Introductory Econometrics for Finance

Chris Brooks
Solutions to Review Questions - Chapter 13

1. (a) The scope of possible answers to this part of the question is limited only by
the imagination! Simulations studies are useful in any situation where the
conditions used need to be fully under the control of the researcher (so that
an application to real data will not do) and where an analytical solution to the
problem is also unavailable. In econometrics, simulations are particularly
useful for examining the impact of model mis-specification on the properties
of estimators and forecasts. For example, what is the impact of ignored
structural breaks in a series upon GARCH model estimation and forecasting?
What is the impact of several very large outliers occurring one after another
on tests for ARCH? In finance, an obvious application of simulations, as well as
those discussed in Chapter 11, is to producing “scenarios” for stress-testing
risk measurement models. For example, what would be the impact on bank
portfolio volatility if the correlations between European stock indices rose to
one? What would be the impact on the price discovery process or on market
volatility if the number and size of index funds increased substantially?

(b) Pure simulation involves the construction of an entirely new dataset made
from artificially constructed data, while bootstrapping involves resampling
with replacement from a set of actual data.

Which technique of the two is the more appropriate would obviously depend
on the situation at hand. Pure simulation is more useful when it is necessary
to work in a completely controlled environment. For example, when examining
the effect of a particular mis-specification on the behaviour of hypothesis
tests, it would be inadvisable to use bootstrapping, because of course the
boostrapped samples could contain other forms of mis-specification. Consider
an examination of the effect of autocorrelation on the power of the regression
F-test. Use of bootstrapped data may be inappropriate because it violates one
or more other assumptions – for example, the data may be heteroscedastic or
non-normal as well. If the bootstrap were used in this case, the result would
be a test of the effect of several mis-specifications on the F-test!

Bootstrapping is useful, however, when it is desirable to mimic some of the

distributional properties of actual data series, even if we are not sure quite
what they are. For example, when simulating future possible paths for price

1
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

series as inputs to risk management models or option prices, bootstrapping is

useful. In such instances, pure simulation would be less appropriate since it
would bring with it a particular set of assumptions in order to simulate the
data – e.g. that returns are normally distributed. To the extent that these
assumptions are not supported by the real data, the simulated option price or
risk assessment could be inaccurate.

(c) Variance reduction techniques aim to reduce Monte Carlo sampling error.
In other words, they seek to reduce the variability in the estimates of the
quantity of interest across different experiments, rather like reducing the
standard errors in a regression model. This either makes Monte Carlo
simulation more accurate for a given number of replications, making the
answers more robust, or it enables the same level of accuracy to be achieved
using a considerably smaller number of replications. The two techniques that
were discussed in Chapter 11 were antithetic variates and control variates.
Mathematical details were given in the chapter and will therefore not be
repeated here.

Antithetic variates try to ensure that more of the probability space is covered
by taking the opposite (usually the negative) of the selected random draws,
and using those as another set of draws to compute the required statistics.
Control variates use the known analytical solutions to a similar problem to
improve accuracy. Obviously, the success of this latter technique will depend
on how close the analytical problem is to the actual one under study. If the
two are almost unrelated, the reduction in Monte Carlo sampling variation will
be negligible or even negative (i.e. the variance will be higher than if control
variates were not used).

(d) Almost all statistical analysis is based on “central limit theorems” and “laws
of large numbers”. These are used to analytically determine how an estimator
will behave as the sample tends to infinity, although the behaviour could be
quite different for small samples. If a sample of actual data that is too small is
used, there is a high probability that the sample will not be representative of
the population as a whole. As the sample size is increased, the probability of
obtaining a sample that is unrepresentative of the population is reduced.
Exactly the same logic can be applied to the number of replications employed
in a Monte Carlo study. If too small a number of replications is used, it is
possible that “odd” combinations of random number draws will lead to results
that do not accurately reflect the data generating process. This is increasingly
unlikely to happen as the number of replications is increased. Put another

2
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

way, the whole probability space will gradually be appropriately covered as the
number of replications is increased.

(e) Computer generated (“pseudo-”) random numbers are not random at all,
but are entirely deterministic since their generation exactly follows a formula.
In intuitive terms, the way that this is done is to start with a number (a “seed”,
usually chosen based on a numerical representation of the computer’s clock
time), and then this number is updated using modular arithmetic. Provided
that the seed and the other required parameters that control how the
updating occurs are set carefully, the pseudo-random numbers will behave
almost exactly as true random numbers would.

(f) Simulation methods are particularly useful when an analytical solution is

unavailable – for example, for many problems in econometrics, or for pricing
exotic options. In such cases, they may be the only approach available.
However, in situations where analytical results are available, simulations have
several disadvantages.

First, simulations may require a great deal of computer power. There are still
many problems in econometrics that are unsolvable even with a brand new
supercomputer! The problems seem to grow in dimension and complexity at
the same rate as the power of computers!

Second, simulations may be inaccurate if an insufficient number of replications

are used. Even if variance reduction techniques are employed, the required
number of replications to achieve acceptable accuracy could be very large.
This is especially true in the case of simulations that require accurate
estimation of extreme or “tail” events. For example, the pricing of deep out of
the money options is difficult to do accurately using simulation since most of
the replications will give a zero value for the option. Equally, it is difficult to
measure the probability of crashes, or to determine 1% critical values
accurately.

Third, a corollary of the second point is that simulations by their very nature
are difficult to replicate. The random draws used for simulations are usually
calculated based on a random number seed that is set according to the clock
time of the computer. So run two simulations, one ten minutes after the other
and you will get two completely different sets or random draws. Obviously,
again this problem would disappear if enough replications are used, but the
required number of replications may be infeasibly large.

3
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

Finally, there is a real danger that the results of a Monte Carlo study will be
specific to the particular sets of parameters investigated. An analytical result,
on the other hand, if available, may be generally applicable. The answer given
below for question 3 shows that, for a multi-dimensioned problem, a lot of
work is required to do sufficient experiments to ensure that the results are
sufficiently general.

2. Although this is a short question, a good answer would be quite involved. Recall the
null and alternatives for the Ljung-Box (LB) test:
H0: 1 = 0 and 2 = 0 and … and m = 0
H0: 1  0 or 2  0 or … or m  0
The question does not state the number of lags that should be used in the LB test, so
it would be advisable to design a framework that examined the results using several
different lag lengths. I assume that lag lengths m = 1, 5, and 10 are used, for sample
sizes T = 100, 500, 1000.

Probably the easiest way to explain what would happen is using pseudo-code. I have
written a single set of instructions that would do the whole simulation in two goes,
although it is of course possible to separate it into several different experiments (e.g.
one for size and a separate one for power, and separate experiments for each sample
size etc).

Part 1: Simulation with no GARCH as benchmark.

1. Generate a sample of length T from a standard normal distribution – call these
draws ut. The size of a test is examined by using a DGP that is correct under the null
hypothesis – in this case, we want a series that is not autocorrelated. Thus the data
for examination would be ut. Set xt = ut to avoid any later confusion.

The power of the test would be examined by using a DGP that is wrong under the null
hypothesis. Obviously, the power of the test should increase as the null becomes
more wrong. To evaluate the power of the test, generate the following data:
yt = yt-1 + ut
with  = 0.1, 0.5, and 0.8 (assume y0 = 0 in each case).

2. For each sample of xt and yt, construct the Ljung-Box test for each lag length m, and
perform the test.

3. Repeat steps 1 and 2 N times, where N is the number of replications. Assume that
N has been set to 10,000. The size of the test will be given by the percentage of times
that the null hypothesis on xt is correctly not rejected. The power of the test will be
given by the percentage of times that the null on yt is correctly rejected.

4
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

Part 2: Simulation with GARCH.

Repeat steps 1 to 3 above exactly, but generate the data so that ut follows a
GARCH(1,1) process. This would be achieved, given that ut has been drawn. Some
parameters for the GARCH process would have to also be assumed. For xt, use the
equations:
t2 = 0.0001 + 0.1xt-12 + 0.8t-12, with 02 = 0.001.
xt = tut

and for yt:

t2 = 0.0001 + 0.1xt-12 + 0.8t-12, with 02 = 0.001.
yt = yt-1 + tut

Finally, the effect of GARCH would be investigated by comparing the size and power
under parts 1 and 2. If the size increases from its nominal value, or the power falls,
when the GARCH is added, it would be concluded that GARCH does have an adverse
effect on the LB test.

3. (a) This would be a very easy simulation to do. The first thing would be to
choose the appropriate values of  to use. It would be tempting to choose a
spread of values, say from 0 to 1 in units of 0.2, (i.e. 0, 0.2, .., 0.8, 1), but this
would not be optimal since all of the interesting behaviour will occur when 
gets close to 1. Therefore the values are better skewed towards 1, e.g. 0, 0.5,
0.8, 0.9, 0.95, 0.99, and 1. This gives 7 values of  which should give the
flavour of what happens without the results being overwhelming or the
exercise becoming too tedious.

Note that the question says nothing about the sample sizes to use and in fact
the impact of a unit root (i.e.  = 1) will not disappear asymptotically. A good
researcher, however, would choose a range of sample sizes that are empirically
relevant (e.g. 100, 500, and 2000 observations) and would conduct the
simulation using all of them. Assuming that a sample size of 500 is used, the
next step would be to generate the random draws. A good simulation would
generate more than 500 draws for each replication, to allow for some start-up
observations that are later discarded. Again, nothing is stated in the question
about what distribution should be used to generate the random draws. Unless
there is a particular reason to do otherwise (for example, if the impact of fat
tails is of particular importance), it is common to use a standard normal
distribution for the disturbances. We also need to select a starting value for y
– call this y1, and set this starting value to zero (note: the reason why we allow

5
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

for start-up observations is to minimise the impact of this choice of initial

value for y).

Once a set of disturbances are drawn, and the initial value for y is chosen, the
next stage is to recursively construct a series that follows the required AR(1)
model:
y2 = y1 + u2
y3 = y2 + u3

y500 = y499 + u500
The next stage would be to estimate an AR(1), and to construct and save the t-
ratios on the  coefficient for each replication. Note that this estimated
regression should include an intercept to allow for a simulated series with a
non-zero mean, even though y has a zero mean under the DGP. It is probably
sensible and desirable to use the same set of random draws for each
experiment that has different values of .

Producing the EViews or RATS code for this problem is left for readers and
should not be difficult given the examples in the chapter! However, some
pseudo-code could be
SET NUMBER OF REPS
SET SAMPLE SIZE
SET VALUES OF PHI – CALL THEM PHI1 TO PHI7
CLEAR ARRAYS FOR Y USED FOR EACH VALUE OF PHI – Y1 TO Y7
SET INITIAL VALUES FOR THE Y’S TO ZER0 – Y1(1)=0, Y2(1)=0, ETC
START REPLICATIONS LOOP
GENERATE RANDOM DRAWS U FOR SAMPLE SIZE PLUS 200 START-UP
OBSERVATIONS
START RECURSIVE Y GENERATION LOOP WITH J=2
Y1(J)=PHI1*Y1(J-1)+U(J)
Y2(J)=PHI2*Y2(J-1)+U(J)
Y3(J)=PHI3*Y3(J-1)+U(J)
Y4(J)=PHI4*Y4(J-1)+U(J)
Y5(J)=PHI5*Y5(J-1)+U(J)
Y6(J)=PHI6*Y6(J-1)+U(J)
Y7(J)=PHI7*Y7(J-1)+U(J)
END Y GENERATION LOOP
RUN A LINEAR REGRESSION OF Y1 ON A CONSTANT AND THE LAGGED
VALUE OF Y1
AND REPEAT THIS FOR EACH OF Y2, Y3 ETC
SAVE THE SLOPE COEFFICIENT T-RATIOS AND STORE THESE

6
© Chris Brooks 2014
Introductory Econometrics for Finance by Chris Brooks

END REPLICATIONS LOOP

This will yield a set of estimated t-ratios for each of the values of . These
could be plotted or summarised as desired.

(b) The broad structure of the appropriate simulation for part (b) would be
similar to that for part (a), but there is now the added dimension of the size of
the sample for each replication. This is therefore a multi-dimensioned
problem: we want to produce a simulated series for each value of  and for
each sample size. The easiest way to achieve this if the answer to part (a) had
already been constructed would be to run separate simulations as above for
different sample sizes. Note also that in part (b), it is the estimated values of
the  coefficients rather than their t-ratios that are of interest. What we would
expect is that the average of the  estimates across the replications would
converge upon their actual value as the sample size increases.

4. Again, this is a fairly simple exercise given the code that was presented in the
chapter. All that is required is to add a clause in to the recursive generation of the
path of the underlying asset to say that if the price at that time falls below the barrier,
then the value of the knock-out call for that replication is zero. It would also save
computational time if this clause also halted the simulation of the path for that
replication and went straight on to the next replication.

7
© Chris Brooks 2014