UV1461

ANALYZING UNCERTAINTY:
PROBABILITY DISTRIBUTIONS AND SIMULATION

In making important business decisions, executives have to combine experience and

instinct with a careful consideration of the consequences of any action they choose. There are
many different categories of important consequences to consider, and the relative importance of
the different categories may vary from decision to decision. The category of economic
consequences, however, is almost invariably of interest: What are the implications of a course of
action for a company’s bottom line? Analyzing these economic consequences is not only a useful
check on intuition, but often provides valuable insights that lead to better, more informed
decisions.

Identifying the economic consequences of a course of action is almost always made

difficult by the presence of uncertainty in how things will play out. Good economic analysis
must somehow deal with this uncertainty. The simplest approach is to just ignore it: Make
assumptions (best guesses) about the outcome of important uncertainties and analyze possible
courses of action on this basis. While simple, this approach causes the uncertainty (and its
resulting risk) to remain largely hidden from view; it is incorporated into the decision-making
process amorphously via intuition and gut feel. By failing to consider the uncertainty explicitly,
we run the risk of (1) improperly valuing the economic consequences of a course of action and
(2) exposing ourselves to unacceptable risk (i.e., if we were aware of its actual magnitude and
likelihood, we would choose a different course of action). We may also miss opportunities to
find better solutions that add value by proactively managing the risk.

A second approach is to consider the endpoints of the ranges of outcomes for individual
uncertainties (best case, worst case). This approach enables us to begin to see the impact of
uncertainty on the consequences of our actions by generating overall best and worst cases. Often,
our decision is “made” by considering the worst case alone: If it is sufficiently bad, we abandon
the alternative. While this approach is appealing in its simplicity and better than simply taking
our best guess of important uncertainties, it suffers from its incompleteness. While we know the
worst-case consequence, how likely is it? How likely is the best case? And what about the
myriad of intermediate possible consequences, all generally more likely than the best or worst
cases?

This technical note was prepared by Robert L. Carraway. It was written as a basis for class discussion rather than to
illustrate effective or ineffective handling of an administrative situation. Copyright  2005 by the University of
Virginia Darden School Foundation, Charlottesville, VA. All rights reserved. To order copies, send an e-mail to
[email protected]. No part of this publication may be reproduced, stored in a retrieval system, used in a
spreadsheet, or transmitted in any form or by any means—electronic, mechanical, photocopying, recording, or
otherwise—without the permission of the Darden School Foundation.

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -2- UV1461

An alternative to either of these approaches is to use a probability distribution to provide

a complete forecast of each of the major uncertainties faced in a decision-making context. These
distributions provide a complete description of our beliefs about the uncertainty, including both
central tendency (the most likely outcome) and variability (how much the actual outcome might
differ from the most likely). We can then use a tool called simulation to analyze the impact of
these characterizations of uncertainty on the economic consequence of any course of action. The
result of “running” a simulation is another probability distribution, this one characterizing all the
possible economic consequences of a particular course of action and their relative likelihoods.

In the rest of this note, we take a simple example and demonstrate how the simulation
approach might work.

Example

Consider a company that is trying to decide whether or not to continue a product line for
the next quarter. Continuing the line would cost the company $500,000 in fixed costs (i.e., by
dropping the line, the company would save $500,000 over current costs, not including the
variable costs associated with making each individual unit of product). Each unit of product sells
for $10 and costs $5 to make (i.e., once production is up and running, each additional unit costs
an additional $5 to actually produce).

The company faces two major uncertainties involving the likely sales of the product in
the upcoming quarter. First, a current client is considering making a major purchase of the
product. Specifically, the client is considering the purchase of 50,000 units. Because of the size
of the order, the company will offer to sell the product to the major client at a discount,
specifically $9 per unit. The client wants the same vendor for the entire purchase, but there are
competitors able and willing to provide the same product. Because of this competition, it is
unclear whether the company will get the contract (assuming of course it continues the product
line). Unfortunately, the company will have to make the decision of whether or not to continue
the product line before knowing whether or not it will receive the major contract. The company
believes that its chances of winning the contract, given that the client is a current customer,
however, are better than fifty-fifty.

Second, there is uncertainty as to the general level of sales apart from the single major
purchase. Since the company has been selling the product for years, there are ample data
available on sales in previous quarters. Based on these data and an analysis of the market, the
company’s best guess is that it will sell 75,000 units if it chooses to continue the product, not
including the major purchase.

We can create a simple spreadsheet to calculate the consequence of continuing the

product line. Assuming we get the major contract (our best guess) and that sales are at our best-
guess level, we get the following economic consequence of continuing the product line for the
upcoming quarter.

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -3- UV1461

Revenue per unit, general $ 10

Revenue per unit, major purchase $ 9
Variable cost per unit $ 5
Fixed costs for the quarter $500,000
General sales in units 75,000
Major purchase sales, in units 50,000

Revenue
General $ 750,000
Major purchase $ 450,000
Total $1,200,000
Costs
Fixed $ 500,000
Variable $ 625,000
Total $1,125,000
Profit $ 75,000

If we fail to land the major purchase, however, we get the following:

Revenue per unit, general $ 10

Revenue per unit, major purchase $ 9
Variable cost per unit $ 5
Fixed costs for the quarter $500,000
General sales in units 75,000
Major purchase sales, in units -

Revenue
General $750,000
Major purchase $0
Total $750,000
Costs
Fixed $500,000
Variable $375,000
Total $875,000
Profit ($125,000)

Clearly, whether or not we get the major purchase is a big deal in terms of whether or not
we should continue the product line. But remember: The 75,000-unit general sales level is our
best guess, and it could also be “off.” Suppose, based on historical data, experience, and a careful
market analysis, we believe that the general sales level could be as few as 50,000 units or as
many as 110,000 units. If we get the major contract but only sell 50,000 units on the general
market, we get the following consequence:

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -4- UV1461

Revenue per unit, general $ 10

Revenue per unit, major purchase $ 9
Variable cost per unit $ 5
Fixed costs for the quarter $500,000
General sales in units 50,000
Major purchase sales, in units 50,000

Revenue
General $500,000
Major purchase $450,000
Total $950,000
Costs
Fixed $500,000
Variable $500,000
Total $1,000,000
Profit ($50,000)

Even with the major contract, we lose money! Similarly, even if we lose the major
contract but sell 110,000 units, we get the consequence:

Revenue per unit, general $ 10

Revenue per unit, major purchase $ 9
Variable cost per unit $ 5
Fixed costs for the quarter $500,000
General sales in units 110,000
Major purchase sales, in units -

Revenue
General $1,100,000
Major purchase $0
Total $1,100,000
Costs
Fixed $ 500,000
Variable $ 550,000
Total $1,050,000
Profit $ 50,000

We lose the contract, but we make money!

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -5- UV1461

We have now identified four possible consequences of continuing the product line:

Major Purchase Sales General Sales Profit

0 75,000 ($125,000)
50,000 50,000 ($ 50,000)
0 110,000 $ 50,000
50,000 75,000 $ 75,000

We can represent the profit of these four possible outcomes graphically as:

And we have not even considered the best case (getting the major contract, general sales
of 110,000 units) or worst case (no major contract, 50,000 units of general sales) consequences.
Clearly, both uncertainties—whether or not we get the major contract and the general sales
level—matter, since we get positive and negative consequences if either varies from our best
guess. But we have no idea of the relative likelihood of even the four consequences we have
identified (beyond that our best-guess scenario—with a profit of $75,000—is probably more
likely than the others), much less the ones we have not even considered yet (including the best
and worst cases).

Rather than continue generating individual scenarios without regard for their likelihood
(which will just lead to more points on the line but precious little additional insight), we can opt
for the third approach outlined in the introduction to this note. Specifically, we can choose to
express the uncertainty in both whether or not we get the major contract and the general sales
level as probability distributions, with the goal of using simulation to fill out the line above with
both all possible consequences and the relative likelihood of each. The probability distributions
we will need are of two fundamentally different types, distinguished by the number of possible
outcomes of each uncertainty.

Getting the Major Contract: Discrete Probability Distribution

There are only two possible outcomes of the major contract uncertainty: We will either
get the contract, or we won’t. Because there are only a few (in this case, two) possible outcomes,
we say that this uncertainty has a discrete number of outcomes. Each outcome has a significant
possibility of occurring, and so we can assign a probability directly to each outcome.

In describing the example, we said that the company believed there was a better-than-
fifty-fifty chance that it would successfully land the contract. Let us assume that the company
believes there is a .6 probability (60% chance) that it will land the contact, and, accordingly, a .4
probability (40%) that it won’t. (For the moment, we won’t worry about how the company
comes up with these probabilities; this will be a topic for later in the course.) Thus, while the

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -6- UV1461

company believes its chances are better than fifty-fifty, it still believes its chances are not as high
as two out of three (a .67 probability).

Thus, conceptually, in the cell for “Major purchase sales, in units,” we want to represent
that there is a 60% chance that it will be 50,000, and a 40% chance that it will be 0. (More on
how this will be used in the section on simulation.)

General Sales Level: Continuous Probability Distribution

Unlike the major contract uncertainty, the possible outcomes for the general sales level
uncertainty are numerous (for example, 74,311 or 81,284 or 58,935). In fact, they are so
numerous that the probability of getting any one particular sales level is virtually nil. For
example, think about how much you would be willing to bet that the general sales level will be
exactly 76,873 in the upcoming quarter!) Thus, the idea of assigning a probability to each
specific outcome makes little sense, both because of how small the resulting probabilities would
be and because of the size of the task of doing so, with so many possible outcomes.

Still, while it makes little sense to talk about the probability of any individual outcome
occurring, we can talk about the probability that the actual sales level will fall within a particular
range of outcomes. For example, we might feel comfortable saying that the general sales level
has a .5 probability of falling between 65,000 and 85,000 units. The language we use to
communicate the idea that we can only meaningfully talk about the probability of the outcome
falling within ranges is that of continuous probability distributions.

(Technically speaking, “continuous” means that we can get any value, including one with
a fractional component, within a given range. In our example, this is clearly not true, since next
quarter’s sales cannot be 72,411.673. Given how many possible outcomes we do in fact have,
however, it will be useful for us to assume it is, for all practical purposes, continuous.)

Let us assume that the company decides that a triangle probability distribution best
captures its beliefs about the values the general sales level might take. (As with the discrete
distribution above, we won’t worry for the moment how the company determines that this
particular distribution is appropriate, a topic for later.) A triangle distribution has three important
values (or parameters in the lingo of probability distributions): the most likely (best guess),
minimum (worst case), and maximum (or best case). The company has already specified these
three values as 50,000 (minimum), 75,000 (most likely), and 110,000 (maximum). Specifying
these three parameters, and the fact that the underlying probability distribution is triangular,
gives rise to the following continuous probability distribution:

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -7- UV1461

This specific triangle distribution completely determines the likelihood of getting a

general sales level within any given range of values. For example, the likelihood of the general
sales level being between 65,000 and 85,000 is the ratio of the area under the curve between
65,000 and 85,000 to the total area under the curve. (OK: We know that the straight lines that
comprise the triangle do not form a curve in the common usage of the word, but in mathematics
we use “curve” to refer to any line, even a completely straight one!) In the picture below, this
equals the ratio of the shaded region to the total area under the curve and the number line.

To find this ratio requires calculus or geometry, but not to worry! You don’t have to use
or even know calculus. Many software programs find this ratio for you automatically. Any of
these programs would calculate the above ratio as 55.3%. This means that by assuming a triangle
distribution and the three given parameters, the company is implicitly assuming that the
probability of the actual general sales level falling between 65,000 and 85,000 is 55.3% or
slightly better than fifty-fifty.

We can similarly calculate the probability of finding the likelihood of the general sales
level falling within any range, given our distribution. For example, the probability of it being
greater than 100,000:

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -8- UV1461

…is 4.7%. The probability of it being less than 80,000:

…is 57.2%.

Thus, conceptually, in the cell for general sales level, in units, we want to represent that
the likelihood of different values is described by a triangle probability distribution with a
minimum of 50,000, maximum of 110,000, and a most likely value of 75,000. Doing so implies
that the likelihood of the sales level being between 65,000 and 85,000 is roughly 55%; of being
greater than 100,000 is roughly 5%; of being less than 80,000 is 57%; and of being within any
range is as specified, using the same process as in the three examples above.

With probability distributions in place for both major purchase units and general sales
level, we can turn to simulation to explore the impact of these uncertainties on the overall
consequence of continuing the product line in the upcoming quarter.

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -9- UV1461

Simulation

Recall our picture summarizing the consequences of the four possible scenarios we
considered earlier:

At the time, we observed the limitations of this picture: It only lists four possible
consequences out of the many possible, and it gives no indication of how likely even these few
consequences are relative to one another. We are now ready to use simulation to accomplish our
goal of fleshing out this picture.

In simulation, we simulate the upcoming quarter happening many times. You can think of
each time we simulate a quarter as generating a scenario defined by a general sales level and
whether or not we get the major contract. The spreadsheet then calculates the profit
(consequence) accruing in that scenario. In simulation terminology, each scenario we generate is
called a trial.

The key to simulation is in how the scenarios corresponding to each trial are generated.
Rather than the haphazard (or even thought-out) manner in which we chose the four scenarios we
have generated so far, scenarios are generated according to their actual likelihood of occurring.
For example, for our discrete distribution of whether or not we get the major contract, for each
trial, simulation assigns a 60% chance to getting the contract and a 40% chance to not getting the
contract. In any particular trial, of course, the company either gets the contract or not. But over a
series of trials, in roughly 60% of them the company will get the contract, and in roughly 40% of
them it won’t. (The percentages are “roughly” because of the vagaries of chance. For example, if
we run ten trials, each with a 60% chance of getting the contract, we might not end up with
exactly six trials with the contract and four without; we could easily end up with seven or five,
not so easily with nine or four, and possibly even—but not likely—with ten or three, or any other
value less than three.)

For the general sales level, the same process is followed, although it is harder to
conceptualize because of the use of ranges rather than individual outcomes. For each trial, an
individual outcome (such as 61,345) has to be selected, and we have already observed that the
probability of getting any specific outcome is virtually nil. With continuous probability
distributions, simulation generates specific outcomes so that over a number of trials the
percentage of specific outcomes falling within ranges is consistent with the given probability
distribution. For example, with our triangle distribution, over a number of trials, roughly (again)
55% of the selected outcomes for general sales level will fall between 65,000 and 85,000, 5%
will exceed 100,000, and 57% will be less than 80,000. How exactly simulation selects
individual outcomes to ensure this result is not as important as that it does so. (For those
interested, almost any introductory textbook on simulation can describe this process.)

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -10- UV1461

Within a single simulation run, we can conduct as many trials as we like. The more trials
we conduct, the more complete our picture is of what might happen. It is not uncommon for
simulation software available today to be able to quickly and easily accommodate thousands of
trials within a single run.

At the end of the run, we have the economic consequence (profit) associated with each of
the many scenarios (remember each trial equals one possible scenario) generated during the run.
We can collect these consequences and report them as a continuous distribution, where the area
under the curve for any range relative to the total area under the curve reflects exactly what
proportion of our consequences fell within that range. For example, running a simulation of
10,000 trials for our example, using the input probability distributions described above, might
produce the following distribution:

For this distribution, the proportion of consequences falling between ($20,000) and
$30,000 is .1260 (12.60%), which is precisely the ratio of the shaded region below to the entire
area under the curve:

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -11- UV1461

Once again, we don’t have to worry about using calculus to find this area: Simulation
software packages do it for us automatically. (And besides, the .1260 proportion simply means
that 1,260 of our consequences fell within this range, and that can be determined by simply
counting the number of such consequences, something a computer can do very quickly!)

Risk Profile and Expected Value

The above distribution can be thought of as the probability distribution of economic

consequences (profit) if we choose to continue the product line in the upcoming quarter. Thus,
we have a 12.60% chance that our profit will be between ($20,000) and $30,000. This
probability distribution of consequences is referred to as the risk profile if we choose to continue
the product line. Notice how this risk profile overcomes the limitations of our earlier, four-
outcome picture: We now have the entire range of possible outcomes, and we can see how likely
our outcome is to fall into different ranges. Thus, the risk profile tells us everything we need to
know about the potential economics of a given course of action.

Different decision makers might be interested in different aspects of this risk profile. For
example, some might want to know the probability of losing money if we proceed. For our
simulation run, this is 42%, as depicted in the graph that follows:

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -12- UV1461

Some students might be interested in the probability of losing a certain amount of money.
Suppose, for example, we are concerned about the likelihood of losing at least $20,000. Again
from our risk profile, this would be 37%.

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.
 -13- UV1461

One particular aspect of the risk profiles that many decision makers want to know about
is the average economic consequence if we proceed. Precisely as defined, this is simply the
average of the economic consequences of all 10,000 trials we included in our run. The average is
called the expected value of profit or, more simply, the expected profit of continuing the product
line. For our simulation run, the expected profit is $12,123. (While this value is not immediately
obvious by looking at the risk profile, it is easily calculated by any simulation software.) Many
decision makers use this number as an indicator of what the odds favor doing in any situation.
For example, if our risk profile above reflected the economics of a game in a casino rather than a
business situation, where a negative outcome represents a payout by the casino and a positive
outcome represents receipt by the casino from the gambler, then the casino would be more than
happy to offer this game to gamblers. Even though there would be many individual gamblers that
win (that the casino would therefore have to pay) over time, there would also be many losers,
and the casino would expect to make $12,123 times the number of times the game was played.
Casinos can afford to be risk neutral (i.e., not care about the variation around the expected value)
because they can afford the losses from individual plays of the game and are willing to accept
those losses in exchange for getting the wins, which on average more than offset the losses (as
reflected by the positive expected value). In contrast, businesses may be more sensitive to the
risk of bad individual outcomes and therefore turn down opportunities with positive expected
value because of the risk involved. These businesses are called risk averse, because they decline
opportunities whose odds are in their favor because of the risk involved.

This document is authorized for use only in PAD's MBA O 22-24 AD at Universidad De Piura (PAD) from Sep 2022 to Mar 2023.