Nonlife Actuarial Models

Chapter 15
Applications of Monte Carlo Methods

Learning Objectives
1. Monte Carlo estimation of critical values and p-values
2. Bootstrap estimation of p-values
3. Bootstrap estimation of bias and mean squared error.
4. Simulation of lognormally distributed asset prices
5. Simulation of asset prices with discrete jumps

15.1 Monte Carlo Simulation for Hypotheses Test

15.1.1 Kolmogorov-Smirnov Test
For the Kolmogorov-Smirnov D statistic, David and Johnson (1948)
show that if the parameters estimated for the null distribution are
parameters of scale or location, and the estimators satisfy certain
general conditions, then the joint distribution of the probabilityintegral transformed observations of the sample will not depend on
the true parameter values.
The Kolmogorov-Smirnov test is based on the distribution function
under the null, which is the probability integral transform of the
sample observations.

 Many commonly used distributions involve parameters of scale and

location. For example, for the N (, 2 ) distribution, is the location
parameter and is the scale parameter. The parameter in the E()
distribution is a location-and-scale parameter.
In these cases the exact distributions of the D statistics under the
null do not depend on the true parameter values, as long as the null
distribution functions are computed using the MLE.
As the null distribution of the D statistic for the normal distribution does not depend on the true parameter values, we may assume
any convenient values of the parameters without aecting the null
distribution.
This gives rise to the following Monte Carlo procedure to estimate
the critical value of D for a given sample size n
4

1. Generate a random sample of n (call this the estimation sample size) standard normal variates x1 , , xn . Calculate the sample
mean x and sample variance s2 , and use these values to compute
the estimated distribution function F (xi ), where F () is the df of
N (
x, s2 ). Then use equation (13.4) to compute D.
2. Repeat Step 1 m times (call this the Monte Carlo sample size)
to obtain m values of Dj , for j = 1, , m.
3. At the level of significance , the critical value of the KolmogorovSmirnov D statistic is computed as the (1)-quantile of the sample
of m values of D, estimated using the method in equations (11.9)
and (11.10).
The following critical values are proposed by Lilliefors (1967) for
testing normal distributions with unknown mean and variance
5

Level of significance

0.10

0.05

0.01

Critical value

0.805

0.886

1.031
.
n

If the null hypothesis is that the sample observations are distributed

as E(), where is not specified, to estimate the critical values of
the Kolmogorov-Smirnov statistic, the following procedure can be
used
1. Generate a random sample of n variates x1 , , xn distributed as
E(1). Calculate the sample mean x and compute the estimated
distribution function F (xi ), where F () is the df of E(1/
x). Then
use equation (13.4) to compute D.
2. Repeat Step 1 m times to obtain m values of Dj , for j = 1, , m.
6

3. At the level of significance , the critical value of the KolmogorovSmirnov D statistic is computed as the (1)-quantile of the sample
of m values of D, estimated using the method in equations (11.9)
and (11.10).
The following critical values are proposed by Lilliefors (1969) for
testing exponential distributions with unknown mean
Level of significance

0.10

0.05

0.01

Critical value

0.96

1.06

1.25
.
n

15.1.2 Chi-square Goodness-of-Fit Test

The asymptotic distribution of the X 2 statistic for the goodnessof-fit test is 2kr1 , where k is the number of groups and r is the
number of parameters estimated using the MMLE method.
7

 This result holds asymptotically for any null distribution. Yet Monte
Carlo simulation can be used to investigate the performance of the
test and improve the estimates of the critical values in small samples
if required.
Example 15.3: Estimate the critical values of the chi-square goodnessof-fit statistic X 2 using Monte Carlo simulation when the null hypothesis
is that the observations are distributed as E(), where is unknown.
Compute the X 2 statistics based on the MLE using individual observations
as well as the MMLE using grouped data.
Solution:
We group the data into intervals (ci1 , ci ], and use the following 4 intervals: (0, 0.4], (0.4, 1], (1, 1.5] and (1.5, ). The MLE of
using the complete individual data is 1/
x. Let n = {n1 , , n4 }, where
ni is the number of observations in the ith interval. Using grouped data,
8

the MMLE is solved by maximizing the log-likelihood function

log L(; n) =

4
X
i=1

ni log [exp(ci1 ) exp(ci )]

with respect to . The X 2 statistic is then computed using equation

(13.8). Using a Monte Carlo simulation with 10,000 samples, we obtain
the estimated critical values of the X 2 statistic summarized in Table 15.3.
Table 15.3:

0.10
0.05
0.01

Results of Example 15.3

n = 50
MLE
MMLE
4.95
4.70
6.31
6.07
9.45
9.25

n = 100
MLE
MMLE
4.93
4.70
6.30
6.07
9.48
9.39

n = 200
MLE
MMLE
4.91
4.61
6.38
6.05
9.60
9.37

n = 300
MLE
MMLE
4.91
4.66
6.30
6.04
9.41
9.14

22,1
4.61
5.99
9.21

The asymptotic critical values 22,1 are shown in the last column. Two
points can be observed from the Monte Carlo results. First, the asymptotic
9

results are very reliable even for samples of size 50, if the correct MMLE
is used to compute X 2 . Second, if MLE is used to compute X 2 , the use
of 22,1 as the critical value will over-reject the null hypothesis.
2

10

15.2 Bootstrap Estimation of p-Value

There are situations for which the distribution of the test statistic
under the null hypothesis depends on some nuisance parameters not
specified under the null. For such problems, tabulation of the critical
values is not viable.
As an alternative, we may use bootstrap method to estimate the
p-value of the test statistic.
Consider a sample of n observations x = (x1 , , xn ) and a test
statistic T (x) for testing a null hypothesis H0 .
Let the computed value of the test statistic for the sample x be t.
Suppose the decision rule of the test is to reject H0 when t is too
large (i.e., on the right-hand extreme tail).
11

 Assume H0 contains a nuisance parameter , which is not specified.

We now consider the estimation the p-value of the test statistic,
which is the probability that T (x) is larger than t if the null hypothesis is true, i.e.,
p = Pr(T (x) > t | H0 ).

(15.1)

As H0 contains the nuisance parameter , we replace the above problem by

p = Pr(T (x) > t | H0 ()),
(15.2)

where is an estimator of . The bootstrap estimate of p can be

computed as follows

1. Let the computed value of T (x) based on the sample x be t, and

let the estimated value of be , which may be any appropriate
estimator, such as the MLE.
12

2. Generate a sample of observations from the distributional assumption of H0 (), call this x . Compute the test statistic using data x
and call this t .
3. Repeat Step 2 m times, which is the bootstrap sample size, to obtain
m values of the test statistic tj , for j = 1, , m.
4. The estimated p-value of t is computed as
1 + number of {tj t}
.
m+1

(15.3)

The above is a parametric bootstrap procedure, in which the

samples x are generated from a parametric distribution.
At level of significance , the null hypothesis is rejected if the estimated p-value is less than .
13

Example 15.4: You are given the following 20 observations of losses

0.114, 0.147, 0.203, 0.378, 0.410, 0.488, 0.576, 0.868, 0.901, 0.983,
1.049, 1.555, 2.060, 2.274, 4.235, 5.400, 5.513, 5.817, 8.901, 12.699.
(a) Compute the Kolmogorov-Smirnov D statistic, assuming the data
are distributed as P(, 5). Estimate the p-value of the test statistic
using bootstrap.
(b) Repeat (a), assuming the null distribution is P(, 40).
(c) Repeat (a), assuming the null distribution is E().
Solution:
For (a), we estimate using MLE, which, from Example
12.9, is given by
20

= P20
,
i=1 log(xi + 5) 20 log(5)
14

and we obtain
= 2.7447. The computed D statistic is 0.1424.
To estimate the p-value, we generate 10,000 bootstrap samples of size 20
each from P(2.7447, 5), estimate and compute the D statistic for each
sample. The proportion of the D values larger than 0.1424 calculated
using equation (15.3) is 0.5775, which is the estimated p-value. Thus,
the P(, 5) assumption cannot be rejected at any conventional level of
significance.
For (b), the MLE of is
20

= P20
= 15.8233.
log(x
+
40)

20
log(40)
i
i=1

The computed D statistic is 0.2138. We generate 10,000 samples of size

20 each from the P(15.8233, 40) distribution and compute the D statistic
of each sample. The estimated p-value is 0.0996. Thus, at the level of
15

significance of 10%, the null hypothesis P(, 40) is rejected, but not at
the level of significance of 5%.
For (c), the MLE of is
1

= = 0.3665,
x
and the computed D value is 0.2307. We generate 10,000 samples of size
20 each from the E(0.3665) distribution using the inversion method. The
estimated p-value of the D statistic is 0.0603. Thus, the assumption of
E() is rejected at the 10% level, but not at the 5% level.
To conclude, the Kolmogorov-Smirnov test supports the P(, 5) distribution assumption for the loss data, but not the P(, 40) and E() distributions.
2

16

15.3 Bootstrap Estimation of Bias and

Mean Squared Error
Bootstrap method can also be used to estimate the bias and mean
squared error of the parameter estimates of a distribution.
Consider the estimation of the parameter (or a function of the
parameter g()) of a distribution using an estimator (or g()),
given a random sample of n observations x = (x1 , , xn ) of X.
In situations where theoretical results about the bias and mean
squared error of (or g()) are intractable, we may use bootstrap
method to estimate these quantities.
When no additional assumption about the distribution of X is made,
17

we may use the empirical distribution define by x as the assumed distribution. We generate a sample of n observations x = (x1 , , xn )
by re-sampling from x with replacement, and compute the estimate
(or g( )) based on x .

We do this m times to obtain m estimates j (or g(j )), for j =

1, , m.

Based on these bootstrap estimates we can compute the bias and

the mean squared error of the estimator (or g()). As x are
generated from the empirical distribution defined by x, we call this
method nonparametric bootstrap.
To illustrate the idea, we consider the use of the sample mean and
the sample variance as estimates of the population mean and population variance 2 of X, respectively.
18

 Let E and 2E be the mean and the variance, respectively, of the

empirical distribution defined by x.
We note that E = x and 2E = (n 1)s2 /n, where x and s2 are the
sample mean and the sample variance of x, respectively
. To use the bootstrap method to estimate the bias and the mean
squared error of x and s2 , we adopt the following procedure
1. Generate a random sample of n observations by re-sampling with
replacement from x, call this x = (x1 , , xn ). Compute the mean
x and variance s2 of x .
2. Repeat Step 1 m times to obtain values xj and s2
j , for j = 1, , m.
3. The bias and the mean squared error of x are estimated, respectively,

19

by

m
1 X
(
xj E )
m j=1

and

m
1 X
(
xj E )2 .
m j=1

(15.4)

4. The bias and the mean squared error of s2 are estimated, respectively, by
m
1 X
2
(s2

E)
m j=1 j

and

m
1 X
2 2
(s2

E) .
m j=1 j

(15.5)

It is theoretically known that x and s2 are unbiased for and 2 ,

respectively.
Furthermore, the expected value of xj is E and the expected value
2
of s2
is

j
E , so that the bootstrap estimate of the biases should
converge to zero when m is large.

20

 The mean squared error of x is

2
MSE(
x) = Var(
x) = ,
n

(15.6)

which is unknown (as 2 is unknown).

On the other hand, the bootstrap estimate of the MSE of x in equation (15.4) converges in probability to 2E /n, which is known given
x. However, when x varies E( 2E /n) = (n 1) 2 /n2 6= MSE(
x).

21

15.4 A General Framework of Bootstrap

We now provide a framework of the theoretical underpinning of the
bootstrap method.
Let X = {X1 , , Xn } be independently and identically distributed
as X with df F (), which may depend on a parameter .
Suppose = (F ) is a quantity of the distribution (e.g., mean,
median, a quantile or a population proportion) and = (X) is an
estimate of based on X.
We define

(X; F ) = (X) (F ),

(15.8)

which is the error in estimating using . Denoting EF as the

22

expectation taken using the df F , the bias of is

EF [(X; F )] = EF [(X) (F )]

(15.9)

and the mean squared error of is

EF [(X; F )2 ] = EF [((X) (F ))2 ].

(15.10)

For another application, let T (X) be a test statistic for a hypothesis H0 and its value computed based on a specific sample x =
(x1 , , xn ) be t = T (x). We now define
(X; F ) = T (X) t.

(15.11)

If H0 is rejected when t is too large, the p-value of the test is

Pr(T (X) t > 0 | F ) = Pr((X; F ) > 0 | F ).
23

(15.12)

 In the above cases, we are interested in the expectation or the population proportion of a suitably defined function (X; F ). This set-up
includes the evaluation of bias and mean squared error of an estimator and the p-value of a test, as well as many other applications.
As F is unknown in practice, the quantities in equations (15.9),
(15.10) and (15.12) cannot be evaluated.
However, we may replace F by a known df F and consider instead
the quantities
EF [(X; F )] = EF [(X) (F )],

(15.13)

EF [(X; F )2 ] = EF [((X) (F ))2 ],

(15.14)

Pr(T (X) t > 0 | F ) = Pr((X; F ) > 0 | F ).

(15.15)

and

24

 The above quantities are called the bootstrap approximations.

The reliability of these approximations depend on how good F is
as an approximation to F .
If F is taken as the empirical distribution defined by x, we have a
nonparametric bootstrap.
If F is taken as F () for a suitable estimator computed from the
sample x, then we have a parametric bootstrap.
As (X) and T (X) may be rather complex functions of X, the evaluation of equations (15.13), (15.14) and (15.15) may remain elusive
even with known or given F .
In the case where the sample size n is small and the empirical distribution is used for F , we may evaluate these quantities by exhausting
25

all possible samples of X.

This approach, however, will not be feasible when n is large or when
a parametric df F () is used. In such situations the quantities may
be estimated using Monte Carlo methods, and we call the solution
the Monte Carlo estimate of the bootstrap approximate, or simply
the bootstrap estimate.

26

15.5 Monte Carlo Simulation of Asset Prices

15.5.1 Wiener Process and Generalized Wiener Process
Let Wt be a stochastic process over time t with the following properties
1. Over a small time interval t, the change in Wt , denoted by Wt =
Wt+t Wt , satisfies the property

Wt = " t,

(15.16)

where " N (0, 1).

2. If Wt1 and Wt2 are changes in the process Wt over two nonoverlapping intervals, then Wt1 and Wt2 are independent.
27

 A continuous-time stochastic process satisfying the above two properties is called a Wiener process or standard Brownian motion.
From the first of these two properties, we can conclude that
E(Wt ) = 0,

(15.17)

Var(Wt ) = t.

(15.18)

and

For the change over a finite interval [0, T ], we can partition the
interval into N nonoverlapping small segments of length t each,
such that
T = N(t)
(15.19)
and
WT W0 =

N1
X

Wi(t) =

i=0

28

X
i

i=1

t,

(15.20)

where i , for i = 1, , n, are iid N (0, 1).

Thus, given the information at time 0 we have
E(WT ) = W0 +

N
X

E( i ) t = W0 ,

(15.21)

i=1

and
Var(WT ) =

N
X

Var( i )t =

i=1

N
X

t = T.

(15.22)

i=1

Hence, WT is the sum of W0 and N iid normal variates, which implies, given W0 ,
WT N (W0 , T ).
(15.23)
The Wiener process can be extended to allow for a drift in the
process and a constant volatility parameter.
29

 A generalized Wiener process or Brownian motion Xt has the

following change over a small time interval t

Xt = a t + b Wt ,

(15.24)

where a is the drift rate and b is the volatility rate (a and b are
constants), and Wt is a Wiener process.
It can be verified that, given X0 , we have
XT N (X0 + aT, b2 T ),
for any finite T .

30

(15.25)

 The Wiener and generalized Wiener processes are continuous-time

processes when t 0 in equations (15.16) and (15.24), respectively. Thus, we shall write the dierentials of these processes as
dWt and
dXt = a dt + b dWt ,
(15.26)
respectively.
15.5.2 Diusion Process and Lognormal Distribution
A further extension of equation (15.26) is to allow the drift and
volatility rates to depend on time t and the process value Xt . Thus,
we consider the process
dXt = a(Xt , t) dt + b(Xt , t) dWt ,
which is called an Ito process or diusion process.
31

(15.27)

 The terms a(Xt , t) and b(Xt , t) are called the drift rate and the
diusion coecient, respectively.
Xt is in general no longer normally distributed.
We consider a specific member of diusion processes, called the geometric Brownian motion.
Let St be the price of an asset at time t. St is said to follow a
geometric Brownian motion if
dSt = St dt + St dWt ,

(15.28)

where , called the instantaneous rate of return, and , called

the volatility rate, are constants.

32

 The above equation can also be written as

1
dSt = dt + dWt .
St

(15.29)

Further analysis (using Itos lemma) shows that

d log St =
2

dt + dWt ,

(15.30)

so that log St follows a generalized Wiener process and hence is normally distributed. Thus, following equation (15.25), we conclude

log St N log S0 +

t, 2 t ,

(15.31)

so that St is lognormally distributed with mean

"

E(St ) = exp log S0 +

2
33

t
t+
= S0 exp(t).
2

(15.32)

Equation (15.31) can also be written as

St
log St log S0 = log
S0
so that

St
log
S0

=
2

t, 2 t ,

t + tZ,

(15.33)

(15.34)

where Z is standard normal.

Note that

St
1
R log
(15.36)
t
S0
is the continuously compounded rate of return over the interval
[0, t].

Thus, from equation (15.34), the expected continuously compounded

34

rate of return over the finite interval [0, t] is

1
St
E log
t
S0

2
= ,
2

(15.37)

which is less than the instantaneous rate of return .

The total return of an asset consists of two components: capital gain
and dividend yield. As St is the price of the asset, as defined in
equation (15.28) captures the instantaneous capital gain only.
If the dividend yield is assumed to be continuous at the rate , then
the total instantaneous return, denoted by , is given by
= + .

(15.38)

Hence, expressed in terms of the total return and the dividend yield,
35

the expected continuously compounded rate of capital gain (assetprice appreciation) is

1
St
E log
t
S0

2
2

=
= .
2
2

(15.39)

We now consider the simulation of asset prices that follow the geometric Brownian motion given in equation (15.28), in which the
parameter captures the return due to asset-price appreciation.
From equation (15.34), we obtain
St = S0 exp

"

t + tZ ,

(15.40)

which can be used to simulate price paths of the asset. In practical

applications, we need the values of the parameters and
2
R = .
2
36

(15.41)

 Suppose we sample return data of the asset over intervals of length

h. Let there be n return observations (computed as dierences of
logarithmic asset prices) with mean xR and sample variance s2R .
The required parameter estimates are then given by

R =

xR
h

and

sR

= .
h

(15.42)

The asset prices at intervals of h can be simulated recursively using

the equation
St+(i+1)h = St+ih exp [
xR + sR Zi ] ,

for i = 0, 1, 2, ,

(15.43)

where Zi are iid N (0, 1).

We use end-of-day S&P500 index values in the period January 3,
2007, through December 28, 2007.
37

 There are in total 250 index values and we compute 249 daily returns,
which are the logarithmic price dierences.
The price index graph and the return graph are plotted in Figure
15.1.
We estimate the parameters of the price process and obtain xR =
0.0068% and sR = 1.0476%.
These values are in percent per day. If we take h = 1/250, the

annualized estimate of is 1.0476 250% = 16.5640% per annum.

The estimate of is

2
xR s2R

=
R +
=
+
= 3.0718% per annum.
2
h
2h

(15.44)

We use these values to simulate the price paths, an example of which

is presented in Figure 15.2.
38

S&P500 index

S&P500 return

1600

4
2
Return in %

Index value

1550
1500
1450

1400
1350

51
101
151
201
2007/01/04 2007/12/28

Histogram of S&P500 return

51
101
151
201
2007/01/04 2007/12/28
Normal probability plot

100

Probability

Frequency

80
60
40
20
0
S&P500 return

Data

Return of simulated series

1700

2
Return in %

1600
1500

0
2

1400
1

51
101
151
201
2007/01/04 2007/12/28

Histogram of simulated return

51
101
151
201
2007/01/04 2007/12/28
Normal probability plot

60
50
40

Probability

Frequency

Simulated value

Simulated price series

1800

30
20
10
0
Return of simulated series

Data

15.5.3 Jump-Diusion Process

Asset prices following a diusion process are characterized by paths
that are continuous in time.
Anecdotal evidence, however, often suggests that stock prices are
more jumpy than what would be expected of a diusion process.
To allow for discrete jumps in asset prices, we introduce a jump component into the diusion process and consider asset prices following
a jump-diusion process.
We consider augmenting the geometric Brownian motion with a
jump component. We assume the occurrence of a jump in an interval has a Poisson distribution, and when a jump occurs, the jump
size is distributed normally.
39

 We define Nt as a Poisson process with intensity (mean per unit

time) .
Nt is the number of jump events occurring in the interval (t, t+t].
We use the notation dNt when t 0.
We now augment the geometric Brownian motion in equation (15.30)
with a jump component as follows

d log St =
2

dt + dWt + Jt dNt J dt,

(15.45)

where Jt N (J , 2J ) and is distributed independently of Nt .

Note that the mean of Jt dNt is J dt, so that the mean of the
augmented component Jt dNt J dt is zero. Thus, the addition of
the term J dt is to center the jump component so that its mean
is equal to zero.
40

 This property is of important significance because jumps are often

assumed to be idiosyncratic and does not aect the expected return
of the stock.
We re-write equation (15.45) as

d log St = J
2

dt + dWt + Jt dNt .

(15.46)

If J > 0, the jump component induces price appreciation on average, and the diusion part of the price will have a drift term adjusted
downwards. On the other hand, if J < 0, investors will be compensated by a higher drift rate to produce the same expected return.
To simulate the jump-diusion process defined in equation (15.46)
we first consider the jump component.
41

 Suppose the time interval of the prices simulated is h, to simulate the

jump component Jt dNt we generate a number m from the PN (h)
distribution, and then simulate m independent variates Zi from the
N (0, 1) distribution. The jump component is then given by
mJ + J

m
X

Zi .

(15.47)

i=1

The diusion component is computed as

J
2

h + hZ,

(15.48)

where Z is a standard normal variate independent of Zi .

To generate the value of St+(i+1)h given St+ih we use the equation
"
!
#
"
#
m
2

h + hZ exp mJ + J
Zi .
St+(i+1)h = St+ih exp J
2
i=1
(15.49)
42

 For illustration we simulate a jump-diusion process using the following parameters: = 3.0718%, = 16.5640%, = 3, J = 2%
and J = 3% (the first 3 quantities are per annum).
Thus, the jumps occur on average 3 times per year, and each jump is
normally distributed with mean jump size of 2% down and standard
deviation of 3%. To observe more jumps in the simulated process,
we simulate 500 daily observations (about 2 years) and an example
is presented in Figure 15.3.

43

Return of simulated series

4

1700

2
Return in %

Simulated value

Simulated series with jumps

1800

1600
1500
1400
1300

0
2
4
6

101

201
301
Time

401

Histogram of simulated return

101

201
301
Time

Normal probability plot

200

Probability

Frequency

150
100
50
0
Return of simulated series

401

Data