The PERT distribution was a milestone in the development of Monte Carlo simulation models in business. It emerged as early as the 1950s and was put together explicitly as a distribution to fulfil one particular task: the evaluation of large projects.
Below we show how the distribution is defined, what it was developed for and what it looks like. Then we look at when it fits and simulate a project with 25 activities in Excel with Risk Kit. This example also shows us where the limits of the PERT distribution lie.
What is a PERT distribution?
A PERT distribution is described by three values: the minimum, the most likely value and the maximum. The most likely value is also called the mode. The PERT distribution is always bounded, and its values lie between the minimum and the maximum.
It was derived from the beta distribution. The beta distribution is extremely flexible, but its parameters are very abstract. For the PERT distribution it was simplified, and its two abstract parameters were traced back to the minimum, the most likely value and the maximum. Let us write a for the minimum, b for the most likely value and c for the maximum. A PERT random variable X then arises from a beta random variable Y that is stretched to the range from a to c:
X = a + (c − a) · Y
The two parameters α and β of the beta distribution follow from the three values:
α = 1 + 4 · (b − a) / (c − a) and β = 1 + 4 · (c − b) / (c − a)
What was the PERT distribution developed for?
The use case that inspired it at the time was the evaluation of large projects. People wanted to carry out projects that consisted of hundreds or thousands of individual activities. In many areas it was not clear how they would work and what exactly had to be done in the individual activities. So these were research projects and production projects at the same time.
Examples are the development of nuclear missiles for submarines or the staging of the Olympic Games in Grenoble. There, many venues had to be developed, over a period of years, by thousands of people and in a complicated environment in the Alps. The PERT distribution was tailor-made for this situation.
How such a project is built as a model
The situation is structured in an exemplary way, and it shows very well how models are built in the first place. The target quantity is the large project as a whole: we want to know how much it can deviate from the plan.
To find out, we do not ask directly about the delay of the overall project. We break the project down into individual activities and look at how these activities are connected in the project plan. The actual model, that is, the factual relationship between the activities, is therefore already given by the project plan. The only source of uncertainty is the individual activity. The uncertainty of the whole system only results from it.
The PERT distribution was tailor-made for this small building block: for the individual activity, the elementary (atomic) uncertainty, so to speak, which cannot be divided any further.
Three values for a single activity
Which properties matter here? A single activity in a project is limited in scope. It has a minimum duration below which it cannot be done at all. But it also has a maximum duration: by then it is finished, or it escalates and is abandoned. So the time span is limited. Within this time span there is a range in which the duration will most likely lie. Overall, the duration is therefore characterised by three values: the minimum, the maximum and the most likely value.
The second important property: the project manager should be able to assess the distribution of the duration personally. He knows the project and knows the order of magnitude of the effort and of the uncertainty. That is why he can name these three corner points. This is exactly why the beta distribution was simplified into the PERT distribution.
What does a project model with PERT distributions look like?
Our example project consists of five milestones and 25 activities. The milestones run one after the other and partly in parallel, so the project as a whole is already quite hard to see through. We assess each activity separately by its minimum duration, its mode, that is, the most likely duration, and its maximum duration.
Figure 1 shows the left part of the Excel sheet. Each blue box is a milestone with five activities, which are called Aufgabe 1 to Aufgabe 25 (task 1 to task 25) here. The columns “Minimum”, “Modus” (mode) and “Maximum” contain the three values of each task in calendar days. The column “Ist” (actual), shaded in blue, contains the duration that results for the task in one simulation run. The row “Gesamt” (total) gives the duration of the milestone. In the columns “Modus” and “Ist” it equals the largest value of its five tasks, so for milestone 1 it is the mode of 98 of task 2. The arrows show the sequence: milestone 1 is followed by milestones 2 and 3, and milestone 4 runs in parallel. Milestone 5 follows further to the right in the sheet. Let us note task 2 in row 7 for later: it has a minimum of 83, a mode of 98 and a maximum of 165 days. Its “Ist” cell F7 is still empty.
Figure 1: The example project in Excel. Each task has a minimum, a mode and a maximum in days. The “Ist” cell F7 of task 2 is still empty (German interface).
What does the PERT distribution look like?
Figure 2 shows the PERT distribution for task 2 with a minimum of 83, a most likely value of 98 and a maximum of 165 days as an orange area. The horizontal axis shows the duration in days, the vertical axis the density. The higher the area above a duration, the more likely durations in that region are. The blue line belongs to the comparison in the next section.
The density is nice and smooth, without corners or kinks. It starts at the minimum with the value zero, rises to its peak at the most likely value of 98 and ends at the maximum at zero again. The long side is the right one here. There the PERT distribution falls off quickly and is barely visible from about 151 days.
Figure 2: Density of the PERT distribution with minimum 83, most likely value 98 and maximum 165 as an orange area and the triangular distribution with the same three values as a blue line (German interface)
How does the PERT distribution differ from the triangular distribution?
The fact that the PERT distribution falls off quickly on the long side is the most important difference from the triangular distribution. For comparison, we lay a triangular distribution with the same corner values over it. In the Risk Kit input dialog, which we show further below, this is done with “Add overlay” (“Overlay hinzufügen” in the German interface): we choose the triangular distribution and enter the same three values.
It then turns out that the triangular distribution falls off much more slowly at the long end. In Figure 2, to the right of the point where the two curves cross, the blue line lies clearly above the orange area and only reaches zero at the maximum. Extremely long durations for an activity are therefore much more frequent with the triangular distribution than with the PERT distribution.
This was exactly what the modellers wanted: extreme values should be possible, but not so frequent that you see them practically all the time. Nobody considered it realistic for extreme durations to occur all the time.
When does a PERT distribution fit?
A PERT distribution fits when the following points apply:
- The uncertain quantity is bounded. It has a minimum below which it cannot fall and a maximum that it does not exceed.
- Between the minimum and the maximum there is a most likely value.
- Someone who knows the matter well can name these three values, just as the project manager can for the duration of an activity.
- It is an elementary uncertainty that cannot be divided any further, like the individual activity in a project.
- Extreme values should be possible, but rare.
For composite quantities that are fed by many individual uncertainties, however, the PERT distribution is not suitable. The same applies to quantities that are open at the top. More on this at the end of this article.
How do you simulate a PERT distribution in Excel?
In short: Risk Kit is an add-in for Microsoft Excel that lets you build and analyse Monte Carlo simulations directly in the spreadsheet. What we show here applies in general, though. Whether you simulate with Risk Kit, with Python or with any other tool: the PERT distribution has the same parameters everywhere and delivers the same results.
We stay with the example project from Figure 1. The three values of task 2 are in row 7: the minimum of 83 in cell C7, the mode of 98 in cell D7 and the maximum of 165 in cell E7. The PERT distributed duration is to be created in cell F7, the “Ist” cell of task 2. The other tasks in the model are already PERT distributed in the same way. As the result, we are interested in the plan deviation, that is, how many days longer or shorter the whole project takes than planned. This cell is already marked as an output in the model and highlighted in yellow. We proceed in four steps:
- Select cell F7, open the button for popular distributions on the Risk Kit toolbar (“Beliebt” in the German interface) and choose “Pert”.
- In the input dialog, minimum, mode and maximum are set as the parameters. In the fields Min, Mode and Max we enter the references C7, D7 and E7. To the right of the fields, Risk Kit shows the values the references point to: 83, 98 and 165.
- Clicking “Insert” (“Einfügen”) makes Risk Kit write the function Pert_variate with these three references into cell F7, that is, =Pert_variate(C7;D7;E7).
- Click “Simulation” on the Risk Kit toolbar.
Figure 3 shows the input dialog. On the left you can see cells C7, D7 and E7 with the values 83, 98 and 165, and the “Ist” cell F7 is still empty. The title bar of the dialog names the target cell and the function: [F7] = Pert_variate(…). Below it, Risk Kit describes the function as a random number from a Pert distribution with the parameters min, mode and max. The fields Min, Mode and Max contain the references C7, D7 and E7, with their values to the right. Further to the right in the same dialog, outside the section shown, Risk Kit displays the density from Figure 2.
Figure 3: Input dialog of the Pert_variate function for cell F7. Min, Mode and Max refer to cells C7, D7 and E7 with the values 83, 98 and 165 (German interface).
Overall, the project has a planned duration of 335 days. It follows from the most likely values, which here match the project plan. What interests us is the deviation from this plan. It can go up or down: we can finish sooner if everything goes very well, and we can also take longer.
Figure 4 shows the right part of the sheet with milestone 5. Below it, on the left, is the planned project duration of 335 days and, on the right, the plan deviation, the output highlighted in yellow. In the figure it is 4 days. This is only the result of a single calculation, because the durations in the column “Ist” are random numbers. Only the simulation shows how large the plan deviation can become.
Figure 4: Milestone 5 with the planned project duration of 335 days and the plan deviation, highlighted in yellow as the output (German interface)
The result of the simulation
After the simulation, Risk Kit shows the statistics for the plan deviation (Figure 5). On the left is the list of outputs and inputs. Under “Ausgabe” (output) is the plan deviation, under “Eingabe” (input) are the random numbers of the model, for example tasks 11 to 20. Our cell F7 appears there with its address because we did not give it a name. On the right are the statistics of the plan deviation, all in days.
The mean is 23 days, which is clearly positive. On average we need considerably longer than planned, roughly 6 to 7% longer than the 335 days of the plan. Finishing ahead of plan can amount to up to 26 days, as the minimum of −26 days shows. But it is relatively unlikely: the 5% quantile is −4 days, and the 10% quantile is already 1 day. So finishing ahead of plan only happens in 5 to 10% of cases. In more than 90% of cases the project takes longer than planned, in the simulation by up to 102 days (minimum −26 plus range 128 days).
Figure 5: Statistics of the plan deviation after the simulation, including a mean of 23 days, a minimum of −26 days, a 5% quantile of −4 days and a 10% quantile of 1 day (German interface)
At the top of the table in Figure 5 there is a small histogram. Figure 6 shows it larger. The height of the bars shows how often plan deviations in that range occurred in the simulation. The horizontal axis shows the days. The dashed line in the middle marks the mean at 23 days. The line on the right marks the 95% quantile at a delay of 56 days: in 95% of cases the project is at most 56 days late. The line on the left is at 0, that is, at the planned value. Its label says 8.93%: we actually beat the planned value in about 9% of cases.
Figure 6: Histogram of the plan deviation with the mean, the planned value 0, which is beaten in 8.93% of cases, and the 95% quantile at 56 days (German interface)
Where are the limits of the PERT distribution?
This example also shows us a limit of the PERT distribution. We used it for an elementary uncertainty, the individual activity. What happens if we want to use it for the overall plan deviation? In practice, people often try exactly this. Because it is so simple, the PERT distribution is used for much more than just elementary risks, including composite risks that are fed by many individual uncertainties.
So we compare the simulated distribution with the PERT distribution. To do this, we fit various distributions to the simulated values in Risk Kit, which is also called calibrating them. We then see that some distributions fit the simulated distribution wonderfully, for example the Poly distribution. The PERT distribution, however, differs clearly from the result of the simulation. It considers very long delays much more likely than they were in the simulation, and very large accelerations as well.
Figure 7 shows the result of the calibration. At the top left are the two calibrated distributions, Poly and Pert, each with a Kolmogorov-Smirnov value, a measure of the deviation from the simulated data: 0.0030 for Poly and 0.0841 for Pert. The smaller the value, the better the distribution fits. Below are the estimated parameters of the PERT distribution, A, B and C, that is, minimum, mode and maximum, at about −26, 18 and 102 days. In the chart on the right, the orange bars are the simulated values of the plan deviation, the blue line is the Poly distribution and the green line is the PERT distribution. The blue line follows the bars almost exactly. The green line lies clearly below the bars in the middle, but above them at the right end and at the far left.
Figure 7: Calibration to the simulated plan deviation. The Poly distribution (blue) follows the data, while the PERT distribution (green) lies above the bars at both ends and below them in the middle (German interface).
Why is that? For the project plan to get completely out of hand, with delays of dozens up to a hundred days, a chain of unfortunate circumstances has to occur. A lot of negative things have to come together at once, and that is very unlikely. In practice it is much more likely that some activities go very well and others go worse. In total this partly evens out, so that we end up in a middle range with a very high probability. The PERT distribution underestimates this.
In the good range it is exactly the same. For us to get there, everything has to go very well. The PERT distribution overestimates the probability of this, so here it is too optimistic. At the long end, on the other hand, it is too pessimistic. So it is too pessimistic and too optimistic at the same time.
This is a subtle but characteristic limit of the PERT distribution. It is not suitable for describing composite risks. There, thin and long tails are the order of the day, and the PERT distribution does not have them.
What is the PERT distribution well suited for?
Conversely, the PERT distribution is very well suited for elementary risks. There, the expert estimates that are usually used for it are not particularly problematic either. This is because very concrete things have to be estimated. Experts know these concrete things much more intimately than if they had to combine many risks into an overall risk in their heads.
This leads to the surprising situation that we have a model in which every single risk factor is PERT distributed, but the overall distribution of the system is not. The superposition of the individual factors has produced a different distribution.
What if the uncertainty has no upper limit?
This limit becomes even clearer when a system is not bounded in its overall impact, that is, when the uncertainty is open at the top. We see this in financial markets, for example. Short-term price movements can sometimes show extreme swings. There we cannot easily state an upper limit.
The PERT distribution, which is always bounded, is not at all well placed for this case. You have to choose a different distribution that can reproduce this openness at the edges.
More distributions
For an overview of all distributions, see our guide “When Do I Use Which Distribution?”.
- When do I use a Bernoulli distribution in a Monte Carlo simulation?Definition, formula and examples: when the Bernoulli distribution fits and what role it plays as a switch in a simulation.
- When do I use a binomial distribution in a Monte Carlo simulation?Definition, formula and examples: how the binomial distribution describes the number of successes in n independent trials and how to calculate it in Excel.
- When do I use a Poisson distribution in a Monte Carlo simulation?Definition, formula and examples: how the Poisson distribution describes the frequency of independent events and when it fits.
- When do I use a normal distribution in a Monte Carlo simulation?Definition, formula and examples: when the normal distribution fits, where its limits lie and how to truncate it.
- When do I use a triangular distribution in a Monte Carlo simulation?Definition, formula and example: when the triangular distribution fits, why the most likely value is not the expected value and where its limits lie.
- Quantifying expert estimates in a Monte Carlo simulation: the Expert distributionQuantifying expert estimates better: how the Expert distribution captures estimates through bounds and quantiles, even without a fixed minimum or maximum.
- When do I use a lognormal distribution in a Monte Carlo simulation?Definition, properties and limits: when the lognormal distribution correctly describes proportional growth processes and how to parametrise it in Excel.
- Custom distributions in a Monte Carlo simulation: the Flex distributionCustom distributions: how the Flex distribution is built from values and probabilities and when it fits, from expert estimates to transferring simulation results.