12

Monte Carlo Simulation

Chapter

and Risk Analysis

iQoncept/Shutterstock.com

Learning Objectives
After studying this chapter, you will be able to:

• Explain the concept and importance of analyzing risk • Explain the significance of the “flaw of averages.”
in business decisions. • Conduct Monte Carlo simulation using historical data
• Use data tables to conduct simple Monte Carlo and resampling techniques.
simulations. • Use fitted distributions to define uncertain variables in
• Use Analytic Solver Platform to develop, implement, a simulation.
and analyze Monte Carlo simulation models. • Define and use custom distributions in Monte Carlo
• Compute confidence intervals for the mean value of simulations.
an output in a simulation model. • Correlate uncertain variables in a simulation model
• Construct and interpret sensitivity, overlay, trend, using Analytic Solver Platform.
and box-whisker charts for a simulation model.

403
404 Chapter 12  Monte Carlo Simulation and Risk Analysis

For many of the predictive decision models we developed in

Chapter 11, all the data—particularly the uncontrollable inputs—were assumed
to be known and constant. Other models, such as the newsvendor, overbook-
ing, and retirement-planning models, incorporated ­uncontrollable inputs, such as
customer demand, hotel cancellations, and annual returns on investments, which
exhibit random behavior. We often assume such variables to be constant to sim-
plify the model and the analysis. However, many situations dictate that random-
ness be explicitly incorporated into our models. This is usually done by specifying
probability distributions for the appropriate uncontrollable inputs. As we noted
earlier in this book, models that include randomness are called s­ tochastic, or
probabilistic, models. These types of models help to evaluate risks associated
with undesirable consequences and to find optimal decisions under uncertainty.
Risk is the likelihood of an undesirable outcome. It can be assessed by
evaluating the probability that the outcome will occur along with the severity of
the outcome. For example, an investment that has a high probability of losing
money is riskier than one with a lower probability. Similarly, an investment that
may result in a $10 million loss is certainly riskier than one that might result in
only a $10,000 loss. In assessing risk, we could answer questions such as, What
is the probability that we will incur a financial loss? How do the probabilities of
different potential losses compare? What is the probability that we will run out
of inventory? What are the chances that a project will be completed on time?
Risk analysis is an approach for developing “a comprehensive understanding
and awareness of the risk associated with a particular variable of interest (be it a
payoff measure, a cash flow profile, or a macroeconomic forecast).”1 Hertz and
Thomas present a simple s­ cenario to illustrate the concept of risk analysis:

The executives of a food company must decide whether to launch a new pack-
aged cereal. They have come to the conclusion that five factors are the deter-
mining variables: advertising and promotion e
­ xpense, total cereal market, share
of market for this product, operating costs, and new capital investment. On the
basis of the “most likely” estimate for each of these variables, the picture looks
very bright—a healthy 30% return, indicating a significantly positive e
­ xpected
net present value. This future, however, depends on each of the “most likely”
estimates coming true in the actual case. If each of these ­“educated guesses”
has, for example, a 60% chance of b
­ eing correct, there is only an 8% chance
that all five will be correct (0.60 * 0.60 * 0.60 * 0.60 * 0.60) if the factors
are assumed to be independent. So the “expected” return, or present value

1David B. Hertz and Howard Thomas, Risk Analysis and Its Applications (Chichester, UK: John Wiley &
Sons, Ltd., 1983): 1.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 405

measure, is actually dependent on a rather unlikely coincidence. The decision

maker needs to know a great deal more about the other values used to make
each of the five estimates and about what he stands to gain or lose from vari-
ous combinations of these values.2

Thus, risk analysis seeks to examine the impacts of uncertainty in the estimates and
their potential interaction with one another on the output variable of interest. Hertz
and Thomas also note that the challenge to risk analysts is to frame the output of
risk analysis procedures in a manner that makes sense to the manager and provides
clear insight into the problem, suggesting that simulation has many advantages.
In this chapter, we discuss how to build and analyze models involving
­uncertainty and risk using Excel. We then introduce ­Analytic Solver Platform to
implement Monte Carlo simulation. We wish to point out that the topic of simula-
tion can fill an entire book. An entirely different area of simulation, which we do not
address in this book, is the simulation of dynamic systems, such as waiting lines,
inventory systems, manufacturing systems, and so on. This requires different mod-
eling and implementation tools, and is best approached using commercial software.
Systems simulation is an important tool for analyzing operations, whereas Monte
Carlo simulation, as we describe it, is focused more on financial risk analysis.

Spreadsheet Models with Random Variables

In Chapter 5, we described how to sample randomly from probability distributions and

to generate certain random variates using Excel tools and functions. We will use these
­techniques to show how to incorporate uncertainty into decision models.

Example 12.1 
Incorporating Uncertainty in the Outsourcing Decision Model
Refer back to the outsourcing decision model we intro- standard_deviation), as described in Chapter 5, to gen-
duced in Chapter 1 and for which we ­developed an Excel erate random values of the demand (Production Volume)
model in Chapter 11. The model is shown again in Fig- by replacing the input in cell B12 of the spreadsheet with
ure 12.1. Assume that the production volume is uncertain. the formula = ROUND(NORM.INV (RAND( ), 1000, 100),0).
We can model the demand as a random variable having The ROUND function is used to ensure that the values will
some probability distribution. Suppose the manufacturer be whole numbers. Whenever the F9 key is pressed (on a
has enough data and information to assume that demand Windows PC) or the Calculate Now button is clicked from
(production volume) will be normally distributed with a the Calculation group in the Formula tab, the worksheet
mean of 1,000 and a standard deviation of 100. We could will be recalculated, and the value of ­demand will change
use the Excel function NORM.INV (­probability, mean, randomly.

Monte Carlo Simulation

Monte Carlo simulation is the process of generating random values for uncertain inputs
in a model, computing the output variables of interest, and repeating this process for many
2Ibid., 24.
406 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.1

Outsourcing Decision Model

Spreadsheet

trials to understand the distribution of the output results. For example, in the o­ utsourcing
decision model, we can randomly generate the production volume and compute the cost dif-
ference and associated decision and then repeat this for some number of trials. Monte Carlo
simulation can easily be accomplished on a spreadsheet using a data table.

Example 12.2 
Using Data Tables for Monte Carlo Spreadsheet Simulation
Figure 12.2 shows a Monte Carlo simulation for the outsourc- The small number of trials that we used in this exam-
ing decision model (Excel file Outsourcing Decision Monte ple makes sampling error an important issue. We could
Carlo Simulation Model). First, construct a data table (see easily obtain significantly different results if we repeat the
Chapter 11) by listing the number of trials down a column simulation (by pressing the F9 key on a Windows PC). For
(here we used 20 trials) and referencing the cells associated example, repeated simulations yielded the following per-
with demand, the difference, and the decision in cells E3, centages for outsourcing as the best decision: 40%, 60%,
F3, and G3, respectively (i.e., the formula in cell E3 is =B12; 65%, 45%, 75%, 45%, and 35%. There is considerable
in cell F3, = B19; and in cell G3, = B20). Select the range variability in the results, but this can be reduced by using
of the table (D3:G23)—and here’s the trick—in the Column a larger number of trials.
Input Cell field in the Data Table dialog, enter any blank cell To understand this variability better, let us construct
in the spreadsheet. This is done because the trial number a confidence interval for the proportion of decisions
does not relate to any parameter in the model; we simply that result in a manufacturing recommendation with the
want to repeat the spreadsheet recalculation independently sample size (number of trials) n = 20 using the data in
for each row of the data table, knowing that the demand will Figure 12.2. Using formula (6.4) from Chapter 6, a 95%
change each time because of the use of the RAND function confidence interval for the proportion is 0.55 ± 1.96
in the demand formula. 0.551 0.452
= 0.55 ± 0.22, or [0.33, 0.77]. Because the
As you can see from the results, each trial has a B 20
randomly generated demand. The data table process CI includes values below and above 0.5, this suggests that
substitutes these demands into cell B12 and finds the as- we have little certainty as to the best decision. However,
sociated difference and decision in columns F and G. The if we obtained the same proportion using 1,000 trials, the
average difference is $535, and 55% of the trials resulted 0.551 0.452
confidence interval would be 0.55 ± 1.96 =
in outsourcing as the best decision; the histogram shows B 1000
the distribution of the results. These results might suggest 0.55 ± 0.03, or [0.52, 0.58]. This would indicate that we
that, although the future demand is not known, the manu- would have confidence that outsourcing would be the
facturer’s best choice might be to outsource. However, better decision more than half the time.
there is a risk that this may not be the best decision.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 407

Figure 12.2
Although the use of a data table illustrates how we can apply Monte Carlo sim-
Monte Carlo Simulation of the ulation to a decision model, it is impractical to apply to more complex problems.
Outsourcing Decision Model For example, in the Moore Pharmaceuticals model in Chapter 11, many of the
model parameters, such as the initial market size, project costs, market-size growth
factors, and market-share growth rates, may all be uncertain. In addition, we need
to be able to capture and save the results of thousands of trials to obtain good
statistical results, and it would be useful to construct a histogram of the results and
calculate a variety of statistics to conduct further analyses. Fortunately, sophisti-
cated software approaches that easily perform these functions are available. The
remainder of this chapter is focused on learning to use Analytic Solver Platform
software to perform large-scale Monte Carlo simulation. We will start with the simple
outsourcing decision model.

Monte Carlo Simulation Using Analytic Solver Platform

To use Analytic Solver Platform, you must perform the following steps:

1.
Develop the spreadsheet model.
2.
Determine the probability distributions that describe the uncertain inputs in
your model.
3.
Identify the output variables that you wish to predict.
4.
Set the number of trials or repetitions for the simulation.
5.
Run the simulation.
6.
Interpret the results.

Defining Uncertain Model Inputs

When model inputs are uncertain, we need to characterize them by some probability
distribution. For many decision models, empirical data may be available, either in his-
torical records or collected through special efforts. For example, maintenance ­records
408 Chapter 12  Monte Carlo Simulation and Risk Analysis

might provide data on machine failure rates and repair times, or observers might col-
lect data on service times in a bank or post office. This provides a factual basis for
choosing the appropriate probability distribution to model the input variable. We can
identify an appropriate distribution by fitting historical data to a theoretical model, as
we illustrated in Chapter 5.
In other situations, historical data are not available, and we can draw upon the prop-
erties of common probability distributions and typical applications that we discussed in
Chapter 5 to help choose a representative distribution that has the shape that would most
reasonably represent the analyst’s understanding about the uncertain variable. For exam-
ple, a normal distribution is symmetric, with a peak in the middle. Exponential data are
very positively skewed, with no negative values. A triangular distribution has a limited
range and can be skewed in either direction.
Very often, uniform or triangular distributions are used in the absence of data.
These distributions depend on simple parameters that one can easily identify based
on managerial knowledge and judgment. For example, to define the uniform distribu-
tion, we need to know only the smallest and largest possible values that the variable
might assume. For the triangular distribution, we also include the most likely value.
In the construction industry, for instance, experienced supervisors can easily tell you
the fastest, most likely, and slowest times for performing a task such as framing a
house, taking into account possible weather and material delays, labor absences, and
so on.
There are two ways to define uncertain variables in Analytic Solver Platform. One is
to use the custom Excel functions for generating random samples from probability distri-
butions that we described in Table 5.1 in Chapter 5. This is similar to the method that we
used for the outsourcing example when we used the NORM.INV function in the Monte
Carlo spreadsheet simulation. For example, the Analytic Solver Platform function that
is equivalent to NORM.INV(RAND( ), mean, standard deviation) is PsiNormal(mean,
standard deviation).

Example 12.3 
Using Analytic Solver Platform Probability Distribution Functions
For the Outsourcing Decision Model, we assume that the distribution of the production volume in the outsourcing
production volume is normally distributed with a mean decision model, we could use the PsiNormal(mean, stan-
of 1,000 and a standard deviation of 100, as in the previ- dard deviation) function. Thus, we could ­enter the formula
ous example. However, we make the problem a bit more =PsiNormal(1000, 100) into cell B12. To ensure that the re-
complicated by ­assuming that the unit cost of purchas- sult is a whole number, we could modify the formula to be
ing from the supplier is also uncertain and has a triangu- =ROUND(PsiNormal(1000,100),0). To model the unit cost,
lar distribution with a m
­ inimum value of $160, most likely we could enter the formula = PsiTriangular(160, 175, 200)
value of $175, and maximum value of $200. To model the in cell B10.

The second way to define an uncertain variable is to use the Distributions button in
the Analytic Solver Platform ribbon. First, select the cell in the spreadsheet for which you
want to define a distribution. Click on the Distributions button as shown in Figure 12.3.
Choose a distribution from one of the categories in the list that pops up. This will display a
dialog in which you may define the parameters of the distribution.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 409

Figure 12.3

Analytic Solver Platform

Distributions Options

Example 12.4 
Using the Distributions Button in Analytic Solver Platform
In the Outsourcing Decision Model spreadsheet, select the Save button at the top of the dialog. Analytic Solver
cell B12, the production volume. Click the Distributions ­Platform will enter the correct Psi function into the cell in
button in the Analytic Solver Platform ribbon and select the spreadsheet and you may close the dialog. For the unit
the normal distribution from the Common category. This cost, select cell B10 and select the triangular d
­ istribution
displays the dialog shown in Figure 12.4. In the pane on from the list. Figure 12.5 shows the completed dialog after
the right, change the values of the mean and stdev under the min, likely, and max parameters have been entered. If
­Parameters to reflect the distribution you wish to model; you double-click an uncertain cell, you can bring up this
in this case, set mean to 1,000 and stdev to 100. Click dialog to perform additional editing if necessary.

Figure 12.4

Analytic Solver Platform

Normal Distribution Dialog
410 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.5

Analytic Solver Platform

Triangular Distribution Dialog

Defining Output Cells

To define a cell you wish to predict and create a distribution of output values from your
model (which Analytic Solver Platform calls an uncertain function cell), first select it,
and then click on the Results button in the Simulation Model group in the Analytic Solver
Platform ribbon. Choose the Output option and then In Cell.

Example 12.5 Using the Results Button in Analytic Solver Platform

For the Outsourcing Decision Model, select cell B19 (the + PsiOutput( ) manually to the cell formula to designate it
cost difference value) and then choose the In Cell op- as an output cell. However, you may choose only output
tion, as we described. Figure 12.6 shows the process. cells that are numerical; thus, you could not choose cell
Analytic Solver Platform modifies the formula in the cell B20, which displays a text result.
to be = B16 − B17 + PsiOutput( ). You may also add

Running a Simulation
To run a simulation, first click on the Options button in the Options group in the ­Analytic
Solver Platform ribbon. This displays a dialog (see Figure 12.7) in which you can specify
the number of trials and other options to run the simulation (make sure the ­Simulation
tab is selected). Trials per Simulation allows you to choose the number of times that
­Analytic Solver Platform will generate random values for the uncertain cells in the model
and recalculate the entire spreadsheet. Because Monte Carlo simulation is essentially sta-
tistical sampling, the larger the number of trials you use, the more precise will be the
­result. U
­ nless the model is extremely complex, a large number of trials will not unduly
tax ­today’s computers, so we recommend that you use at least 5,000 trials (the educational
version restricts this to a maximum of 10,000 trials). You should use a larger number of
trials as the number of uncertain cells in your model increases so that the simulation can
generate representative samples from all distributions for assumptions. You may run more
than one simulation if you wish to examine the variability in the results.
The procedure that Analytic Solver Platform uses generates a stream of random num-
bers from which the values of the uncertain inputs are selected from their probability
 Chapter 12  Monte Carlo Simulation and Risk Analysis 411

Figure 12.6

Analytic Solver Platform

Results Options

Figure 12.7

Analytic Solver Platform

Options Dialog

d­ istributions. Every time you run the model, you will get slightly different results because
of sampling error. However, you may control this by setting a value for Sim. Random
Seed in the dialog. If you choose a nonzero number, then the same sequence of random
numbers will be used for generating the random values for the uncertain inputs; this will
guarantee that the same values will be used each time you run the model. This is useful
when you wish to change a controllable variable in your model and compare results for the
412 Chapter 12  Monte Carlo Simulation and Risk Analysis

same assumption values. As long as you use the same number, the assumptions generated
will be the same for all simulations.
Analytic Solver Platform has alternative sampling methods; the two most common
are Monte Carlo and Latin Hypercube sampling. Monte Carlo sampling selects random
variates independently over the entire range of possible values of the distribution. With
Latin Hypercube sampling, the uncertain variable’s probability distribution is divided
into i­ntervals of equal probability and generates a value randomly within each interval.
Latin Hypercube sampling results in a more even distribution of output values because
it samples the entire range of the distribution in a more consistent manner, thus achiev-
ing more ­accurate forecast statistics (particularly the mean) for a fixed number of Monte
Carlo ­trials. However, Monte Carlo sampling is more representative of reality and should
be used if you are interested in evaluating the model performance under various what-if
scenarios. Unless you are an advanced user, we recommend leaving the other options at
their default values.
The last step is to run the simulation by clicking the Simulate button in the Solve
­Action group. When the simulation finishes, you will see a message “Simulation finished
successfully” in the lower-left corner of the Excel window.

Viewing and Analyzing Results

You may specify whether you want output charts to automatically appear after a simula-
tion is run by clicking the Options button in the Analytic Solver Platform ribbon, and
either checking or unchecking the box Show charts after simulation in the Charts tab. You
may also view the results of the simulation at any time by double-clicking on an output
cell that contains the PsiOutput() function or by choosing Simulation from the Reports
button in the Analysis group in the Analytic Solver Platform ribbon. This displays a win-
dow with various tabs showing different charts to analyze results.

Example 12.6 Analyzing Simulation Results for the Outsourcing

Decision Model
Figure 12.8 shows the Frequency tab in the simula- ­ ifference. From the chart, we see that there is about a
d
tion results window. This is a frequency distribution of 59% chance of a negative value for outsourcing, whereby
the cost difference for the 5,000 trials using the Monte in-house manufacturing would be best. The red line that
Carlo sampling method. You can see that the distribu- divides the r­ egions in the chart is called a marker line. You
tion is somewhat negatively skewed. In the Statistics can move it with your mouse to calculate different areas
pane on the right, we see that the mean cost difference of probability. As you do, the values in the Chart Statis-
is − $3,068, which suggests that, on average, it would tics section will change. You may right-click on a marker
be better to manufacture in-house than to outsource. We line to remove it; you may also add new marker lines by
also see that the minimum cost difference was − $43,222 ­right-clicking to show probabilities between marker lines
and the maximum difference was $24,367. These are es- in the chart. If you specify both a Lower Cutoff and Up-
timates of the best- and worst-case results that can be per Cutoff value, marker lines will be added at both values,
expected, lending further evidence that it might be better and the ­Likelihood statistic will be the probability between
to manufacture in-house. them. The other tabs in the results window display a cu-
In the Chart Statistics section of the Statistics mulative f­requency distribution and a reverse cumulative
pane, you may specify a Lower Cutoff, Likelihood, or frequency distribution, as well as a sensitivity chart and
­Upper ­Cutoff value. These options help you analyze the scatter plots, which we discuss in other examples. The
­frequency chart. For example, if we set the Upper ­Cutoff to best way to learn to analyze the charts is by experimenting.
0, we obtain the chart shown in Figure 12.9. This illustrates In addition, you can change the display in the right
the probability of a negative (as well as a positive) cost pane by selecting other o ­ ptions in the drop-down menu
 Chapter 12  Monte Carlo Simulation and Risk Analysis 413

Figure 12.8

Simulation Results—
Cost Difference
Frequency
Distribution

Figure 12.9

Probability of a
Negative Cost
Difference

by clicking on the down arrow to the right of the Statis- of the cumulative distribution of the output; for exam-
tics header. The options are Percentiles, Chart Type, ple, the 10th percentile in these simulation results was
Chart Options, Axis ­O ptions, and Markers. The Per- − $16,550 (not shown). This means that 10% of the simu-
centiles ­o ption displays percentiles of the simula- lated cost differences were less than or equal to − $16,550.
tion results and is ­e ssentially a numerical tabulation The other ­options are simply for customizing the charts.

In the remainder of this chapter, we present several additional examples of Monte

Carlo simulation using Analytic Solver Platform. These serve to illustrate the wide range
of applications in which the approach may be used and also various features of Analytic
Solver Platform and tools for analyzing simulation models.
414 Chapter 12  Monte Carlo Simulation and Risk Analysis

New-Product Development Model

The Moore Pharmaceuticals spreadsheet model to support a new-product development deci-

sion was introduced in Chapter 11; Figure 12.10 shows the model again. Although the values
used in the spreadsheet suggest that the new drug would become profitable by the fourth year,
much of the data in this model are uncertain. Thus, we might be interested in evaluating the
risk associated with the project. Three questions we might be interested in are as follows:

1.
What is the risk that the net present value over the 5 years will not be positive?
2.
What are the chances that the product will show a cumulative net profit in the
third year?
3.
What cumulative profit in the fifth year are we likely to realize with a prob-
ability of at least 0.90?

Suppose that the project manager of Moore Pharmaceuticals has identified the fol-
lowing uncertain variables in the model and the distributions and parameters that describe
them, as follows:

• Market size: normal with mean of 2,000,000 units and standard deviation of
400,000 units
• R&D costs: uniform between $600,000,000 and $800,000,000
• $30,000,000
Clinical trial costs: lognormal with mean of $150,000,000 and standard deviation

• Annual market growth factor: triangular with minimum = 2%, maximum = 6%,
and most likely = 3%
• Annual market share growth rate: triangular with minimum = 15%,
maximum = 25%, and most likely = 20%

Figure 12.10

Moore Pharmaceuticals
Spreadsheet Model
 Chapter 12  Monte Carlo Simulation and Risk Analysis 415

Example 12.7 Setting Up the Simulation Model for Moore Pharmaceuticals

As we learned earlier, we may use either the Psi func-
tions or the Distribution buttons in the Analytic Solver
• Annual market growth factor (cells C18 to F18):
=PsiTriangular(2%, 3%, 6%)
Platform ­r ibbon to specify the uncertain variables. Al-
though the result is the same, the Psi functions are often
• Annual market share growth rate (cells C20 to
F20): =PsiTriangular(15%, 20%, 25%)
easier to use. To model the market size, we could use the
Because the annual market-growth factors and
PsiNormal(mean, standard deviation) function. Thus, we
­market-share-growth rates use the same distributions,
could enter the formula = PsiNormal(2000000, 400000)
we need enter them only once and then copy them to the
into cell B5. Similarly, we could use the following func-
other cells.
tions for the remaining uncertain variables:
We define the cumulative net profit for each year
• R&D Costs (cell B11): =PsiUniform(600000000,
800000000)
(cells B28 through F28) and the net present value (cell
B30) as the output cells.
• Clinical trial costs (cell B12):
=PsiLognormal(150000000, 30000000)

Now we are prepared to run the simulation and analyze the results. If your simulation
model contains more than one output function, then a Variables Chart containing fre-
quency graphs of up to 9 output functions and uncertain variables will appear as shown in
Figure 12.11. In this case, the Variables Chart shows the frequency charts for all 6 uncer-
tain functions (cells B28:F28 and B30) and 3 of the uncertain inputs (B5, B11, and B12) in
the Moore Pharmaceutical model. You may customize this by checking or unchecking the
boxes in the Filters pane; for example, you can remove the uncertain input distributions
and only show the six outputs. As noted earlier in this chapter, you may also suppress the
automatic display of the chart in the Charts tab after clicking the Options button.
In this example, we used 10,000 trials. We may use the frequency charts in the simu-
lation results to answer the risk analysis questions we posed earlier.

Figure 12.11

Variables Chart for Simulation

Results
416 Chapter 12  Monte Carlo Simulation and Risk Analysis

Example 12.8  Risk Analysis for Moore Pharmaceuticals

1. 
What is the probability that the net present value 3. What cumulative profit in the fifth year are we
over the 5 years will not be positive? Double-click likely to realize with a probability of at least 0.90?
on cell B30 to display the simulation results for the An easy way to answer this question is to view the
net present value output. Enter the number 0 for the ­Percentiles results (see Figure 12.14). Therefore, we
Upper Cutoff value in the Statistics pane. The re- can expect a cumulative net profit of about $180,000
sults are shown in Figure 12.12; this shows about an or more with 90% certainty. Another way is to set
18% chance that the NPV will not be positive. the lower cutoff in the Chart Statistics field to some
number smaller than the minimum value and then
2. What are the chances that the product will show a cu-
set the likelihood to 10%. Analytic Solver Platform
mulative net profit in the third year? Double-click cell
will calculate and draw a marker line for the value of
D28, the cumulative net profit in year 3. Enter the value
the upper cutoff that provides a certainty less than
0 for the Lower Cutoff value, as illustrated in Figure
the upper cutoff of 10% and, consequently, a cer-
12.13. This shows that the probability of a positive cu-
tainty of 90% greater than the upper cutoff.
mulative net profit in the third year is only about 9%.

Figure 12.12

Probability of a Nonpositive
Net Present Value

Figure 12.13

Probability of a Non-Positive
Cumulative Third-Year Net
Profit
 Chapter 12  Monte Carlo Simulation and Risk Analysis 417

Figure 12.14

Percentiles for Fifth-Year

Cumulative Net Profit

Confidence Interval for the Mean

Monte Carlo simulation is essentially a sampling experiment. Each time you run a simu-
lation, you will obtain slightly different results as we observed in Example 12.2 for the
outsourcing decision model. Therefore, statistics such as the mean are a single observation
from a sample of n trials from some unknown population. In Chapter 6, we discussed how
to construct a confidence interval for the population mean to measure the error in estimat-
ing the true population mean. We may use the statistical information to construct a confi-
dence interval for the mean using a variant of formula (6.3) in Chapter 6:

x { za>21s> 1n2 (12.1)

Because a Monte Carlo simulation will generally have a very large number of trials (we
used 10,000), we may use the standard normal z-value instead of the t-distribution in the
confidence interval formula.

Example 12.9 A Confidence Interval for the Mean Net Present Value
We will construct a 95% confidence interval for the mean NPV This means that if we ran the simulation again with dif-
using the simulation results from the Moore Pharmaceuticals ferent random inputs, we could expect the mean NPV to
example. From statistics shown in Figure 12.12, we have generally fall within this interval. To reduce the size of the
confidence interval, we would need to run the simulation
mean = $200,608,120
for a larger number of trials. For most risk analysis appli-
standard deviation = $220,980,564
cations, however, the mean is less important than the ac-
n = 10,000
tual distribution of outcomes.
For a 95% confidence interval, zA>2 = 1.96. Therefore,
­using formula (12.1), a 95% confidence interval for the
mean would be

$200,608,120 ± 1.961 220,980,564 , !10,000 2,

or [ $196,276,901, $204,939,339]
418 Chapter 12  Monte Carlo Simulation and Risk Analysis

Sensitivity Chart
The sensitivity chart feature allows you to determine the influence that each uncertain
model input has individually on an output variable based on its correlation with the output
variable. The sensitivity chart displays the rankings of each uncertain variable according to
its impact on an output cell as a tornado chart. A sensitivity chart provides three benefits:

1.
It tells which uncertain variables influence output variables the most and
which would benefit from better estimates.
2.
It tells which uncertain variables influence output variables the least and can
be ignored or discarded altogether.
3.
By providing understanding of how the uncertain variables affect your model,
it allows you to develop more realistic spreadsheet models and improve the
accuracy of your results.

The sensitivity chart can be viewed by clicking the Sensitivity tab in the results window
(see Figure 12.15).

Example 12.10  Interpreting the Sensitivity Chart for NPV

Figure 12.15 shows the sensitivity chart and the net pres- uncertain variable cells have a negligible effect. This means
ent value output cell (B30). The uncertain variable cells are that if you want to reduce the variability in the distribution
ranked from top to bottom, beginning with the one having of NPV the most, you would need to obtain better informa-
the highest absolute value of correlation with NPV. In this tion about the estimated market size and use a probability
example, we see that cell B5, the market size, has a cor- distribution that has a smaller variance. The small correla-
relation of about 0.95 with NPV; the R&D cost (cell B11) has tions between NPV and the market-growth factors suggest
a negative 0.255 correlation, and the clinical trial cost (cell that using constant values instead of uncertain probability
B12) has a negative 0.130 correlation with NPV. The other distributions would have little effect on the results.

Overlay Charts
If a simulation has multiple related forecasts, the overlay chart feature allows you to
superimpose the frequency distributions from selected forecasts on one chart to compare
differences and similarities that might not be apparent.

Figure 12.15

Sensitivity Chart for Net

Present Value
 Chapter 12  Monte Carlo Simulation and Risk Analysis 419

Example 12.11  Creating an Overlay Chart

To create an overlay chart, click the Charts button in the which correspond to the cumulative net profit for years
Analysis group in the Analytic Solver Platform ­r ibbon. 1 and 5. Figure 12.17 shows the overlay chart for the
Click Multiple Simulation Results (do not choose ­Multiple distributions of cumulative net profit for years 1 and 5.
Simulations!) and then choose Overlay. In the Reports This chart makes it clear that the mean value for year 1
dialog that appears, select the output variable cells you is smaller than for year 5, and the variance in year 5 is
wish to include in the chart and move them to the right much larger than that in year 1. This is to be expected
side of the dialog using the arrow buttons (see Figure because there is more uncertainty in predicting farther in
12.16). In this example, we selected cells B28 and F28, the future, and the model captures this.

Figure 12.16

Reports Dialog for Selecting

Output Cells for an Overlay
Chart

Figure 12.17

Overlay Chart for Year 1 and

Year 5 Cumulative Net Profit
420 Chapter 12  Monte Carlo Simulation and Risk Analysis

Trend Charts
If a simulation has multiple output variables that are related to one another (such as over
time), you can view the distributions of all output variables on a single chart, called a
trend chart. In Analytic Solver Platform, the trend chart shows the mean values as well
as 75% and 90% bands (probability intervals) around the mean. For example, the band
representing the 90% band range shows the range of values into which the output variable
has a 90% chance of falling.

Example 12.12  Creating a Trend Chart

To create a trend chart for the Moore Pharmaceuticals chart and move them to the right side of the dialog using
example, click the Charts button in the Analysis group the arrow buttons. In this example, we selected cells B28
in the Analytic Solver Platform ribbon. Click Multiple through F28, which correspond to the cumulative net
­S imulation Results and then choose Trend. (Be care- profit for all years. Figure 12.18 shows a trend chart for
ful not to confuse “Multiple Simulation Results” with these variables. We see that although the mean net cumu-
“­M ultiple Simulations” in the drop-down menu; these lative profit increases over time, so does the variation, in-
are different options.) In the Reports dialog that appears, dicating that the uncertainty in forecasting the future also
­select the output variable cells you wish to include in the increases with time.

Box-Whisker Charts
Finally, Analytic Solver Platform can create box-whisker charts to illustrate the statistical prop-
erties of the output variable distributions in an alternate fashion. A box-whisker chart shows
the minimum, first quartile, median, third quartile, and maximum values in a data set graphi-
cally. The first and third quartiles form a box around the median, showing the middle 50% of
the data, and the whiskers extend to the minimum and maximum values. They can be created
by clicking on the Charts button similar to the overlay and trend charts. Figure 12.19 shows an
example for the cumulative net profits in the Moore Pharmaceuticals simulation.

Figure 12.18

Trend Chart for Cumulative

Net Profit Over 5 Years
 Chapter 12  Monte Carlo Simulation and Risk Analysis 421

Figure 12.19

Example of Analytic Solver

Platform Box-Whisker Chart

Simulation Reports
Analytic Solver Platform allows you to create reports in the form of Excel worksheets that
summarize a simulation. To do this, click the Reports button in the Analysis group in the
Analytic Solver Platform ribbon, and choose Simulation from the options that appear. The
report summarizes basic statistical information about the model, simulation options, un-
certain variables, and output variables, most of which we have already seen in the charts.
It is useful to provide a record of the simulation for quick reference.

Newsvendor Model

In Chapter 11, we developed the newsvendor model to analyze a single-period purchase

decision. Here we apply Monte Carlo simulation to forecast the profitability of different
purchase quantities when the future demand is uncertain.
Let us suppose that the store owner kept records for the past 20 years on the number
of boxes sold at full price, as shown in the spreadsheet in Figure 12.20 (Excel file News-
vendor Model with Historical Data). The distribution of sales seems to be some type of
positively skewed unimodal distribution.

The Flaw of Averages

You might wonder why we cannot simply use average values for the uncertain inputs in
a decision model and eliminate the need for Monte Carlo simulation. Let’s see what hap-
pens if we do this for the newsvendor model.

Example 12.13 Using Average Values in the Newsvendor Model

If we find the average of the historical candy sales, we construct a data table to evaluate the profit for each of the
obtain 44.05, or, rounded to a whole number, 44. Using historical values (also shown in ­Figure 12.21), we see that
this value for demand and purchase quantity, the model the average profit is only $255.00.
predicts a profit of $264 (see ­Figure 12.21). However, if we
422 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.20

Newsvendor Model with

Historical Data

Figure 12.21

Example of the Flaw of

Averages

Dr. Sam Savage, a strong proponent of spreadsheet modeling, coined the term the
flaw of averages to describe this phenomenon. Basically what this says is that the evalu-
ation of a model output using the average value of the input is not necessarily equal to the
average value of the outputs when evaluated with each of the input values. The reason this
occurs in the newsvendor example is because the quantity sold is limited to the smaller of
the demand and purchase quantity, so even when demand exceeds the purchase quantity,
the profit is limited. Using averages in models can conceal risk, and this is a common ­error
among users of analytic models. This is why Monte Carlo simulation is valuable.

Monte Carlo Simulation Using Historical Data

We can perform a Monte Carlo simulation by resampling from the historical sales
­distribution—that is, by selecting a value randomly from the historical data as the demand
in the model.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 423

Example 12.14 Simulating the Newsvendor Model Using Resampling

In the Newsvendor Model with Historical Data spread- profit cell B17 as an uncertain function cell, set the simula-
sheet, we have the historical data listed in the range tion options (we chose 5,000 trials), and run the simulation.
D2:D21. All we need to do is to define the distribution of Figure 12.22 shows the results; for the purchase quantity
demand in cell B11 using the PsiDisUniform function in of 44, the mean profit is $255.00. The frequency chart,
Analytic Solver Platform. This function will sample a value also shown in Figure 12.22, looks somewhat odd. How-
from the historical data for each trial of the simulation. ever, recall that if demand exceeds the purchase quantity,
Enter the formula = PsiDisUniform(D2:D21) into cell B11. then sales are limited to the number purchased, which
Now, you may set up the simulation model by defining the ­explains the large spike at the right of the distribution.

Monte Carlo Simulation Using a Fitted Distribution

While sampling from empirical data is easy to do, it does have some drawbacks. First, the
empirical data may not adequately represent the true underlying population because of
sampling error. Second, using an empirical distribution precludes sampling values outside
the range of the actual data. Therefore, it is usually advisable to fit a distribution and use
it for the uncertain variable. We can do this by fitting a distribution to the data using the
techniques we described in Chapter 5.

Example 12.15 Using a Fitted Distribution for Monte Carlo Simulation

Following the steps in Example 5.42, first highlight the if you wish to accept the fitted distribution. Click Yes, and a
range of the data in the Newsvendor Model with Historical pop-up will allow you to drag and place the function into a
Data spreadsheet, and click Fit from the Tools group in the cell in the spreadsheet. Place the Psi function for the nega-
­Analytic Solver Platform ribbon. Because the number of tive binomial distribution in the first cell of the data (cell D2).
sales is discrete, select the Discrete radio button in the Fit To use this for the simulation, simply reference cell D2 in cell
Options dialog and click Fit. Figure 12.23 shows the best-­ B11, corresponding to the demand in the model. Figure 12.24
fitting distribution, a negative ­binomial distribution. When you shows the results, which are quite similar to the results found
attempt to close the dialog, Analytic Solver Platform will ask by resampling in Example 12.14.

Figure 12.22

Newsvendor Model
Simulation Results Using
Resampling for Purchase
Quantity = 44
424 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.23

Best-Fitting Distribution for

Historical Candy Sales

Figure 12.24

Newsvendor Simulation
Results Using the Negative
Binomial Distribution for
Purchase Quantity = 44

Analytic Solver Platform has a feature called Interactive Simulation. Whenever the
Simulate button is clicked, you will notice that the lightbulb in the icon turns bright. If
you change any number in the model, Analytic Solver Platform will automatically run the
simulation for that quantity; this makes it easy to conduct what-if analyses. For example,
changing the purchase quantity to 50 yields the results shown in Figure 12.25. The mean
profit drops to $246.05. You could use this approach to identify the best purchase quantity;
however, a more systematic method is described in the online Supplementary Chapter B.

Overbooking Model

In Chapter 11, we developed a model for overbooking decisions (Hotel Overbooking

Model). In any realistic overbooking situation, the actual customer demand as well as the
number of cancellations would be random variables. We illustrate how a simulation model
can help in making the best overbooking decision and introduce a new type of distribution
in Analytic Solver Platform, a custom distribution.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 425

Figure 12.25

Newsvendor Simulation
Results for Purchase
Quantity = 50

Figure 12.26

Hotel Overbooking
Simulation Model and
Demand Distribution

The Custom Distribution in Analytic Solver Platform

Let us assume that historical data for the demand have been collected and summarized in
a relative frequency distribution, but that the actual data are no longer available. These are
shown in columns D and E in Figure 12.26 (Excel file Hotel Overbooking Monte Carlo
Simulation Model with Custom Demand). We also assume that each reservation has a
­constant probability p = 0.04 of being canceled; therefore, the number of cancellations
(cell B14) can be modeled using a binomial distribution with n = number of reservations
made and p = probability of cancellation.

Example 12.16 Defining a Custom Distribution in Analytic Solver Platform

To use the relative frequency distribution to define the uncer- corresponds to the demand, then click on the Distributions
tain demand in the Hotel Overbooking Model with Custom button in the Analytic Solver Platform ribbon and choose
Demand (note that this spreadsheet is already completed; Discrete from the Custom category. In the dialog, edit the
to follow along, copy columns D and E to the original ­Hotel range for “values” and “weights” in the Parameters section
Overbooking Model worksheet) first select cell B12 that in the fields on the right. Values correspond to the range
(continued )
426 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.27

Custom Discrete Distribution

Dialog

Figure 12.28

Binomial Distribution Dialog

of demand in cells D2:D13, and weights are the relative trials must be the value in cell B13. This is critical in this
frequencies or probabilities in cells E2:E13. The dialog will example, because the number of reservations made will
then display the actual form of the distribution, as shown in change, depending on the customer demand in cell B12.
Figure 12.27. Alternatively, you could use the function =Psi Therefore, in the Parameters section of the dialog, we
Discrete($D$2:$D$13,$E$2:$E$13) in cell B12. must reference cell B13 and not use a constant value,
To model the number of cancellations in cell B14, as shown in F­ igure 12.28. Alternatively, we could use the
choose the binomial distribution from the Discrete cat- function = PsiBinomial(B13, 0.04) in cell B14. Define cells
egory in the Distributions list. Note that the number of B17 and B18 as output cells and run the model.

Figures 12.29 and 12.30 show frequency charts of the two output variables—number
of overbooked customers and net revenue—for accepting 310 reservations. There is about
a 14% chance of overbooking at least one customer. Observe that there seem to be two
different distributions superimposed over one another in the net revenue frequency distri-
bution. Can you explain why this is so? As with the newsvendor problem, we can easily
change the number of reservations made, and the Interactive Simulation capability will
quickly run a new simulation and change the results in the frequency charts.

Cash Budget Model

Cash budgeting is the process of projecting and summarizing a company’s cash inflows
and outflows expected during a planning horizon, usually 6 to 12 months.3 The cash bud-
get also shows the monthly cash balances and any short-term borrowing used to cover

3Douglas R. Emery, John D. Finnerty, and John D. Stowe, Principles of Financial Management (Upper
Saddle River, NJ: Prentice Hall, 1998): 652–654.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 427

Figure 12.29

Frequency Chart of Number

of Overbooked Customers

Figure 12.30

Frequency Chart of Net

Revenue

cash ­shortfalls. Positive cash flows can increase cash, reduce outstanding loans, or be used
elsewhere in the business; negative cash flows can reduce cash available or be offset with
additional borrowing. Most cash budgets are based on sales forecasts. With the inherent
uncertainty in such forecasts, Monte Carlo simulation is an appropriate tool to analyze
cash budgets.
Figure 12.31 shows an example of a cash budget spreadsheet (Excel file Cash
Budget Model). The highlighted cells represent the uncertain variables and outputs
we want to predict from the simulation model. The budget begins in April (thus, sales
for April and subsequent months are uncertain). These are assumed to be normally
­distributed with a standard deviation of 10% of the mean. In addition, we assume that
sales in adjacent months are correlated with one another, with a correlation coefficient
of 0.6. On average, 20% of sales are collected in the month of sale, 50%, in the month
following the sale, and 30%, in the second month following the sale. However, these
figures are uncertain, so a uniform distribution is used to model the first two values
(15% to 20% and 40% to 50%, respectively), with the assumption that all remaining
revenues are collected in the second month following the sale. Purchases are 60% of
428 Chapter 12  Monte Carlo Simulation and Risk Analysis

Figure 12.31
sales and are paid for 1 month prior to the sale. Wages and salaries are 12% of sales
Cash Budget Model and are paid in the same month as the sale. Rent of $10,000 is paid each month. Addi-
tional cash operating expenses of $30,000 per month will be incurred for April through
July, decreasing to $25,000 for August and September. Tax payments of $20,000 and
$30,000 are expected in April and July, respectively. A capital expenditure of $150,000
will occur in June, and the company has a mortgage payment of $60,000 in May. The
cash balance at the end of March is $150,000, and managers want to maintain a mini-
mum balance of $100,000 at all times. The company will borrow the amounts necessary
to ensure that the minimum balance is achieved. Any cash above the minimum will be
used to pay off any loan balance until it is eliminated. The available cash balances in
row 25 of the spreadsheet are the output variables we wish to predict.

Example 12.17  Simulating the Cash Budget Model without Correlations

Build the basic simulation model by defining distributions the uniform distribution = PsiUniform(15%, 20%), and
for each of the uncertain variables. First, specify the sales for the previous month collections rate in cell B8, use
for April through October (cells E5:K5) to be normally =PsiUniform(40%, 50%). Define the available balances in
­distributed with means equal to the values in the spread- row 25 as output variables in the simulation model. The
sheet and standard deviations equal to 10% of the means. Excel file Cash Budget Monte Carlo Simulation Model
For example, use the function =PsiNormal(600000,60000) provides the completed simulation model.
in cell E5. For the current collections rate in cell B7, use

Figure 12.32 shows the results of Example 12.17 in the form of a trend chart. We see
that there is a high likelihood that the cash balances for the first 3 months will be negative
before increasing. Viewing the frequency charts and statistics for the individual months
will provide the details of the distributions of likely cash balances and the probabilities
 Chapter 12  Monte Carlo Simulation and Risk Analysis 429

Figure 12.32

Cash Balance Simulation

Trend Chart

of requiring loans. For example, in April, the probability that the balance will not exceed
the minimum of $100,000 and require an additional loan is about 0.70 (see Figure 12.33).
This actually worsens in May and June and becomes zero by July.

Correlating Uncertain Variables

Unless you specify otherwise, Monte Carlo simulation assumes that each of the uncertain
variables is independent of all the others. This may not be the case. In the cash budget
model, if the sales in April are high, then it would make sense that the sales in May
would be high also. Thus, we might expect a positive correlation between these variables.
In this scenario, we assume a correlation coefficient of 0.6 between sales in successive
months. The following example shows how to incorporate this assumption into the simu-
lation model.

Figure 12.33

Likelihood of Not Meeting

Minimum Balance in April
430 Chapter 12  Monte Carlo Simulation and Risk Analysis

Example 12.18 Incorporating Correlations in Analytic Solver Platform

To correlate the uncertain variables in the Cash Budget does not satisfy this property, it will ask you if you want to
Monte Carlo Simulation Model, first click the ­Correlations adjust the correlations so that it does. Always choose Yes.
button in the Simulation Model group in the Analytic Solver Click the Update Matrix button (you can make changes
Platform ribbon. This brings up the Create new correlation manually, but we recommend this only for advanced users)
matrix dialog shown in Figure 12.34 that lists the uncertain and then Accept Update. The adjusted matrix is shown in
variables in the model. In this example, we are only correlat- Figure 12.37. Note that the correlations between successive
ing the variables in the range E5:K5. In the left pane, hold the months are close to 0.6, but that the matrix now includes
Ctrl key and click on each of the distributions in the range some small correlations between other months. This en-
E5:K5, or click on $E5$, hold the Shift key and then click sures the mathematical consistency needed to run the sim-
on $K$5 to select them. Then click on the right arrow. (The ulation. You may now close the dialog.
double right arrow selects all of them, which we do not want The cell range of the correlation matrix is used
in this example.) This creates an initial correlation matrix in the function PsiCorrMatrix(cell range, position, in-
as shown in Figure 12.35. The numerical values show the stance), where position corresponds to the number
correlations (initially set to zero); the green distributions are of the uncertain variables in the correlation matrix
those used in the uncertain cells, and the blue scatterplots and instance refers to the name given to the correla-
show visual representations of the correlations between the tion matrix. Analytic Solver Platform adds these func-
variables. Replace the zeros by the correlations you want tions to the distributions for the uncertain variables
in the model. In this example, we will assume a 0.6 correla- that are correlated. For example, the ­formula in
tion between each successive month. In boxes 2 and 3, you cell E5 for April sales is changed to: = PsiNormal
can name the correlation matrix and specify the location to (600000,60000,PsiCorrMatrix($B$33:$H$39,1, “Monthly
place it in the spreadsheet. This is shown in Figure 12.36. Correlations”)). The ­f ormula in cell F5 for May sales is
Now, it is very important to ensure that the correlations changed to: = PsiNormal(700000,70000,PsiCorrMatrix
are mathematically consistent with each other (a mathemat- ($B$33:$H$39,2, “Monthly Correlations”)), and so on.
ical property called positive semidefinite). You can select Now set the simulation options and run the model. The
the Validate button in the Manage Correlations dialog, or Excel file Cash Budget Monte Carlo Simulation Model
Analytic Solver Platform will perform an ­automatic check for with Correlations provides the completed model for this
this when you try to close the dialog. If the correlation ­matrix example.

Figure 12.34

Create New Correlation

Matrix Dialog
 Chapter 12  Monte Carlo Simulation and Risk Analysis 431

Figure 12.35

Initial Correlation Matrix

Figure 12.36

Completed Correlation
Matrix

Figure 12.37

Adjusted Correlations
432 Chapter 12  Monte Carlo Simulation and Risk Analysis

You will observe some slight differences in the results when uncertain variables
are correlated. For example, the standard deviation for the September balance is lower
when correlations are included in the model than when they are not. Generally, induc-
ing correlations into a simulation model tends to reduce the variance of the predicted
outputs.

Analytics in Practice: I mplementing Large-Scale Monte Carlo

Spreadsheet Models4
Implementing large-scale Monte Carlo models in because the entire spreadsheet must be recalculated
spreadsheets in practice can be challenging. This ex- both for each iteration of the simulation and each in-
ample shows how one company used Monte Carlo dividual asset (or transaction) within the portfolio. This
simulation for commercial real estate credit-risk pushes the limits of stand-alone Excel models, even
analysis but had to develop new approaches to effec- for a single asset. Moreover, because the bank is usu-
tively implementing spreadsheet analytics across the ally interested in analyzing its entire portfolio of thou-
company. sands of assets, in practice, it becomes impossible to
Based in Stuttgart, Germany, Hypo Real Estate do so using stand-alone Excel.
Bank International (Hypo), with a large portfolio in Therefore, Hypo needed a way to implement the
commercial real estate lending, undertakes some of complex analytics of simulation in a way that its global
the world’s largest real estate transactions. Hypo was offices could use on all their thousands of loans. In
faced with the challenge of complying with Basel II addition to the computational intensity of simulation
banking regulations in Europe. Basel II was a new reg- analytics, the option to build the entire simulation
ulation for setting the minimum capital to be held in framework in Excel can lead to human error, which
reserve by internationally active banks. If a bank is
able to comply with the more demanding require-
ments of the regulation, it can potentially save
E20–E60 million per year in capital costs. To qualify
however, Hypo needed new risk models and report-
ing systems. The company also wished to upgrade
its internal reporting and management framework to
provide better analytical tools to its lending officers,
who were responsible for structuring new loans, and
to provide its managers with better insights into the
risks of the overall portfolio.
Monte Carlo simulation is the only practical ap-
proach for analyzing risk models the bank needed. For
example, in one commercial real estate application,
200 different macroeconomic and market variables are
Vladitto/Shutterstock.com

typically simulated over 20 years. The cash-flow mod-

eling process can be even more complex, particularly if
the effects of all the intricate details of the transaction
must be quantified. However, the computational pro-
cess of Monte Carlo simulation is numerically intensive

4Based on Yusuf Jafry, Christopher Marrison, and Ulrike Umkehrer-Neudeck, “Hypo International
Strengthens Risk Management with a Large-Scale, Secure Spreadsheet-Management Framework,” Interfaces,
38, 4 (July–August 2008): 281–288.
 Chapter 12  Monte Carlo Simulation and Risk Analysis 433

they called spreadsheet risk. Spreadsheet risks that Given these potential problems, Hypo deemed
Hypo wished to minimize included the following: a pure Excel solution as i­mpractical. Instead, they
used a consulting firm’s proprietary software, called
• Proliferation of spreadsheet models that are
stored on individual users’ desktop computers
the Specialized Finance System (SFS), that embeds
spreadsheets within a ­high-performance, server-based
throughout the organization are untested and lack system for enterprise applications. This eliminated the
version data, and the unsanctioned manipulation spreadsheet risks but allowed users to exploit the flex-
of the results of spreadsheet calculations. ible programming power that spreadsheets provide,
• Potential for serious mistakes resulting from typo-
graphical and “cut and copy-and-paste” errors when
while giving confidence and trust in the results. The new
system has improved management reporting and the
entering data from other applications or spreadsheets. efficiency of internal processes and has also provided
• Accidental acceptance of results from incomplete
calculations.
insights into structuring new loans to make them less
risky and more profitable.
• Errors associated with running an insufficient
number of Monte Carlo iterations because of data
or time constraints.

Key Terms
Box-whisker chart Risk
Flaw of averages Risk analysis
Marker line Sensitivity chart
Monte Carlo simulation Trend chart
Overlay chart Uncertain function

­ ­