two-return
hypothesis? series is - 6 0 % . Are the expected returns on these two
assets the sam e? Does your answer change if the correla
6.12 Using the Excel function N O R M .S.D IST or a normal prob
ability table, what are the critical values when testing the tion is 0 ?
null hypothesis, H0: /z = /z0, against 6.16 If a p-VaR model is well specified, HITs should be iid
a. a one-sided lower alternative using a size of 1 0 %? Bernoulli^ — p). W hat is the probability of observing two
HITs in a row? Can you think of how this could be used to
b. a one-sided upper alternative using a size of 2 0 %?
c. a two-sided alternative using a size of 2 %? perform a test that the model is correct?
d. a two-sided alternative using a size of . 1 %? 6.17 A data m anagem ent group wants to test the null hypoth
esis that observed data is N(0,1) distributed by evaluating
6.13 Find the p-value for the following t-test statistics of the
the mean of a set of random draws. However, the actual
null hypothesis, H0: /z = /z0.
underlying data is distributed as N(1, 2.25).
a. Statistic: 1.45, Alternative: Two-sided
b. Statistic: 1.45, Alternative: One-sided upper a. If the sam ple size is 10, what is the probability of a
Type II error and the power of the test? Assum e a 90%
c. Statistic: 1.45, Alternative: One-sided lower
d. Statistic: —2.3, Alternative: One-sided upper confidence level on a two-sided test.
b. How many sam ples would need to be taken to reduce
e. Statistic: 2.7, Alternative: Two-sided
the probability of a Type II error to less than 1%?
6.14 If you are given a 99% confidence interval for the mean
return on the Nasdaq 100 of [2.32% , 12.78% ], what is the
6.18 Suppose that for a linear regression, an estimation of
sam ple mean and standard error? If this confidence inter slope is stated as having a one-sided p-value of 0.04.
W hat does this mean?
val is based on 37 years of data, assumed to be iid, what is
the sam ple standard deviation?
Chapter 6 Hypothesis Testing ■ 95

ANSW ERS

6.1 The alternative specifies the range of values where the 6.7 The p-value measures the probability that the observed
null should be rejected. In most tests, the alternative is data would have been generated if the null is true. It is
the natural com plem ent to the null, although this does the size where the test would just reject the null in favor
not have to be the case. It is possible to specify a test of the alternative.
where neither the null nor the alternative is true.
6.8 If the null is outside of the confidence interval, then it
6.2 A one-sided alternative is helpful when there is some should be rejected in favor of the alternative.
guidance about the value of the param eter when the null 6.9 The VaR is a measure or the m agnitude of the loss that
is false. A leading exam ple occurs when testing whether
the portfolio will lose with some specified probability
a risk premium is 0 against a one-sided alternative that (e.g ., 5%) over some fixed horizon (e.g ., one day or one
the premium is positive. w eek). The p-VaR is form ally defined as the value where:
6.3 The size is the probability of wrongly rejecting the null Pr(L > VaR) = 1 - p, (6 . 6 )
when it is in fact true. This is equivalently the probability
where L is the loss of the portfolio of the selected time
of a Type I error.
horizon and 1 — p is the probability that a large loss
6.4 A small test size reduces the probability of a Type I error,
occurs.
that is, rejecting the null when it is true. It also lowers the
power of the test, that is the chance of rejecting a false null 6.10 First, the exact distribution that com putes the exact
distribution of the number of VaR violations (HITs) over
because smaller sizes correspond to larger critical values.
some tim e period. Second, the asym ptotic method that
6.5 The power of a test the probability of rejecting a false exam ines the average number of violations and uses the
test when the alternative is true. It is one minus the prob
CLT to constrict the asym ptotic distribution. Finally, the
ability of a Type II error.
likelihood ratio that exploits the Bernoulli distribution
6.6 The critical value determ ines the rejection region for a that underlies the 0-1 H IT variables.
test. Test statistics larger than the critical value indicate
the null should be rejected in favor of the alternative.
Solved Problems
6.11 The null is that the IG (investm ent grade) default rate d. This size corresponds to the same critical value as a
is the same for S&P rated firms as it is for Fitch-rated .05% upper-tailed test, which is 3.29.
firms. If Sj are default indicators for S&P rated firms, and
6.13 a. In a two-sided test, the p-value is the area under
Fj are default indicators for Fitch-rated firm s, then the the normal curve for values greater than 1.45 or less
null is H q : As = A f , which states that the mean values than - 1 .4 5 . This is tw ice the area less than - 1 .4 5 , or
are the sam e. The null can be equivalently expressed 2 X 0.074. = .148.
as H0: a s — A f = 0. The alternative is Hy. /j l $ ^ a f or b. In a one-sided upper test, we need the area above
equivalently H^: /xs — a f ^ 0 - The data required to test 1.45. This is half the previous answer, or 0.074.
this hypothesis would be binary random variables where
1 indicates that an IG bond defaulted and 0 if it did not
c. In a one-sided lower test, we need the area under
the curve for values less than 1.45. This is just
within a fixed tim e fram e (e.g ., a quarter).
1 - 0.074 = .926.
6.12 a. The Iower-tailed critical value is —1.28.
d. Here we need to probability under the normal curve
b. The one-sided upper-critical value is 0.84.
for values above —2.3, which is 1 minus the area less
c. A two-sided alternative with a size of 2% uses the than - 2 .3 , or 1 — 0.01 = 0.99.
same critical value as a one-sided upper test, which is
e. Because it is a two-sided alternative, we need that
2.32. Each tail has probability 1% when using the criti
area for values larger than 2.7 or less than —2.7,
cal value.

which is tw ice the area less than —2.7. This value is ± 1.96 using a size of 5%. The null is not rejected. If the
2 X .0034 = .0068. correlation was 0 , then the variance estim ate would be
6.14 The mean is the midpoint of a sym m etric confidence 0.0366, and the test statistic is 1.69. The null would still
not be rejected if the size was 5%, although if the test
interval (the usual type), and so is 7.55% . The 99%
Cl is constructed as [/x — c X & ,jl + c X a] and so size was 10%, then the critical value would be ± 1.645
and the correlation would matter.
c X a = 12.78% — 7.55% = 5.23% . The critical value for
a 99% Cl corresponds to the point where there is 0.5% in 6.16 The probability of a H IT should be 1 — p if the model
5.23% is correct. They should also be independent, and so
each tail, or 2.57, and so a 2.03% needs to
2.57 the probability of observing two HITs in a row should
be surveyed.
be (1 — p)2. The can be form ulated as the null that
6.15 The null hypothesis is H0: /j l e = /j l b . The alternative is Hb: E[H IT i X H ITi+i] = (1 — p ) 2 and tested against the
H-|: fx E ¥= /j l b . The test statistic is based on the difference alternative H-p E[H/T(- X H/T/+1] ^ (1 — p)2. This can be
of the average returns, 8 = 7.3% — 2.7% = 4.6% . im plem ented as a sim ple test of a mean by defining the
The estim ator of the variance of the difference is random variable X/ = HIT} X H/T( + 1 and then testing the
a E + a B - 2(t be, which is 0 .1 842 + 0.0532 - 2 X - 0 .6 X null H0: fJLx = (1 — p ) 2 using a standard test of a mean.
0.184 X 0.053 = 0.0483. The test statistic is
g
6.17 a. W hen the null hypothesis is false, the probability of
—, = 1.47. The critical value for a two-sides test is a Type II error is equal to the probability that the
/O043
hypothesis fails to be rejected, as per the diagram in
V 50
the chapter:
Further recall that for a 90% confidence level on a N(0,1) distribution, the cut-off points are + /—1.65.
-1.65 1.65
Chapter 6 Hypothesis Testing 97
Now, if there are 10 sam ples taken from an N(0,1) In actuality, the true distribution is N (1,2.25), so
then the standard deviation is reduced the a = V 2 .2 5 = 1.5. For a sam ple size of 10, the
expected sam ple standard deviation is
°H> — = = 0.316
V 10
/V
A 5
0.474
O s a m p le
VTo
Therefore, the cut-off points are
Calculating the equivalent distance of + /— 1.65 in this
± 1 .6 5 * 0 .3 1 6 = ± 0 .5 2 2 distribution com pared to a standard N(0,1) yields
Hypothesis Testing for Normal Distribution Parameters
N(0,0.316A2) --------- N(1,0.474A2)
- 0 .5 2 2 - 1 Therefore, the requirem ent becom es

-3 .2 1
0.474
1 - Pr(> right) = 0.01
and
This occurs at a Z-score of (using the Excel function
+ 0.522 - 1 NORM SINV) - 2 .3 2 .
1.00
0.474
-
Accordingly, the following equations need to be

The probability of being on the left-hand side is prac
solved
tically zero. For the right,
Pr(> right) = 1 - $ (- 1 .0 0 ) = 1 - 15.9% = 84.1% . - 2 .3 2

So total probability of a Type II error is 1 - the prob
ability of being in the two tails is
Plugging in K yields:
Pr(Non — R e je c tio n | H0 is false) = 1 - [Pr(< left)
+ Pr(> right)] « 1 - 84.1% = 15.9%
Therefore, the power of the test is 84.1% .
b. The requirem ent is to have

1 - [Pr(< left) + Pr(> right)] = 1%
Clearly, as n increases from 10, the probability of

being in the left-hand tail will only decrease from And because partial observations are not allowed,
already being close to zero. n = 27.

6 .1 8 The p-value tells us the "probability of observing a test will correspond to the value of x that solves the following
statistic that is more extrem e than the one com puted equation:
from the observed data when the null hypothesis is tru e." 1 - O(x) = 0.04
In other words, if the estim ate of the slope is true, then in
Using the excel function N O RM SIN V gives x = 1.75.
a randomized trial the probability of getting a larger new
slope estim ate is 4% .Assum ing a normal distribution, this
Chapter 6 Hypothesis Testing ■ 99

QUESTIONS

7.1 W hat are the three requirem ents of linear regression? 7.4 How is a t-test related to the t-statistic of a param eter?
7.2 You suspect that the CAPM held on all days except 7.5 The R 2 from regressing Y on X is the same as the R2 of
those with a Federal Open M arkets Com m ittee (FO M C) regressing X on Y. True or false?
announcem ent, and on these days the f3 is different. How
7.6 W hat is R2 if /3 = 0?
can a dummy be used to capture this effect? W hat could
you do if you suspected that both a and (3 are different on
7.7 The return on an optimally hedged portfolio is independent
of the return of the hedging instrument. True or false?
FO M C days?
7.8 W hat are the consequences of using a poor benchm ark to
7.3 W hat does a confidence interval for a regression param
evaluate a fund m anager?
eter measure?
Practice Questions
7.9 Find the O LS estim ators for the following data: 7.11 In a CAPM that regresses W ells Fargo on the m arket, the
coefficients on monthly data are a = 0.1 and [3 = 1. 2.
X y W hat is the expected excess return on W ells Fargo when
0 - 1 .4 6 the excess return on the market is 3.5% ?
1 0.35 7.12 You fit a CAPM that regresses the excess return of
Coca-Cola on the excess m arket return using 20 years
2 6.46
of monthly data. You estim ate a = 0.71, (3 = 1.37,
3 4.09 s 2 = 20.38, <xj = 19.82 and jl x = 0.71.
4 7.34 a. W hat are the standard errors of a and /3?
A
5 6.18 b. W hat are the t-statistics for a and (3?

c. W hat is the 99% confidence interval for /§?
6 14.97
7.13 Suppose you estim ated a model that regressed the vol
7 14.28
ume of put and call options traded on the S&P 500 on a
8 2 0 .2 0
measure of overnight news, VOL-, = a + (3NEWSj + e,-.
9 21.24 Your statistical package returned a 90% confidence inter
val for (3 of [0.32,1.89].
7.10 In running a regression of the returns of Stock XYZ
a. Is /3 statistically significant when using a 5% test?
against the returns on the m arket, the standard d evia
b. W hat is the p-value of the t-statistic of (3?
tion for the returns of Stock X YZ is 20% , and that of the
m arket returns is 15%. If the estim ated beta is found to 7.14 If (3 = 1.05 and its 95% confidence interval is [0.43,1.67],
be 0 .7 5 : what is the t-statistic of [3? W hat is the p-value of the
t-statistic?
a. W hat is the correlation between the returns of Stock
X YZ and those of the market? 7.15 How does adding leverage (borrowing at a risk-free rate,
b. If the market falls by 2%, what is the expected return which is not random) to a portfolio change the optimal
on Stock XYZ? hedge ratio? For exam ple, if the portfolio returns were
c. W hat is the maximum possible value of beta given that doubled so that R p = 2Rp, what is the optimal hedge
the standard deviation of the returns of Stock X YZ is ratio for the Rp? Hint: Use the formula for the O LS estim a
2 0 % and those of the m arket is 15%? tor of (3.
116 Financial Risk Manager Exam Part I: Quantitative Analysis

ANSW ERS

7.1 The three requirem ents are that the model is linear in the 7.5 True. The R2 in a regression with a single explanatory
unknown coefficients, the error is additive, and there are no variable is the squared sam plejcprrelation between X
missing values. and Y.
7.2 The model that allowed differences in the slope would 7.6 The only tim e /3 = 0 is when Cov(X,V) = 0 and so the
A
be Rj = a + p R mii + ylFOMCRm,i + €i where If o mc is two variables are uncorrelated. The R2 is the squared cor
1 on FO M C days and 0 otherw ise. If y is not zero, relation and so must be 0 .
then the slope is different on FO M C days. This can be 7.7 False. The optim ally hedged portfolio is uncorrelated
extended to both param eters by estim ating the model
with the return on the hedge. It is not necessarily inde
Ot y IFO M C "f P R m ,i y ^ F O M C ^ m .i "f pendent. It would be independent if the returns were
7.3 The 1 — a confidence interval contains the set of null jointly normally distributed.
hypotheses that could not be rejected using a test size 7.8 If the benchm ark is poor, so that the evaluation model is
of a.
m isspecified, then the slope /3 will be m ism easured. As a
7.4 The t-stat is a t-test of the null H0:f3 = 0 against the alterna result, some of the com pensation for system ic risk (which
tive H-\:(3 9^ 0. is captured by (3) may be attributed to skill a.
Solved Problems
7.9 Doing the basic needed calculations:
X y x —x y-y 1 .x-xf (x -x )(y -y ) ( y - ( a + /be) ) 2
0 - 1 .4 6 - 4 .5 0 - 10 .8 6 20.25 48.85 0.25
1 0.35 - 3 .5 0 - 9 .0 4 12.25 31.66 0.04
2 6.46 - 2 .5 0 -2 .9 3 6.25 7.33 11.41
3 4.09 - 1 .5 0 - 5 .3 0 2.25 7.96 2.31
4 7.34 - 0 .5 0 -2 .0 5 0.25 1.03 0.62
5 6.18 0.50 - 3 .2 2 0.25 -1.61 20.06
6 14.97 1.50 5.57 2.25 8.36 3.19
7 14.28 2.50 4.89 6.25 1 2 .2 2 2.03
8 2 0 .2 0 3.50 10.80 12.25 37.82 3.88
9 21.54 4.50 12.14 20.25 54.64 0.61
A verage 4.50 9.40 8.25 20.83
SUM 82.50 208.25 44.41
Chapter 7 Linear Regression ■ 117

/V X ; = 1 (X; ~ X)(Y; ~ Y) 20.83 confidence interval is then [1.37 — 2.58 X 0.065,1.37

2.52 + 2.58 X 0.065] = [1.20,1.54].
2/=i (x, - x)2 8.25
7.13 a. Yes. The null hypothesis value of H0:(3 = 0 is not in the
a = Y - f3X = 9.40 - 2.52 * 4 .5 0 = - 1 .9 6
confidence interval, and so the null is rejected using a
Continuing to estim ate the variance: test size of 1 0 %.
n
1 1 b. The range of the confidence interval is 2c X se{(3),
= 1 0 - 2 19.87 = 5.55
.2 _
= n - 2 (t where c is the critical value used to construct

i= 1
PxyO-2 ^ the 90% confidence interval (5% in each tail,

PxyVy
7.10 a. (3 = 0.75 = - ^ - - p xy = 56.25%
A
so 1.645). The estim ate of [3 is the m idpoint

0 \
of the confidence interval so 1.105. Because
b. (3 * 2% = 0.75 * 2% = 1.5% /x — 1.645se(/3) = 0.32 and jx + 1.645se(/3) = 1.89,
c. The maximum correlation is 1, therefore: these can be solved for se{(3) = .477. The t-stat is
1 * <T,
A
P
0.20 then = 2.32. Finally, the two-sided p-value is
Pm ax = 1.33 se{(3)
o\ 0.15
2(1 — <E»( 1 1 1 )) = 2.06% , where <I>( •) is the standard
7.11 The expected return is 0.1 + 1.2 X 3.5% = 14.2%. This
normal C D F function. This p-value confirms the rejec
value is the fitted value from the linear regression when
tion in the previous step because it is less than 1 0 %.
the m arket return is 3.5% .
1.67 - 0.43
7.12 a. The standard error for (3 is 7.14 The standard error is — ------—— = 0.316.
2 X 1.96
20.38 The t-stat is then 3.31. The two-sided p-value is
= 0.065.
n X & x 240 X 19.82 2(1 - 0 ( |t |) ) = 0.09% .
The standard error for a is C ov(Rp/ Rh)
7.15 The O LS [3 estim ator is — ^ — . Scaling Rp
20.38 X (0.712 + 19.82) Var[Rh)
0.295.
240 X 19.82 by a leverage v, the optimal hedge would be
Cov(vRp, Rh) CoviRp, Rh)
1.37 — \r-u D \
— = v xr u D \ = VP ' where P ls the led9e
The t-stat for (3 is 20.9. Var(Rh) Var(Rh)
0.065
ratio from the unlevered portfolio. Levering up increases
0.71
The t-stat for a is 2.40. exposure to risk that can be hedged and to the hedge
0.295
ratio must account for this.
c. A 99% two-sided confidence interval uses the
critical value from a one-sided test with a size
of 1% (0.5% in each tail), which is 2.57. The

QUESTIONS

8.1 How does the correlation between pairs of explanatory 8.5 You estim ate a regression model
variables affect the interpretation of the coefficients in a Yj = a + f3-\X-| j + i82X 2l + s,. Using the F-stat of the model,
regression with multiple explanatory variables? you reject the null H0 :/31 = yS2 = 0 but fail to reject either
8.2 Both R2 and R2 only increase when adding a new variable. of the nulls H q :/3^ = 0 or H q :/32 = 0 using the t-stat of the
coefficient. Which values of p = C o rr[X 1f X 2] make this
True or false?
scenario more likely?
8.3 W hen is R2 less than 0? Can this measure of fit be larger
than 1 ? 8.6 Suppose you fit a model where both t-statistics are exactly
+ 1.0. Could an F-statistic lead to rejection in this circum
8.4 The value of an F-test can be negative. True or false?
stance? Use a diagram to help explain why or why not.
Practice Questions
8.7 Construct 95% confidence intervals and p-values for the 8.11 Using the data below:
coefficients in the shipping containers industry portfolio
returns.
Trial Number y *1 *2
8.8 The models were estimated on monthly returns. W hat is the 1 - 5 .7 6 -3 .4 8 -1 .3 7

estimate of the annualized a for the beer and liquor indus 2 0.03 - 0 .0 2 -0 .6 2
try portfolio? W hat is the t-statistic for the annualized a?
3 -0 .2 5 - 0 .5 -1 .0 7
8.9 Suppose Yj and X , followed a bivariate normal so that:
4 - 2 .7 2 -0 .1 8 - 1.0 1
V; -0 .8 2 0.39
5 -3 .0 8
6 -7.1 -2 .0 8 1.39
W hat is the R2 in a model that regresses Y on X? W hat
7 -4.1 - 1 .0 6 0.75
would the R2 be in a model that regressed X on V?
0.14 -0 .6 3
8.10 A model was estim ated using daily data from the S&P 500
8 0 .0 2
from 1977 until 2017 which included five day-of-the-week 9 -6 .1 3 - 1.6 6 1.31
dummies (n = 10,087). The R2 from this regression was 10 0.74 0 .6 8 -0 .1 5
0.000599. Is there evidence that the mean varies with the
a. A pply O LS linear regression to find the param eters for
day of the w eek?
the following equation:
Y, = a + p-\XVl + (32X 2i + €j
b. W hat is the R2, R2, and F statistic of this regression?

ANSW ERS

8.1 Regression coefficients are interpreted as holding all other
variables constant, and so the correlation between the regres
sors is immaterial to the interpretation of the coefficients.
8.2 False. R2 will always increase. R2 can decrease if the

degree-of-freedom adjusted residual variance estimator,
sf
— , does not decrease when adding a new regressor.
This can happen because k is the number of regressors
in the model and so if the num erator does not change,
adding an additional regressor will increase k and so
decrease the denominator.
8.3 R2 is always less than R2 and so cannot be larger than 1. 8.6 Yes. When the regressors are very positively correlated,
It can be sm aller than 0 if the model explains virtually then this can happen. The t-stats are small because the
nothing and the degree-of-freedom loss is sufficient to variables are highly co-linear, so that the variation in the
push its value negative. left-hand-side variable cannot be uniquely attributed to
8.4 False. An F-test measures the reduction in the sum of either. However, the F-stat is a joint test and so if there is
squared residuals that occurs in the expanded model and an effect in at least one, the Fe an reject even if the t-stats
depends on SSRu — S S R r . This value is always positive do not. See the image below that shows the region where
because the unrestricted model must fit better than the the F would fail to reject, and the two t-stats.
restricted model.
8.5 This is most likely to occur when the regressors are highly
correlated. If the regressors are positively correlated, then
the parameter estimators of the coefficients will be nega
tively correlated. If both values are positive, this would lead
to rejection by the F-test. Similarly, if the regressors were
negatively correlated, then the estimators are positively
correlated and the F will reject if one t is positive and the
other is negative. The figure below shows the case for pos
itively correlated regressors. The shaded region is the area 3 •-------- •-------- --------- 1-------- 1-------- •
where the F would fail to reject. The t-stats are outside this -3 - 2 - 1 0 1 2 3
area even though neither is individually significant.
Solved Problems
8.7 The confidence interval is ft ± 1.96 X se. The p-value is
2(1 — <E>(| 1 1)) where <I>( -) is the standard normal CDF.
Coefficient Estimate t-stat Std. Err. Conf. Int. P-value
8.8 The annualized a is 12a = 12 X 0.517 = 6.204% . The
Pm 1 . 0 1 1 18.63 0.054 [0.905, 0.000 t-stat is unaffected because the standard error is also
1.117]
scaled by 1 2 .
- - 0 .1 6 0.075 [- 0 .1 5 9 , 1.128
Ps 0 .0 1 2
8.9 Because this is a regression with a single explanatory
0.135]
variable, the R2 is the squared correlation. The correla
Pv 0.276 3.51 0.079 [0 . 1 2 1 , 0 .0 0 2
tion is 0 .9 / ^ 3 = .519 and so the R2 = 0.27. The R2 is
0.431]
Chapter 8 Regression with Multiple Explanatory Variables ■ 135

the same in both regressions because it only depends on

4 - 2 .7 2 -0 .1 8 - 1.0 1
the correlation between the variables.
5 -3 .0 8 -0 .8 2 0.39
8.10 The model estim ated is
6 -7.1 -2 .0 8 1.39
Y, = /3 -|D-| + (32D2 + /33 D 3 + /34 D 4 + /34 D 5 + e„ where
7 -4.1 - 1 .0 6 0.75
Dj is a dummy that takes the value 1 if the index of the
w eekday is / (e.g ., Monday = 1, Tuesday = 2, . . . ). 8 0.14 0 .0 2 -0 .6 3
The restriction is that H q :/3-\ = = = /34 = ySs 9 -6 .1 3 - 1.6 6 1.31
so there this is no day-of-the-week effect.
0.74 -0 .1 5
This model can be equivalently written as
10 0 .6 8
Yj = /jl + 82D2
S3D3 + 84D4 + S4D5 + Sj, therefore,
+ A verage -2 .8 2 3 -0 .9 1 0 - 0 .1 0 1
here the null is H0:82 = 83 = 84 = 8 5 . In the two m odels,

H = /3-|, and /jl + 8, = (3,. The second form of the model
is a more standard null for an F-stat.
Trial Number y - y x1 — x1 *2 “ *2
The F-stat of the regression is
1 - 2 .9 3 7 -2 .5 7 0 -1 .2 6 9
^0.000599
R 2- 0/
4 2 2.853 0.890 -0 .5 1 9
1.51. The distribution is
1- R 2 1 -0.000599
/n —5 / 3 2.573 0.410 -0 .9 6 9
/(10087 —5)
an ^4 , 1 0 0 8 2 and the critical value using a 5% size is 2.37. 4 0.103 0.730 -0 .9 0 9
The test statistic is less than the critical value, therefore, 5 -0 .2 5 7 0.090 0.491
the null that all effects are 0 is not rejected.
6 -4 .2 7 7 -1 .1 7 0 1.491
8.11 a. Start by setting up the basic calculations: 7 -1 .2 7 7 -0 .1 5 0 0.851
Trial Number y *1 *2 8 2.963 0.930 -0 .5 2 9
1 - 5 .7 6 - 3 .4 8 -1 .3 7 9 -3 .3 0 7 -0 .7 5 0 1.411
2 0.03 - 0 .0 2 -0 .6 2 10 3.563 1.590 -0 .0 4 9
3 -0 .2 5 - 0 .5 -1 .0 7 A verage
N>
Trial Number (*! - X t )2 (x2 - x2)2 (x-, - Xt )(y - y) (x2 - x2)(y - y) (*! - )(*2 - * 2)
<<
I
1 8.626 6.605 1.610 7.548 3.727 3.261
2 8.140 0.792 0.269 2.539 -1.481 -0 .4 6 2
3 6.620 0.168 0.939 1.055 -2 .4 9 3 -0 .3 9 7
4 0.011 0.533 0.826 0.075 -0 .0 9 4 -0 .6 6 4
5 0.066 0.008 0.241 -0 .0 2 3 -0 .1 2 6 0.044
6 18.293 1.369 2.223 5.004 -6 .3 7 7 -1 .7 4 4
7 1.631 0.023 0.724 0.192 -1 .0 8 7 -0 .1 2 8
8 8.779 0.865 0.280 2.756 -1 .5 6 7 -0 .4 9 2
9 10.936 0.563 1.991 2.480 -4 .6 6 6 -1 .0 5 8
10 12.695 2.528 0.002 5.665 -0 .1 7 5 -0 .0 7 8
A verage 7.580 1.345 0.911 2.729 -1 .4 3 4 -0 .1 7 2

This gives y 0 = fiy - rj 1/xx, = - 2 .8 2 3 - (2 .0 2 9 )(-0 .9 1 ) = - 0 .9 7 6
jxy = —2.823 /Xx, = —0.91 /zx2 = —0.101 So:

cry = 7.58 ox, = 1.345 o$2 = 0.911
Trial
& x,x = - 0 .1 7 2 (Jy(Jxy = 2.729 o-yox = -1 -4 3 4 ~2
2 2
Number k *2 kx2 *2
First regressing on the last independent variable: 1 2.274 -1 .5 9 8 -3 .6 3 5 2 .5 5 5
X1(- — S'0
n + 5iX->; + X 1/
1^2/ 2 1.047 -0 .4 0 6 -0 .4 2 5 0.164
X2 - 0 .1 7 2
0.189 3 1.741 -0 .9 1 7 -1 .5 9 6 0.841
0.911
4 -1 .3 7 8 -0 .8 1 6 1.125 0.666
5q = /xXl - ^i Mx2 = -0-91 - 0.189*(—0.101) = -0.929
5 -0 .4 4 0 0.502 -0.221 0.252
Vi ~ To + 71^2/ + 6 -1 .9 0 5 1.341 -2 .5 5 4 1.798
- 1 .4 3 4
- 1 .5 7 4 7 -0 .9 7 4 0.831 -0 .8 0 9 0.691
0.911
8 1.076 -0 .4 1 0 -0 .4 4 2 0.168
7o = Mr - fiM x2 = - 2 .8 2 3 - (—1 .574*—0.101 ) = - 2 .9 8 2
9 -1 .7 8 7 1.315 -2 .3 4 9 1.728
Trial 10 0.338 0.154 0.052 0.024
Number ~2
y xi y*1 xi A verage -1 .0 8 5 0.889
1 - 4 .9 3 4 -2 .8 1 0 13.865 7.896
and
2 2.036 0.792 1.612 0.627
C o v ( K ,X 2) = - 1 .0 8 5 o-| = 0.889
3 1.048 0.227 0.238 0.051
- 1 .0 8 5
4 -1 .3 2 8 0.558 -0.741 0.311 = - 1.221
0.889
f t =
5 0.516 0.183 0.094 0.033 Tying things together:

6 -1 .9 3 0 -0 .8 8 8 1.715 0.789 V, - ftX -,, - /32X 2( = a + e, = - 1 .8 7 3 X 1; + 1.221 X 2l
7 0.063 0.011 0.001 0.000
8 2.130 0.830 1.768 0.689 Trial Number a + €j 4
9 -1 .0 8 6 -0 .4 8 3 0.525 0.234 1 -0 .9 1 3 0.328 0.108
10 3.486 1.581 5.510 2.498 2 -0 .6 8 9 0.553 0.306
A verage 2.459 1.313 3 -0 .6 1 9 0.623 0.388
4 -3 .6 1 5 -2 .3 7 3 5.632
C o v (Y ,X J 2.459 5 -1 .0 6 8 0.173 0.030
1.873
1.313 6 -1 .5 0 8 -0 .2 6 7 0.071
Repeating the process for X 2: 7 -1 .2 0 0 0.042 0.002
*2 i = io + + x 2i 8 -0 .6 6 6 0.576 0.331
- 0 .1 7 2 9 -1 .4 2 3 -0.181 0.033
- 0 .1 2 8
1.345
10 -0 .7 1 7 0.525 0.276
£0 = Mx 2 - ^iM x, = -0 .1 0 1 - (- 0 .1 2 8 ) - (- 0 .9 1 ) = - 0 .2 1 7
A verage -1 .2 4 2 0.718
Yj = fjo + + K So a = - 1 .2 4 2 , and the variance of the residuals
2.729 is 0.718.
2.029
1.345
Chapter 8 Regression with Multiple Explanatory Variables ■ 137

Finally, asserting the presumption of normality: So, the TSS = 75.797 and the ESS = 68.598.
Y; = - 1 .2 4 2 + 1.873*X 1(- - 1.221*X2, p2 _ E S S _ 68.598

t\ = 90.5%
+ £// £/ N(0, 0.718) T SS ~ 75.797
n - 1
Note that the variance for e, is biased and the unbi- R2 = 1 - (1 - R2)
n - k - 1
10
ased figure is this figure multiplied by — ---- — = 1.429 10 - 1
1I (1 - 0.905) = 0.878
or 1.206. 10 - 2 -
b. (.R SS r - RSSu)/q (7.58 - 0.72)/2
F = 33.47
RSSu/(N - ku - 1) 0.72/(10 - 2 - 1 )
Trial
Number
A
y y (y - y )2 (y - y )2
1 -5 .7 6 0 -6 .0 8 9 8.626 10.664
2 0.030 -0 .5 2 3 8.140 5.290
3 -0 .2 5 0 -0 .8 7 3 6.620 3.802
4 -2 .7 2 0 -0 .3 4 7 0.011 6.131
5 -3 .0 8 0 -3 .2 5 4 0.066 0.185
6 -7 .1 0 0 -6 .8 3 4 18.293 16.085
7 -4 .1 0 0 -4 .1 4 2 1.631 1.741
8 0.140 -0 .4 3 6 8.779 5.698
9 -6 .1 3 0 -5 .9 4 9 10.936 9.774
10 0.740 0.215 12.695 9.227
A verage -2 .8 2 3
SUM 75.797 68.598

QUESTIONS

9.1. Define hom oscedasticity and heteroskedasticity. When irrelevant regressors are included when the regressors are
might you exp ect data to be hom oscedastic? uncorrelated?
9.2. W hat conditions are required for a regression to suffer 9.5. W hy does a variable with a large variance inflation factor
from om itted variable bias? indicate that a model may be im proved?
9.3. W hat are the costs and benefits of dropping highly collin- 9.6. The sam ple mean is an O LS estim ator of the model
ear variables? Yj = a + 6,. W hat does the BLU E property imply about
the mean estim ator?
9.4. If you use a general-to-specific model selection with a
test size of a, what is the probability that one or more 9.7. W hat is the strongest justification for using O LS to esti
mate model param eters?
Practice Questions
9.8. In a model with a single explanatory variable, what value there are nonlinear effects of some of these variables,
of the R2 in the second step in a test for heteroskedasticity and so use a R E S E T test including both the squared and
indicates that the null would be rejected for sam ple sizes cubic term . The R2 of the original model is 68.6% , and the
of 100, 500, or 2500? (Hint: Look up the critical values of a R2 from the model that includes both additional term s is
Xq, where q is the number of restrictions in the test.) 68.9% . You have 456 observations. W hat do you conclude
about the specification of the model?
9.9. Suppose the true relationship between Y and two explan
atory variables is Yj = 2 + 1.2X-I,- — 2.1X2/ + e/- 9.12. You are investigating the determ inants of book leverage
of firms. Your model includes an intercept and four addi
a. W hat is the population value of (3-\ in the regression
tional variables: the natural logarithm of sales revenue, the
Yj = a + (3-jX-1 + €j if pxiX2 = 0-6 and ox = 1 and
book-to-market ratio, the ratio of EBlTD A-to-book assets,
<>i = 1 / 2 ?
and Net property-plant-and-equipm ent (Net PPE). Your
b. W hat value of pXlx2 would make f3-\ = 0?
initial model groups firm s in all industries. Suppose you
9.10. In a model with two explanatory variables,
want to test w hether the four slope coefficients are con
Y-, = a + jS-|X1( + /32X 2; + €„ what does the correlation
stant across energy sector firms.
need to be between the two regressors, X | and X 2, for the
variance inflation factor to be above 10? a. W hat additional model would you estim ate to im ple
ment the Chow test?
9.11. You are interested in understanding the determ inants of
b. If your sam ple has 433 observations, what is the distri
the yield spread of corporate bonds above a maturity
bution of the Chow test?
matched sovereign bond. You include three explanatory
c. If the R2 in the original specification is 16.6% , how
variables: the leverage defined as the ratio of long-term
large would the R2 have to be in the regression with
debt to the book value of assets, a dummy variable for
the dummy interactions for the null to be rejected
high yield, and a measure of the volatility of the profit
using a 5% test size?
ability of the issuer. You are interested in testing whether

9.13. G iven the data set: where the data was used for fitting the model versus the
blue indicating the data that was held back):
Y X 1 X2
M1 m2 m3
1 -2 .3 5 3 -0 .4 0 9 -0 .0 0 8
2 -0 .1 1 4 0.397 -1 .2 1 6 1 2.1 0.6 1.1
3 -1 .6 6 5 -0 .8 5 6 -0.911 2 -0.1 - 0 .4 2.5
4 -0 .3 6 4 1.682 0.366 3 - 0 .8 -0.1 1
5 -0.081 0.455 -0 .6 3 9 4 - 2 .6 - 0 .5 -0.1
6 -0 .7 3 5 - 1 .3 9 -1 .0 8 6 5 -1 .1 - 0 .6 1.3
7 -2 .5 0 7 0.954 0.67 6 - 0 .4 0.9 0.5
8 -1 .1 4 4 1.021 0.238 7 0.6 0 0
9 -2 .4 1 9 -0 .1 5 6 -0 .0 5 5 8 0.9 -1 - 0 .4
10 -3.151 1.382 1.148 9 1.2 0.8 -0.1
11 -2 .0 8 5 -0 .5 6 2 -0 .1 3 5 Using the principles of m-fold cross validation, which
12 -2 .9 7 2 -1 .5 5 4 -0 .2 9 9 model should be selected?
13 -0 .6 3 3 -1 .1 2 3 -1 .0 2 7 9.15. Consider the following data provided in the table and

plotted below:
14 -2 .6 7 8 -0 .1 2 4 0.331
15 -7 .0 9 5 0.284 2.622 Y X
a. Find the param eters for the model: 1 -1 0 .4 2 -5

Y = a + /3-|X-|. 2 -1 4 .1 2 -4
b. Find the param eters for the model: 3 -1 4 .7 2 -3
Y = a + ^2X 2 -
4 -8 .2 5 -2
c. Using the results from part a and part b, and the corre
5 2.3 -1
lation between the two explanatory variables, estim ate
the param eters for the full model: 6 1.59 0
Y = a + f a X y + ^2X 2 . 7 5.87 2
d. C heck the answer to part c by using the Excel function 8 14.13 3
LIN EST or the Excel regression macro.
9 6.78 4
9.14. A given set of data was divided into three equal parts.
10 32 1
Three separate models were developed— each model
using two of the three parts for fitting. Errors were cal
culated for each model. The diagram below shows the
standard errors for each model run (the orange highlights
Chapter 9 Regression Diagnostics 151

-10
i----------------------1---------------------- 1---------------2 0 — a—
M anagem ent is concerned that the last data point in the

table is an outlier.
a. Calculate Cook's Distance for the last data point.

Use built-in regression functions in Excel to assist as
needed, such as LINEST.
b. W hat is the interpretation of this result? W hat does it
say about the status of the last data point as an outlier

ANSW ERS

9.1. Hom oscedasticity is a property of the model errors able to clearly attribute m ovem ent in the left-hand-side
where they have the same variance. Heteroskedasticity is variable to the remaining variables.
a property of the errors where their variance changes sys 9.4. If the regressors are uncorrelated, then their joint dis
tem atically with the explanatory variables in the model. tribution has no correlation, and so the probability that
Experim ental data are highly likely to be hom oscedastic. one-or-more irrelevant regressors is included is 1 minus
In general, the more hom ogeneous the data, the more the probability that none are. The probability of includ
likely the errors will be hom oscedastic. In finance, we
ing a single irrelevant regressor is 1 — a, where a is the
often use data with substantially different scales, for test size. This is the definition of the size of a test. The
exam ple, corporate earnings or leverage ratios. This het probability that no irrelevant regressors are included is
erogeneity is frequently accom panied by heteroskedas
(1 — a)p, where p is the number of variables considered.
ticity in model errors.
9.5. W hen a variable has a large VIF, it is highly correlated
9.2. There are two conditions. First, the variable must be with the other variables in the model. The param eter on
om itted in the sense that it has a non-zero coefficient in
this variable is usually poorly estim ated, and so dropping
the correct model. Second, the om itted variable must be it tends to have little im pact on the fit of the model.
correlated with an included variable. If the X variables are
9.6. The sam ple mean estim ator is the BLU E because it is a
uncorrelated, then the coefficients can be estim ated con
special case of O LS if the data are hom oscedastic.
sistently even if there is an om itted variable.
9.7. W hen the errors are iid normally distributed, then the
9.3. The cost is that the model may be m isspecified if the
O LS estim ator of the regression param eters {a and (3, but
dropped variable should be in the true model. The ben
not s2) is an M VUE (Minimum Variance Unbiased Estim a
efit is that the param eter estim ates of the remaining
tor), and so there is not a better estim ator available.
variables will improve because the O LS estim ator will be
Solved Problems
9.8. W hen there is a single explanatory variable in the original b. The om itted variable formula shows that when a vari
m odel, the auxiliary regression used for W hite's test will able is excluded, the coefficient on the included vari
have an intercept and two explanatory variables— the ables is /3-i + y/32, where y is the population regression
original variable and the variable squared. W hite's test coefficient from the regression X 2( = 8 + y X i; + I?;.
statistic is nR2, and the null has a X 2 distribution. When This coefficient depends on the correlation between
using a test with a 5% size, the critical value is 5.99. Solv the two and the standard deviations of the two vari-
5.99 ox, VT72
ing for the R2, R2 < . The maximum value of R2 that ables, so that y = pXlx2— = 0-6 = 0.424. The
n
would not reject is 0.0599, 0.011, and 0.0023 when n is coefficient is then 1.2 — 2.1 X 0.424 = 0.309.
100, 500, and 2500, respectively.
p X V T X VTT2
Solving 1.2 — 2.1 X 0 so that
9.9. a. Using the om itted variable form ula, when 1
X 2 is om itted from the regression,
if p = ---------- — = 0.522 then 13i would be 0 in the
<0 /0.6 x VT x V T 2 \ 2.1 X V T 2
f t = 1.2 - 2.1 x ( ----------- ------------ j = - 0 .1 8 ,
om itted variable regression.
, 0.6 x VT x VT2 . .
w h e re ----------- ------------ is the regression slope from
a regression of X 2 on X-|.
Chapter 9 Regression Diagnostics ■ 153

1 the slope coefficient to vary in the energy industry, so

9.10. The variance inflation factor i s ------- t t , where R? mea-
1 - R? 1 that the extended model is
sures how well variable j is explained by the other Yj = a + (3s \n SALES, + [3bB T M ) + (3EEBITDA , + foN etPPE }
variables in the model. Here there are two variables, y si E n e r g y ^ l^ S A L E S j + y BlE n e r g y ^ B TM j + y ElE n e r g y
and so Rf = Px,x2- The variance inflation factor of 10 X EBITD A i + yplEnergy X N etPPE, + e„
solves 10 = 1/(1 - p^x2), so that (1 - px,x2) = 1/10 or
where l Energy is a dummy that takes the value 1 if a firm
Px,x2 = 1 — 1/10 = 0.9. Correlations greater than (in
is in the energy industry.
absolute value) V (T 9 ~ 0.948 would produce values
above 10. b. The Chow test is an F-test with degrees of freedom
equal to the number of coefficients being tested (four
9.11. The R ES ET test exam ines whether the two additional
here, because there are four interaction terms) and
explanatory variables that squared and cubed fitted
the number of observations minus the number of
values have zero coefficients. It is im plem ented using an
coefficients in the extended model (433 - 9 here), so
F-test:
that the test statistic has a F4 424 distribution.
0.689 - 0.686 /1 - 0.689 \
F 2,450 c. The F-test statistic is
V 456 - 6 )
The F-test exam ines the difference between the R2 in the

two m odels. The critical value for an F2450 is 3.01 (F.INV.
The critical value is 2.392 (F.IN V.RT(0.05,4,424)). The
RT(0.05,2,450) in Excel). The value of the test statistic is
value of Rfj can be found by trial-and-error, and values
2.17, which is less than the critical value, and so the null
equal to 18.5% or larger lead to rejection.
that the coefficient on the squared and cubic term s is 0 is
not rejected. 9.13. a. Using the standard form ula:
C o v iY .X ,)
9.12. a. The original model is Y, = a + /%In S A L E S , + (3b BTM ,
0i = v w
+ (3e EB\TD A, + PpNetPPEj + er The Chow test allows
Y X1 (Y -Y ) ( X ,—X . H Y - Y ) ( X ,- X ,) 2
1 -2 .3 5 3 -0 .4 0 9 -0 .4 0 9 -0 .3 5 3 0.145 0.167
2 -0 .1 1 4 0.397 0.397 1.886 0.749 0.158
3 -1 .6 6 5 -0 .8 5 6 -0 .8 5 6 0.335 -0 .2 8 7 0.733
4 -0 .3 6 4 1.682 1.682 1.636 2.751 2.829
5 -0.081 0.455 0.455 1.919 0.873 0.207
6 -0 .7 3 5 - 1 .3 9 -1 .3 9 0 1.265 -1 .7 5 8 1.932
7 -2 .5 0 7 0.954 0.954 -0 .5 0 7 -0 .4 8 4 0.910
8 -1 .1 4 4 1.021 1.021 0.856 0.874 1.042
9 -2 .4 1 9 -0 .1 5 6 -0 .1 5 6 -0 .4 1 9 0.065 0.024
10 -3.151 1.382 1.382 -1.151 -1.591 1.910
11 -2 .0 8 5 -0 .5 6 2 -0 .5 6 2 -0 .0 8 5 0.048 0.316
12 -2 .9 7 2 -1 .5 5 4 -1 .5 5 4 -0 .9 7 2 1.511 2.415
13 -0 .6 3 3 -1 .1 2 3 -1 .1 2 3 1.367 -1 .5 3 5 1.261
14 -2 .6 7 8 -0 .1 2 4 -0 .1 2 4 -0 .6 7 8 0.084 0.015
15 -7 .0 9 5 0.284 0.284 -5 .0 9 5 -1 .4 4 7 0.081
Average -2 .0 0 0 0 .0 0 0 0 .0 0 0 0.933

Therefore,
0.000
0.000
0.933
And, of course,
a = Y - f r X , = - 2 .0 0 0
Y = -2.000 最新cfafrm加微信286982279
b. Proceeding as in part a gives:
Y X2 (X 2 - X 2) (Y -Y ) (X 2 - X 2) CY - Y ) (X 2 - x 2f
1 -2 .3 5 3 -0 .0 0 8 -0 .0 0 8 -0 .3 5 3 0.003 0.000
2 -0 .1 1 4 -1 .2 1 6 -1 .2 1 6 1.886 -2 .2 9 3 1.478
3 -1 .6 6 5 -0.911 -0.911 0.335 -0 .3 0 5 0.830
4 -0 .3 6 4 0.366 0.366 1.636 0.599 0.134
5 -0.081 -0 .6 3 9 -0 .6 3 9 1.919 -1 .2 2 6 0.408
6 -0 .7 3 5 -1 .0 8 6 -1 .0 8 6 1.265 -1 .3 7 3 1.179
7 -2 .5 0 7 0.67 0.670 -0 .5 0 7 -0 .3 4 0 0.449
8 -1 .1 4 4 0.238 0.238 0.856 0.204 0.057
9 -2 .4 1 9 -0 .0 5 5 -0 .0 5 5 -0 .4 1 9 0.023 0.003
10 -3.151 1.148 1.148 -1.151 -1 .3 2 2 1.318
11 -2 .0 8 5 -0 .1 3 5 -0 .1 3 5 -0 .0 8 5 0.012 0.018
12 -2 .9 7 2 -0 .2 9 9 -0 .2 9 9 -0 .9 7 2 0.291 0.089
13 -0 .6 3 3 -1 .0 2 7 -1 .0 2 7 1.367 -1 .4 0 4 1.055
14 -2 .6 7 8 0.331 0.331 -0 .6 7 8 -0 .2 2 5 0.110
15 -7 .0 9 5 2.622 2.622 -5 .0 9 5 -1 3 .3 6 0 6.875
A ve ra g e -2 .0 0 0 0 .0 0 0 -1 .3 8 1 0.934
- 1 .3 8 1
- 1 .4 7 9
0.934
a = Y - (32X 2 = - 2 .0 0 0
y = -2 .0 0 0 - 1 . 4 7 9 X 2
c. The key statem ent from the chapter is "the estim ator
/3-| converges to (3-\ + (328 "
So, part a gives that:
/3i + (328^ = 0
And part b gives that:
(32 + (3-\82 —1.479

X1 X2 (X2- X 2) (X1- X 1)(X2- X 2) (X, - X, )2 (X2 - X2)2
1 -0 .4 0 9 -0 .0 0 8 -0 .4 0 9 -0 .0 0 8 0.003 0.167 0.000
2 0.397 -1 .2 1 6 0.397 -1 .2 1 6 -0 .4 8 3 0.158 1.478
3 -0 .8 5 6 -0.911 -0 .8 5 6 -0.911 0.780 0.733 0.830
4 1.682 0.366 1.682 0.366 0.616 2.829 0.134
5 0.455 -0 .6 3 9 0.455 -0 .6 3 9 -0.291 0.207 0.408
6 - 1 .3 9 -1 .0 8 6 -1 .3 9 0 -1 .0 8 6 1.510 1.932 1.179
7 0.954 0.67 0.954 0.670 0.639 0.910 0.449
8 1.021 0.238 1.021 0.238 0.243 1.042 0.057
9 -0 .1 5 6 -0 .0 5 5 -0 .1 5 6 -0 .0 5 5 0.009 0.024 0.003
10 1.382 1.148 1.382 1.148 1.587 1.910 1.318
11 -0 .5 6 2 -0 .1 3 5 -0 .5 6 2 -0 .1 3 5 0.076 0.316 0.018
12 -1 .5 5 4 -0 .2 9 9 -1 .5 5 4 -0 .2 9 9 0.465 2.415 0.089
13 -1 .1 2 3 -1 .0 2 7 -1 .1 2 3 -1 .0 2 7 1.153 1.261 1.055
14 -0 .1 2 4 0.331 -0 .1 2 4 0.331 -0.041 0.015 0.110
15 0.284 2.622 0.284 2.622 0.744 0.081 6.875
Average 0 .0 0 0 0 .0 0 0 0.467 0.933 0.934
C o v [X i,X 2] 0.467 Solving yields that

0.501
V [X ,] 0.933 f t = 0.989 f t = - 1 .9 7 3 3
and
And plugging these back into the equation to find
C o v[X lfX 2] 0.467 alpha:
So = = 0.500
V [X 2] 0.934
Y = a + ftX-| + f t X 2 —2.000
Note: In this case these quantities are essentially
equal, but that will not usually be the case. Y = - 2 .0 0 0 + 0.989X-, - 1.9733X2
Plugging back in the 2 X 2 system of equations: d. All approaches should provide the same answer.
f t + 0.501 f t = 0
0 .5 0 0 ft + f t = - 1 .4 7 9

9.14. T h e fi rst task is to calculate the squared residuals:
M1 M2 M3 M1 M2 M3
1 2.1 0.6 1.1 4.41 0.36 1.21
2 -0.1 0.4 2.5 0.01 0.16 6.25
3 - 0 .8 0.1 1 0.64 0.01 1
4 - 2 .6 0.5 -0.1 6.76 0.25 0.01
5 -1 .1 - 0 .6 1.3 1.21 0.36 1.69
6 -0.4 - 0 .9 0.5 0.16 0.81 0.25
7 0.6 0 0 0.36 0 0
8 0.9 -1 - 0 .4 0.81 1 0.16
9 1.2 0.8 -0.1 1.44 0.64 0.01
T O T A L RSS 15.8 3.59 10.58
C V RSS 5.06 1.42 0.17
The model selected is the one that has the sm allest RSS Y1 is the estim ate using the entire data set
within the blue out-of-sample boxes— this is M3.
Y2 is the estim ate using the first nine observations
9.15. a. Following the m ethodology of the first exam ple:
So the Cook's distance for the last data point is
106.552
Alpha Beta 0.61
2 *8 7 .7 8 3
Entire 3.26 3.49 b. The last data point does not have a major influence on
First nine 0.10 2.96 the O LS param eters. However, it may still be an out
lier, and other tests would need to be conducted to
Com pleting the table: draw such a conclusion.
Y X Y1 Y2 (Y1 —Y2)A2 Y 1 -Y
1 -1 0 .4 2 -5 -14.181 -1 4 .7 0 7 0.277 -3.761
2 -1 4 .1 2 -4 -1 0 .6 9 2 -1 1 .7 4 5 1.107 3.428
3 -1 4 .7 2 -3 -7 .2 0 4 -8 .7 8 3 2.491 7.516
4 -8 .2 5 -2 -3 .7 1 6 -5.821 4.428 4.534
5 2.3 -1 -0 .2 2 8 -2 .8 5 8 6.919 -2 .5 2 8
6 1.59 0 3.260 0.104 9.963 1.670
7 5.87 2 10.236 6.028 17.713 4.366
8 14.13 3 13.724 8.990 22.418 -0 .4 0 6
9 6.78 4 17.213 11.952 27.676 10.433
10 32 1 6.748 3.066 13.561 -2 5 .2 5 2
Sum 106.552
SumSq/10 87.783

QUESTIONS

10.1. W hat are the three com ponents of a tim e series? 10.6. W hat is the maximum number of non-zero autocorrela
tions in an MA(q)?
10.2. D escribe the four features of a tim e series that violate the
assum ptions of covariance stationarity. 10.7. Rewrite the MA(2), Yt = 0.3 + 0.8et_-| + 0.16et_2 + et,
using a lag polynomial.
10.3. How does Gaussian white noise differ from general white
noise? 10.8. W hy does the A IC select a model that is either the same
10.4. W hat are the properties of the lag operator? size or larger than the model selected using the BIC ?
10.5. W hat are the key features of the A C F and PA C F of an 10.9. W hat steps should be taken if the A C F and/or PA C F of
model residuals do not appear to be white noise?
MA(q) process?
Practice Questions
10.10. In the covariance-stationary AR(2), Yt = 0.3 + c. Calculate the tim e series for the following ARM A(1,1)
1.4Yt_-| — 0 .6 Y t_ 2 + et, where et ~ WN(0, a 2), what is m odel, taking e0 = 0, q0 = 0, rj = 0.5, and
the long-run mean E[Yt] and variance V[Yt]? (f> = 0 = 0.375:
10.11. If (1 — 0.4L)(1 — 0 .9 L4)Y t = et, what is the form of this 10.14. In the MA(2), Yt = 4.1 + 5et_-| + 6.25et_2 + et, where
model if written as a standard AR(5) with Yt_ 1f. . . , Yt_ 5 et ~ WN[0, a 2), what is the A C F ?
on the right-hand side of the equation?
10.15. Suppose all residual autocorrelations are 1 .5 / V ^ where
10.12. For the equation: T is the sam ple size. Would these violate the confidence
15 5 3 bands in an A C F plot?
Y‘ = 3 i + ? y * - ' - 8 Y* - 2 + e‘ 10.16. Suppose you observed sam ple autocorrelations

W hat is the lag polynomial? Pi = 0.24, p2 = —0.04, and p3 = 0.08 in a tim e series
10.13. G iven the following data for a set of innovations: with 100 observations. W ould a Ljung-Box Q statistic
reject its null hypothesis using a test with a size of
et 5%? W ould the test reject the null using a size of 10%
1 -0.58 or 1%?
2 1.62 10.17. Suppose you fit a range of A R M A models to a data

set. The A R M A orders and estim ated residual variance
3 -1.34 -2
a are
4 0.88
5 -1.11 Order c r 2 Order c r 2 Order c r 2
6 0.52 (1 ,0 ) 1.212 (1, D 1.147 (1 ,2 ) 1.112
7 0.81 (2, 0) 1.175 (2, 1) 1.097 (2, 2) 1.097
8 -0.65 (3, 0) 1.150 (3, 1) 1.097 (3, 2) 1.068
9 0.85 (4, 0) 1.133 (4 ,1 ) 1.096 (4, 2) 1.065
10 -0.88 Which model is selected by the A IC and BIC if the sample size
T = 250?
a. Calculate the tim e series for the following AR(1)
model, taking y0 = 0, S = 0.5, and </> = 0.75: 10.18. The 2018Q 4 value of real G D P growth is
b. Calculate the tim e series for the following MA(1) RGDPGj - t = 2.793. W hat are the forecasts for 2019Q1
model, taking e0 = 0, ju = 0.5, and 0 = 0.75: - 2019Q 4 using the AR(2) model in Table 10.1?
Chapter 10 Stationary Time Series ■ 183

ANSWERS
Short Concept Questions b. Lag operators multiply so that LL = L2 and

L2Y, = L ( L Y , ) = L Y ,_ , .
10.1. The three com ponents are trend, seasonal, and cyclical.
c. The lag operator applied to a constant is just the con
10.2. a. Random walks— The variance of the process changes stant, L8 = 8.
over tim e because a shock is never "forgotten" by the
process.
d. The pth order lag polynomial is written 4>[L) =
1+<
/>
-| L + 4>2L2 + ••• + (f>pLp.
b. Determ inistic trends— The mean of the process is
e. Lag polynomials can be multiplied to produce another
changing over tim e due to a tim e trend; for exam ple,
lag polynomial.
the series has a positive growth rate.
f. Under some assum ptions, lag polynomials can be
c. Determ inistic seasonality— The mean of the process
inverted.
depends on the time observation due to a seasonal,
repeating pattern. 10.5. The A C F of an MA(q) is non-zero until q lags, and then is
zero for all values above q. The PA C F declines slowly in
d. Structure breaks— The model param eters change over
the lag and may oscillate, but there is no lag above which
tim e.
the PA CF is always zero.
10.3. G eneral white noise can have any distribution and does
10.6 . q, the order of the model.
not have the be independent across tim e. Gaussian
white noise is iid, and each shock is normally distributed 10.7. Y t = et(1 + 0.8L + 0.16L2)
N(0, a 2). 10.8. The BIC has a sharper penalty for all sam ple sizes larger
10.4. a. The lag operator shifts the tim e index back one than 7, and so will always prefer few er lags.
observation. 10.9. The model should be expanded to include additional AR
or M A coefficients.
Solved Problems
10.10. Because this process is covariance-stationary 10.13. a. Yt — 8 + 4>Yt_i+et
0.3
E [Y t] = A = = 1.5 yt
1 - 1.4 - (0.6)
0 .3 : 1 -0 .0 8
V [Y t] = y 0 = = 0.45
1 - 1.4 - (0.6) 2 2.06
10.11. (1 - 0.4L)(1 - 0 .9 L4)Y t = (1 - 0 .4 L - 0 .9 L4 + 0 .3 6 L3)Yt 3 0.71
=> Yt - 0 .4 Yt_ ! - 0 .9 Y t_ 4 + 0 .3 6 Yt_ 5 = et 4 1.91
Yt = 0 .4 Y t_-| + 0 .9 Y t_ 4 - 0 .3 6 Yt_ 5 + et 5 0.82
6 1.64
10.12.1 — — f c L 2 = 1 - f l + fL*
7 2.54
8 1.76
9 2.67
10 1.62

Each value is generated via the form ula: Each value is generated via the form ula:
yy = 0.5 + 0.75 * 0 - 0.58 = - 0 .0 8 qy = 0.5 + 0.375 * 0 + 0.375 * 0 - 0.58 = - 0 .0 8
y2 = 0.5 + 0.75 * ( - 0 .0 8 ) + 1.62 = 2.12 z2 = 0.5 + 0 .3 7 5 (-0 .0 8 ) + 0 .3 7 5 (-0 .5 8 ) + 1.62 = 1.87

Z2
and so forth. and so forth.
2\ 2
Z{ = fJL + 0 € {- y + 10.14. = (1 + 52 + _
.2
and y 2 = 6cr . The autocorrelations are then
Zt Po = 1, py = 0.557 and p2 = 0.096.
1 -0 .0 8 10.15. No, the confidence bands are ± 1 .96/^ / t and so these

would not. If many autocorrelations are jointly relatively
2 1.69
large, then this series is likely autocorrelated. A Ljung-
3 0.38
Box test would likely detect the joint autocorrelation.
4 0.38
1 0 .1 6 . The test statistic is
5 0.05
6 0.19 Qlb = T 2 ( ) P?
/= 1 T — i
7 1.70
= 1001 ^ ) (0.24)2 + 100 [ ^ j (- 0 .0 4 )2 + 100 ( ^ ) (0.08)2
8 0.46 99 98 97
9 0.86 = 5.93 + 0.16 + 0.67 = 6.77
10 0.26 The critical values for tests sizes of 10%, 5%, and 1% are
6.25, 7.81, and 11.34, respectively. (CHISQ.INV(1 -a, 5)
Each value is generated via the form ula: in Excel, where a is the test size). The null is only be
rejected using a size of 10%.
Zy = 0.5 + 0.75 * 0 - 0.58 = - 0 .0 8
10.17. The ICs and the selected model in bold:
z2 = 0.5 + 0 .7 5 (—0.58) + 1.62 = 1.69
and so forth. Order AIC Order AIC Order AIC
c. qt = n + <fiqt-y + 0et_ , + et (1 ,0 ) 50.068 (1, D 38.287 (1 ,2 ) 32.540
(2, 0) 44.317 (2, 1) 29.145 (2, 2) 31.145

qt
(3, 0) 40.940 (3 ,1 ) 31.145 (3, 2) 26.447
1 -0 .0 8
(4, 0) 39.217 (4, 1) 32.917 (4, 2) 27.744
2 1.87
3 0.47 Order BIC Order BIC Order BIC

4 1.05
(1 ,0 ) 53.589 d, D 45.330 (1 ,2 ) 43.104
5 0.12
(2, 0) 51.360 (2, 1) 39.709 (2, 2) 45.231
6 0.65
(3, 0) 51.505 (3, 1) 45.231 (3, 2) 44.054
7 1.75
(4, 0) 53.303 (4 ,1 ) 50.524 (4, 2) 48.872
8 0.81
9 0.41
10 0.47
Chapter 10 Stationary Time Series ■ 185

10.18. All forecasts are recursively com puted starting with the The three- and four-step depend entirely on other
first: forecasts:
1.765 + 0.319 X 2.564 + 0.114 X 2.793 = 2.90. 1.765 + 0.319 X 2.98 + 0.114 X 2.90 = 3.05.
The two-step forecast uses the one-step forecast: 1.765 + 0.319 X 3.05 + 0.114 X 2.98 = 3.08.
1.765 + 0.319 X 2.90 + 0.114 X 2.564 = 2.98.

QUESTIONS

11.1 W hen modeling In Yt using a tim e trend model, what is 11.3 W hy does a unit root with a tim e trend,
the relationship between exp E T[ln YT+h] and E T[Y T+h] Yt = 61 + Yt_-] + et not depend explicitly on t?
for any forecasting period h? Are these ever the same?
11.4 W hat are the consequences of excluding a determ inistic
Assum e the error term s is normally distributed around a term in an A D F test if it is needed? W hat are the conse
mean of zero.
quences of including an extra determ inistic term that is
11.2 Suppose an hourly tim e series has a calendar effect where not required to fit the data?
the hour of the day m atters. How would the dummy vari 11.5 W hy doesn't the A D F test reject if the t-statistic is large
able approach be im plem ented to capture this calendar
and positive?
effect? How could differencing be used instead to remove
the seasonality?
Practice Questions
3
11.6 A Iinear tim e trend model is estim ated on annual real 11.8 The seasonal dummy model Yt = 8 + 2 ; = 1 Yjljt + et is
euro-area GDP, measured in billions of 2010 euros, estim ated on the quarterly growth rate of housing starts,
using data from 1995 until 2018. The estim ated model is and the estim ated param eters are y i = 6.23, 72 = 56.77,
RG D P t = —234178.8 + 121.3 X t + et. The estim ate of 73 = 10.61, and 8 = —15.79 using data until the end
the residual standard deviation is a = 262.8. Construct of 2018. W hat are the forecast growth rates for the four
point forecasts and 95% confidence intervals (assuming quarters of 2019?
Gaussian white noise errors) for the next three years. Note
11.9 A D F tests are conducted on the log of the ten-year
that t is the year, so that in the first observation, t = 1995, US governm ent bond interest rate using data from
and in the last, t = 2018.
1988 until the end of 2017. Th e results of the A D F
11.7 A log-linear trend model is also estim ated on annual euro with different configurations of the determ inistic term s
area G D P for the same period. The estim ated model is are reported in the table below. The final three co l
In R G D P t = —18.15 + .0136 t + et, and the estim ated umns report the num ber of lags included in the test
standard deviation of et is 0.0322. Assum ing the shocks as selected using the A IC and the 5% and 1% critical
are normally distributed, what are the point forecasts of values that are appropriate for the sam ple size and
G D P for the next three years? How do these compare included determ inistic term s. Do interest rates contain
those from in the previous problem ? a unit root?
Deterministic r So Lags 5% CV 1% CV
None - 0 .0 0 3 7 - 1 .9 4 2 - 2 .5 7 2
( - 1 .6 6 6 )
Constant - 0 .0 0 9 0 .0 1 0 4 - 2 .8 7 0 - 3 .4 4 9
(-1 .4 2 5 ) (1.027)
Trend - 0 .0 8 5 0.188 - 0 .0 0 0 3 - 3 .4 2 3 - 3 .9 8 4
(-4 .3 7 8 ) (4.260) (-4 .1 0 9 )

11.10 An AR(2) process is presented: The data is to be evaluated using the following regres
Yt = kYt_ i + m Yt_ 2 + et sion with a Dickey-Fuller (DF) Test:
It is determ ined that the equation has a unit root. Yt = Y t - 1 + e t.
The specific regression to be evaluated is

a. W hat conditions must k and m satisfy for the first dif
ference series to be stable? yt - yt- i = myt-'i + et
b. Express the equation as an AR(1) process. The relevant critical values for the D F test are 1%: —2.60,
11.11 The following data is to be evaluated for being a 5%: - 1 .9 5 .
random walk:
a. W hat is the appropriate null hypothesis and counter
hypothesis if a one-sided D F test is em ployed?
t Yt
b. Calculate the D F t-statistic, taking Vq = 0.
1 -1.81 Hint: For a regression with no intercept, y,- = mx,- + e,-:
2 -2.31 Ix ^
3 -1 .6 5
4 -0.81
5 -0.31
6 -0 .3 7
7 0.47
8 1.10
9 0.53
10 0.47
11 1.00
12 1.68
13 0.90
14 1.99
15 1.31
16 0.76
17 0.43
18 -1 .0 1
19 0.27
20 -2.21
21 -3.21
22 - 1 .9 0
23 -0 .5 3
24 - 0 .6 0
25 0.34
Chapter 11 Non-Stationary Time Series ■ 205

ANSWERS

11.1 A tim e trend model for In Yt can be stated as: where /Jt = 1 when t(m od 24) = j, and lJt = 0 when
t(m od 24) 7^ j
In Yt = g(t) + €tl e ~ N(0, a 2),
where g(t) is a function of t. Differencing this series can be achieved by looking at

observation 24 periods (hours) apart from each other
So,
(the following presumes that the error term s are iid and
E r [ln Vr+h] = g (T + h), normal):
which gives y t + 2 4 ~ y t = g (t + 24) - g(t) + et+24 - et
exp E r [ln Yr+h] = exp [g (T + h)] O nce the determ inistic time trend is removed the
On the other hand: remaining is a covariance-stationary MA(1) process.
EA yr+ h l = E T[e x p (g (T + h) + eT+h)] = exp (g (T + h) 11.3 The time trend becom es apparent as the series is propa
+ Eyfexp eT+h)j, gated backwards:
which equals yt — 6-| + yt_! + €t — 2§-| + yt_ 2 + et + et_i — • • •

2
t
E T[Y T+h] = exp [g (T + h)] + y
= tS, + Y0 + 2 6/
/= 1
And so:
11.4 Excluding required determ inistic term s m akes it
2
harder to reject the null, and the pow er of the test
E T [yT+h} = exp E r [ln y T+h] + y
will fall tow ards zero. On the other hand, including
These will be equal if the variance is zero (in other words, superfluous determ inistic term s skew s the distribu
if the process is com pletely determ inistic. tion to the left and also reduces the pow er of the test.
1 1 .2 Let s = 24 represent the hour of the day in military time 11.5 The A D F focuses on the left side of the distribution and
(e.g. 13 = 1 p.m .). Then rejection occurs if the test statistic is to the left of the
Yt = 9(t) + 71*11 + 72*2t + • • • + 7 2 3 /2 3 t + € t>

dem arcation point (e.g ., large and negative).
Solved Problems
11.6 Note that there is no AR or M A com ponent, so the vari And the error bounds on the In are + /— 1.96*0.0322, so
ance remains constant. Therefore, the 95% confidence the bounds are given in proportional term s rather than
interval is + / — 1.96*262.8 = + / — 515.1 about the fixed values.
expected value. Bounds_M ultiplier = exp( ± 1.96 * 0.0322)
A s for the expected means: = e xp (± 0.0631) = 0.939,1.065
E [R G D P 2o i 9] = -2 3 4 1 7 8 .8 + 121.3 X 2019 = 10,725.9 Calculating E[ln R G D P J:
E [R G D P 202o ] = -2 3 4 1 7 8 .8 + 121.3 X 2020 = 10,847.2 E[ln R G D P2019] = - 1 8 .1 5 + 0.0136 * 2019 = 9.308
E [R G D P 202-\] = -2 3 4 1 7 8 .8 + 121.3 X 2021 = 10,968.5
E[ln R G D P2020] = - 1 8 .1 5 + 0.0136 * 2020 = 9.322
11.7 In this case: E[ln R G D P202i ] = - 1 8 .1 5 + 0.0136 * 2020 = 9.336
a
ET[YT] = exp ET[ln Y t ] + Furtherm ore,
0.03222
0.0005
2

(which will only make a small im pact in this exam ple) For the trend m odel, both of these have t-stats with
absolute values > 4, well within the bounds of statisti
So:
cal significance at even the 99% level. Accordingly, the
E [R G D P 2o i 9] = exp(9.308 + 0.0005) = 11,031.4
proper model to study is the last one.
E [R G D P 2 0 2 0 ] = exp(9.322 + 0.0005) = 11,186.9
For this m odel, the t-stat on -y is to the right of the 1%
E[R G D P 202i ] = exp(9.336 + 0.0005) = 11,344.6 CV — therefore, the null hypothesis of having a unit root
And the 95% confidence bands are given as: is rejected at the 99% confidence level. Note that if the
proper model were either the constant or no-trend then
95% ^CBo_ = [0 .9 3 9 * 11031.4,1.065 * 11031.4]J
d RGDP 2019 1 '
the null hypothesis would not be rejected.
= [10,358.5,11,748.4]
11.10 a. G iven that the equation has a unit root means that the
95% c b RGDP2020 = [10,504.5,11,914.0] factorization of the characteristic equation must be of
95% c b RGDp2019 = [10,652.6,12,082.0] the form:
In comparison with #1, the bands are growing in size and (z - 1)(z — c) = z2 - (c + 1)z + c,
overall the results are a bit bigger.
where 0 < < 1
1 1 .8 Though there is variance from quarter-to-quarter, the
The characteristic equation for the original relation
expected value of Y t is the same for any two observa
ship is
tions of the same quarter, regardless of the year.
z2 - k i
Accordingly:
r n
3 Therefore, it is required that | m < 1

E [Y Q1] = 8 + 2 Jjljt = - 1 5 .7 9 + 6.23 * 1 + 56.77 * 0
7=1 and
+ 10.61 * 0 = - 9 .5 6 k = ( - m + 1) or (1 - m)
3
E [Y Q2] = 8 + 2 7 jljt = - 1 5 .7 9 + 6.23 * 0 + 56.77 * 1
b. Express the equation as an AR(1) process:
7=1
Yt = (—m + 1)Vt_ 1 + m Yf-2 +
+ 10.61 * 0 = 40.98
Yt - Yt_-| = —m Yt_ 1 + m Yt_2 + et
3
E [Y Q3] = 8 + = - 1 5 .7 9 + 6.23 * 0 + 56.77 * 0 = —m(Yt_-| - Yt_ 2) + €t
y=i
Defining Z t = Yt — Yt_-|
+ 10.61 * 1 = - 5 .1 8
yields
3
E [Y Q4] = 8 + 2 y fjt = - 1 5 .7 9 + 6.23 * 0 + 56.77 * 0 Zt = + et
y=i
+ 10.61 * 0 = - 1 5 .7 9 11.11 a. H0: m = 0
Hfi m < 0
1 1 .9 T h e fi rst step is to select the appropriate model to use.
For these, the statistical significance of the param eters The counterhypothesis is different from usual practice
for the constant and trend must be taken into account. because only the left tail is being evaluated.

b. Com pleting the table yields
t yt yt - 1 A yt yt - 1 * A y t y?-i
1 -1 .8 1 0 -1.81 0.00 0.00
2 -2.31 -1.81 -0 .5 0 0.91 3.28
3 - 1 .6 5 -2.31 0.66 - 1 .5 2 5.34
4 -0.81 - 1 .6 5 0.84 - 1 .3 9 2.72
5 -0.31 -0.81 0.50 -0.41 0.66
6 -0 .3 7 -0.31 -0 .0 6 0.02 0.10
7 0.47 -0 .3 7 0.84 -0.31 0.14
8 1.10 0.47 0.63 0.30 0.22
9 0.53 1.10 -0 .5 7 -0 .6 3 1.21
10 0.47 0.53 -0 .0 6 -0 .0 3 0.28
11 1.00 0.47 0.53 0.25 0.22
12 1.68 1.00 0.68 0.68 1.00
13 0.90 1.68 -0 .7 8 -1.31 2.82
14 1.99 0.90 1.09 0.98 0.81
15 1.31 1.99 -0 .6 8 -1 .3 5 3.96
16 0.76 1.31 -0 .5 5 -0 .7 2 1.72
17 0.43 0.76 -0 .3 3 -0 .2 5 0.58
18 -1.01 0.43 - 1 .4 4 -0 .6 2 0.18
19 0.27 -1.01 1.28 - 1 .2 9 1.02
20 -2.21 0.27 -2 .4 8 -0 .6 7 0.07
21 -3.21 -2.21 - 1 .0 0 2.21 4.88
22 - 1 .9 0 -3.21 1.31 -4.21 10.30
23 -0 .5 3 - 1 .9 0 1.37 - 2 .6 0 3.61
24 - 0 .6 0 -0 .5 3 -0 .0 7 0.04 0.28
25 0.34 - 0 .6 0 0.94 -0 .5 6 0.36
SUM -5.81 0.34 -1 2 .5 0 45.76
So m can be calculated as:
_ 1 x,y, _ - 1 2 .5 0
- 0 .2 7 3
m “ S x ,2 “ 45.76

Returning to calculate the errors:
t y t y t - 1 A yt y t - 1 * A yt yf-i
1 -1.81 0 -1.81 0.00 0.00 -1.81
2 -2.31 -1.81 -0 .5 0 0.91 3.28 - 0 .9 9
3 -1 .6 5 -2.31 0.66 - 1 .5 2 5.34 0.03
4 -0.81 -1 .6 5 0.84 - 1 .3 9 2.72 0.39
5 -0.31 -0.81 0.50 -0.41 0.66 0.28
6 -0 .3 7 -0.31 - 0 .0 6 0.02 0.10 - 0 .1 4
7 0.47 -0 .3 7 0.84 -0.31 0.14 0.74
8 1.10 0.47 0.63 0.30 0.22 0.76
9 0.53 1.10 -0 .5 7 -0 .6 3 1.21 -0 .2 7
10 0.47 0.53 - 0 .0 6 -0 .0 3 0.28 0.08
11 1.00 0.47 0.53 0.25 0.22 0.66
12 1.68 1.00 0.68 0.68 1.00 0.95
13 0.90 1.68 -0 .7 8 -1.31 2.82 -0 .3 2
14 1.99 0.90 1.09 0.98 0.81 1.34
15 1.31 1.99 -0 .6 8 -1 .3 5 3.96 - 0 .1 4
16 0.76 1.31 -0 .5 5 -0 .7 2 1.72 -0 .1 9
17 0.43 0.76 -0 .3 3 -0 .2 5 0.58 -0 .1 2
18 -1.01 0.43 - 1 .4 4 -0 .6 2 0.18 - 1 .3 2
19 0.27 -1.01 1.28 - 1 .2 9 1.02 1.00
20 -2.21 0.27 - 2 .4 8 -0 .6 7 0.07 -2.41
21 -3.21 -2.21 - 1 .0 0 2.21 4.88 - 1 .6 0
22 - 1 .9 0 -3.21 1.31 -4.21 10.30 0.43
23 -0 .5 3 - 1 .9 0 1.37 - 2 .6 0 3.61 0.85
24 - 0 .6 0 -0 .5 3 -0 .0 7 0.04 0.28 -0.21
25 0.34 - 0 .6 0 0.94 -0 .5 6 0.36 0.78
SUM -5.81 0.34 -1 2 .5 0 45.76 -1 .2 5
St Dev 1.36 0.95
Using the STD EV function in E X C E L , As the t-stat is less than the 5% point on the D F distri
The standard error of this is bution, the null is rejected at a 95% confidence level,
but the t-stat is not sufficiently negative to enable
(T, 0.95
= 0.140 rejection at the 99% confidence level.
nay<-i 5 * 1.36
So the t-statistic is
m - 0 .2 7 3
t - stat = = - 1 .9 6
(T,m 0.140

QUESTIONS

12.1. If returns are fat-tailed, and we instead assume that they 12.3. Variances are usually transform ed into volatilities by tak
are normally distributed, is the probability of a large loss ing the square root. This changes the units from squared
over- or under-estimated? returns to returns. W hy isn't the same transform ation used
12.2. Is Iinear correlation sensitive to a large outlier pair? W hat on covariances?
about rank correlation and Kendall's r?
Practice Questions
12.4. On Black Monday in O ctober 1987, the Dow Jo n es Indus 1 2 .1 1 . The following data is collected for four distributions:
trial A verage fell from the previous close of 2,246.74 to
1,738.74. W hat was the sim ple return on this day? W hat Dataset Skew Kurtosis T
was the log return? A 0.85 3.00 50
12.5. If the annualized volatility on the S&P 500 is 20% , what is B 0.85 3.00 51
the volatility over one day, one w eek, and one month? C 0.35 3.35 125
12.6. If the mean return on the S&P 500 is 9% per year, what is D 0.35 3.35 250
the mean return over one day, one w eek, and one month?
W hich of these datasets are likely (at the 95% confidence
12.7. If an asset has zero skewness, what is the maximum kurto-
level) to not be drawn from a normal distribution?
sis it can have to not reject normality with a sam ple size of
100 using a 5% test? W hat if the sam ple size is 2,500? 12.12. Calculate the rank correlations for the following data:
12.8. If the skewness of an asset's return is —0.2 and its kurtosis 1
•
X Y
is 4, what is the value of a Jarque-Bera test statistic when
1 0.22 2.73
T = 100? W hat if T = 1,000?
2 1.41 6.63
12.9. Calculate the simple and log returns for the following data:
3 - 0 .3 0 - 2 .1 9
Time Price 4 - 0 .5 9 -6.51
0 100 5 -3 .0 8 -0 .9 9
1 98.90 6 1.08 2.63
2 98.68 7 -0 .4 5 - 3 .4 0
3 99.21 8 0.40 5.10
4 98.16 9 -0 .7 5 - 5 .1 4
5 98.07 10 0.24 1.14
6 97.14
12.13. Find Kendall's Tau for the following data:
7 95.70
X V
•
8 96.57 J
1 3.12 2.58
9 97.65
2 - 1 .2 6 -0 .0 5
10 96.77
3 2.08 -0 .7 2
12.10. The implied volatility for an ATM money option is
4 -0 .2 8 -0 .5 2
reported at 20% , annualized. Based upon this, what
would be the daily implied volatility? 5 - 1 .9 6 - 0 .4 0

ANSW ERS

12.1. We will underestim ate the probability of a large loss correlation and Kendall's t are both less sensitive to a
because the normal has thin tails. For exam ple, the prob single outlier because the effect is bounded and is aver
ability of a 4a loss from a normal is .003% (3 in 100,000). aged out by the other observations. If n is large, then
The probability of a 4rj loss from a standardized Stu a single outlier will have an immaterial im pact on the
dent's t4 is .24% or 80 tim es more likely. estim ates.
12.2. Yes. Linear correlation is sensitive in the sense that mov 12.3. Covariance may have either sign, and so the square
ing a single point in X 1f X 2 space far from the true linear root transform ation is not reliable if the sign is negative.
relationship can severely affect the correlation. Effec Transformation to (3 or correlation is preferred when
tively a single outlier has an unbounded effect on the exam ining the magnitude of a covariance.
covariance, and ultimately the linear correlation. Rank
Solved Problems
^7 2 8 7 4 _ 2246 74 12.9. S imple returns use the form ula:
12.4. The simple return i s ------- ------- = —22.6% . The
2246.74
P t ~ P t-
log return is In 1738.74 - In 2246.74 = 25.6% . The Rt =
P t-
difference between the two is increasing in the
m agnitude. Log returns use the form ula:
12.5. The scale factors for the variance are 252, 52, and 12, Pt
rt = In Pt — In Pt_ 1 = In
P t-
respectively. The volatility scales with the square root of
i i r i 20% 20%
the scale factor, and so — ----- = 1.26% , — — = 2.77% , Time Price Simple Log
V252 \/5 2
20 % 0 100
and — 7 = = 5.77% .
V12
1 98.90 -1 .1 0 % -1 .1 1 %
12.6. The mean scales linearly with the scale factor, and so
9% 9% . . 9% . 2 98.68 -0 .2 2 % -0 .2 2 %
— — = .036% , —— = 0.17% , and —— = 0.75% .
252 52 12 3 99.21 0.54% 0.54%
12.7. The Jarque-Bera has X 2 distribution, and the criti
4 98.16 -1 .0 6 % -1 .0 6 %
cal value for a test with a size of 5% is 5.99. The
f S 2 (k - 3)2 5 98.07 -0 .0 9 % -0 .0 9 %
Jarque-Bera statistic is J B = (T - 1)( — -I-----— —
6 97.14 -0 .9 5 % -0 .9 5 %
so that when the skewness S = 0, the test statistic is
7 95.70 -1 .4 8 % -1 .4 9 %
(k - 3)2
(T — 1)— —— . In order to not reject the null, we need 8 96.57 0.91% 0.90%
24
(k - 3)2 5.99 9 97.65 1 .12% 1.11%
( T - 1) < 5.99, and so (k — 3)2 < 24 X
24 T - 1
10 96.77 -0 .9 0 % -0 .9 1 %
5.99
and k —3 T .»/ 24 X . W hen T = 100, this value
T - 1 12.10. Using the equation
is 4.20. W hen T = 2,500 this value is 3.24. This shows
®annual = V 2 5 2 X o Ldaily
;
that the J B test statistic is sensitive to even mild excess
kurtosis. 0.2 = V 2 5 2 X daily
12.8. Using the formula in the previous problem , the value of 0.2
&daily = 1.26%
the J B is 4.785 when T = 100 and 48.3 when T = 1000. V 2 5 2
Chapter 12 Measuring Returns, Volatility, and Correlation ■ 221

12.11. The appropriate test statistic is the Jarque-Bera statistic. 12.13. The first step is to rank the data:
1 X Y Ry/
•
R x,
1 3.12 2.58 5 5
This will be distributed as ar|d for a 5% test size the
critical value is 5.99 2 - 1 .2 6 -0 .0 5 2 4
3 2.08 -0 .7 2 4 1
Dataset Skew Kurtosis T JB
4 -0 .2 8 -0 .5 2 3 2
A 0.85 3.00 50 5.90
5 - 1 .9 6 - 0 .4 0 1 3
B 0.85 3.00 51 6.02
The next step is to see which pairs are concordant, which
C 0.35 3.35 125 3.16
are discordant, and which are neither.
D 0.35 3.35 250 6.35
W hich leads to the pair-by-pair com parison:
So, as the J B figure is greater than 5.99 for datasets

B & D, their respective null hypotheses (normally distrib
uted) are rejected.
1 2 .1 2 . Following the procedure outlined in the text (and exam

ple), the first step is to com plete the table:
•
1 X Y R x, Ry; RX , ~ R Xi
1 0.22 2.73 6 8 - 2
2 1.41 6.63 10 10 0 For the first observation, because both the rank of X & Y
3 - 0 .3 0 - 2 .1 9 5 4 1 are the maximum, every other observation will autom ati
cally be concordant with respect to this one.
4 - 0 .5 9 -6.51 3 1 2
By contrast, looking at the second and third observa
S -3 .0 8 -0 .9 9 1 5 -4
tions, the X rank increases while the Y rank decreases.
6 1.08 2.63 9 7 2 Therefore, this pair is discordant.
7 -0 .4 5 - 3 .4 0 4 3 1 „ _ nc - nd _ 5 -5 _
8 0.40 5.10 8 9 -1 T n(n - 1)/2 10
9 -0 .7 5 - 5 .1 4 2 2 0
10 0.24 1.14 7 6 1
Sum Squares 32
6 S / = i ( Rx; ~ Ry/): 6*32

1 0.81
10(100 1)
= -
n(n 2 - 1) -

QUESTIONS

13.1 W hat is the main difference between a simulation and the 13.5 How are bootstrap sam ples generated using the iid
bootstrap? bootstrap?
13.2 How is the seed of a pseudo-random number generator 13.6 How are bootstrap sam ples generated using the circular
useful? block bootstrap (CBB)?
13.3 W hat are antithetic variables and control variates? 13.7 W hen is the C B B needed?
13.4 W hy do antithetic variables and control variates improve 13.8 W hat are the key limitations of the bootstrap?
the accuracy of a simulation?
Practice Questions
13.9 Suppose you are interested in approximating the a. Calculate the option payoffs using the equation
expected value of an option. Based on an initial sample M ax(x — 0.5,0).
of 100 replications, you estimate that the fair value of the b. W hat are the antithetic counterparts for each sce
option is USD 47 using the mean of these 100 replications. nario and their associated values?
You also note that the standard deviation of these 100 c. Assum ing that the sam ple correlation holds for the
replications is USD 12.30. How many simulations would entire population, what is the change in expected
you need to run in order to obtain a 95% confidence inter sam ple standard error attainable by using antithetic
val that is less than 1% of the fair value of the option? How sam pling?
many would you need to run to get within 0.1% ? d. How does the answer to part c change if the payoff
13.10 Plot the variance reduction when using antithetic random function is Max(| x — 0.5 |, 0)?
variables as a function of the correlation between pairs of

observations.
13.11 Plot the variance reduction when using control variates as

a function of the correlation between the control variate
and the random variable used in the simulation.
13.12 A random draw from an N(1,5) distribution is 1.2. If this

was generated using the inverse of the cum ulative dis
tribution approach, what is the original draw from the
U(0,1) distribution?
13.13 G iven the data below:
N(0,1)
1 - 0 .6 5
2 - 0 .1 2
3 0.92
4 1.28
5 0.63
6 - 1 .9 8
7 0.22
8 0.4
9 0.86
10 1.74
Chapter 13 Simulation and Bootstrapping

ANSW ERS

13.1 A simulation uses draws from a specific distribution to correlation with the simulated values. The variance reduc
simulate the shocks and other quantities in a model error. tion works in the same manner because the variance of
A simulation requires a fully specified model that provides the sum depends on the sum of the variances.
the distribution of all random quantities. A bootstrap uses 1 3 .5 A set of n integers are randomly generated with replace
random sampling from the observed data to produce a
ment from 1 , 2 , . . . , n. These are used to select the indi
new dataset that has sim ilar characteristics. In a bootstrap, ces of original data to produce the bootstrapped sam ple.
the random sam ples are used to generate indices when
1 3 .6 The C B B uses sam ples from the original data in blocks
selecting the data to include in the bootstrap sam ple.
of size m. Each block is chosen by selecting the index for
1 3 .2 The seed is a value that sets the internal state of a the first observation in the block. A total of \m] blocks
pseudo-random number generator. Setting the seed are selected, where n is the sam ple size where [ • ] selects
allows the same sequence of pseudo-random numbers to the sm allest integer larger than its input. If the number of
be generated, which allow simulations to be reproduced. observations selected is larger than m due to rounding,
1 3 .3 A ntithetic variables are sim ulated variables that have only the first n are kept. If a block starts in the last m — 1
been constructed to have a negative correlation with a observations, the block wraps around to the start of
previously sam pled set of sim ulated values. Control vari the sam ple.
ates are mean zero values that are correlated with the 1 3 .7 The C B B is generally needed when the data being boot
quantity whose expectation is being com puted via sim u strapped have tim e series dependence. A block bootstrap
lation. The control variates are then com bined with the is required when the statistic being bootstrapped is sen
sim ulated value using an optimal w eight to reduce the sitive to the serial correlation in the data. For exam ple,
simulation variance. Control variates should be cheap to when bootstrapping the mean, the C B B is required if the
produce. data are serially correlated. If bootstrapping the variance,
the C B B is required if the squares are serially correlated—
1 3 .4 Both introduce some negative correlation in the sim ula
that is, if the data has conditional heteroskedasticity.
tion. Antithetic variables are selected to have negative
correlation across sim ulations. Because the variance of 1 3 .8 The most important limitation is that the bootstrap cannot
the sum depends on the covariance, and the covariance generate data that has not occurred in the sample. This
is negative, the variance of the sum is less than the case is sometimes called the "Black Swan" problem. A simula
if the variables were independent. Control variates add tion can potentially generate shocks larger than historically
a mean-zero term to a simulation that has a negative observed values if the assumed distribution has this feature.
Solved Problems
1 3 .9 The standard deviation is USD 12.30, and a 95% confi = 0.47 => V n = 2 x 1•0 4 7x 1230 = 102.5 (so 103). Using
dence interval is [/x — 1.96 X A + 1.96 X and .1% , we would need 1,025.8 (replace 0.47 with 0.047)
so the width is 2 X 1.96 X If we want this and so 1,026 replications.

vn
12.30
value to be 1% of USD 47.00, then 2 x 1 . 9 6 X
Vn

13.10
13.
13.12 Essentially, the question is to determ ine the cumulative b. Recall that for a N(0,1) distribution:
distribution up to 1.2 on an N(1,5) distribution.
Fxiu rf = —Fx( 1 - Ut )
This can be done with the Excel function
Accordingly, the table is populated as follows:
N O RM D IST(1.2,1 ,sqrt(5),True) and returns the value 0.535.
13.13 a. The formula is simply im plem ented in Excel. For the Option Antithetic Option
first entry: N(0,1) Payoff Sample Payoff
M a x (-0 .6 5 - 1,0) = 0
1 - 0 .6 5 0 0.65 0.15
N(0,1) Option Payoff 2 - 0 .1 2 0 0.12 0
1 - 0 .6 5 0 3 0.92 0.42 - 0 .9 2 0
2 - 0 .1 2 0 4 1.28 0.78 - 1 .2 8 0
3 0.92 0.42 5 0.63 0.13 - 0 .6 3 0
4 1.28 0.78 6 - 1 .9 8 0 1.98 1.48
5 0.63 0.13 7 0.22 0 - 0 .2 2 0
6 - 1 .9 8 0 8 0.4 0 - 0 .4 0
7 0.22 0 9 0.86 0.36 - 0 .8 6 0
8 0.4 0 10 1.74 1.24 - 1 .7 4 0
9 0.86 0.36
10 1.74 1.24
Chapter 13 Simulation and Bootstrapping ■ 237

c. Using Excel, the correlation is —27% . Therefore, the The correlation between the two option value vec
reduction in error is tors is + 25% , so the IN C R EA S E in standard error is
1 - V i + p = 15% V i T p - 1 = 12%.
For the same number of simulations.
d. The M ax in the payoff function is
M a x(|x — 0.5 |,0) = |x — 0.5
The table is now as follows:
Option Antithetic Option

N(0,1) Payoff Sample Payoff
1 - 0 .6 5 1.15 0.65 0.15
2 - 0 .1 2 0.62 0.12 0.38
3 0.92 0.42 - 0 .9 2 1.42
4 1.28 0.78 - 1 .2 8 1.78
5 0.63 0.13 - 0 .6 3 1.13
6 - 1 .9 8 2.48 1.98 1.48
7 0.22 0.28 - 0 .2 2 0.72
8 0.4 0.1 - 0 .4 0.9
9 0.86 0.36 - 0 .8 6 1.36
10 1.74 1.24 - 1 .7 4 2.24

Questions Short Concept Questions

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

Questions Short Concept Questions

Uploaded by

Copyright:

Available Formats

The following questions are intended to help candidates understand the material.

They are not actual FRM exam questions.

Short Concept Questions

1.6 Continue the application of Bayes' rule to com pute the

1.9 If fund managers beat their benchmark at random (meaning

8 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

Chapter 1 Fundamentals of Probability ■ 9

20% X 20% X 10% action is flagged is .001% X 90% + 99.999% X 0.0001%

.625% = .000999% . The P r(F ra u d n F la g ) = .001% X 90%

10 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

related to these values of X when Y = a + bX ?

2 .9 Consider the following data on a discrete random distribution:

10 0.296 a. Com pute the mean, variance, and standard deviation.

11 - 0 .9 7 9 b. Verify your result in (a) by computing E[X2] directly and

a. Calculate the mean and variance of X.

Chapter 2 Random Variables ■ 23

Short Concept Questions

2.2 The PM F m easures the probability of a single value.

2.3 A PM F must be non-negative and sum across the support

2.4 E[Y] = E[a + bX]

The variance of V is V[a + bX] = V[bX] = b2V[X]

The standard deviation of V is simply the square root of

24 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

5 1.737 0.977 Standard 2.600 1.000

4 1.531 0.897 0.723 0.649

5 1.737 0.977 0.932 0.910

8 1.532 0.898 0.724 0.650

9 2.550 1.289 2.143 2.764

10 0.296 0.423 0.075 0.032

13 2.810 1.389 2.682 3.726

15 - 0 .5 7 5 0.088 0.001 0.000

Mean - 0 .8 0 3 0.000 -0 .5 8 1 2.773

Variance 6.762 1.000

Standard Deviation 2.600 1.000

Hence, the skew of X is —0.581 and the kurtosis of X is 2.773.

Chapter 2 Random Variables ■ 25

The variance is Var[X] = £ (x — E[X])2 Pr(X = x)

The standard deviation is \ / Var[X] = 2.355% .

e. The median is the value where at least 50% probabil­

Thus the skewness is 0.021, and the distribution has a

26 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

3.11 If the return on a stock, R, is normally distributed with a M y) = % exp ( - % ) , y 2: 0,

44 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

N O RM .S.IN V(.95). —3.09 X a + /jl = —3.86% .

b. The same idea an be used here where z = —2.32

Chapter 3 Common Univariate Random Variables ■ 45

3.12 a. The monthly return is 2%, and the monthly standard k

3.13 In a Student's t, the kurtosis depends only on the degree

of freedom and is k = 3 ----—. This can be solved so

46 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

Small Firm (X2) Return - 1 0 % 0 % + 1 0 %

USD 100M 0 % 3.09% 6.58% 6.16% series are unrelated?

58 ■ Financial Risk Manager Exam Part I: Quantitative Analysis

Short Concept Questions

4.9 It would be if these are independent, because this is a

Xi mally distributed with 0 correlation, the sum of iid normal

USD 4M % 19.5% 41.6% 38.9%

Chapter 4 Multivariate Random Variables ■ 59

e. The median is the value where at least 50% probabil

Trial Number y 1 2 8 2.963 0.930 -0 .5 2 9