Monte Carlo Simulation in Projects: P50, P80, Schedule and Cost

Learn how to apply Monte Carlo simulation in engineering projects to analyze schedule and cost risk, interpret P50/P80, calculate contingency, and identify drivers.

Check it out!

Monte Carlo simulation is a quantitative technique that runs a large number of possible scenarios based on uncertainty distributions and risk events. In engineering projects, it makes it possible to move away from a single deterministic date or value and work with a distribution of outcomes, estimating the probability of meeting a given schedule, budget, or contingency level.

Instead of stating that a project “will finish on June 30” or “will cost $50 million,” the analysis may show, for example, that the date has a 35% chance of being met or that a budget corresponds to the P60 percentile of the distribution. This information changes the quality of the decision because it makes explicit the confidence level embedded in the commitment.

Monte Carlo is not a tool for manufacturing precision. The result is reliable only if the schedule, distributions, risks, correlations, and assumptions adequately represent the project. A poor model run ten thousand times is still a poor model.

What is Monte Carlo simulation

Monte Carlo uses repeated random sampling to represent uncertainty. In each iteration, the model draws possible values for the variables according to previously defined distributions, calculates the result, and stores that scenario.

After hundreds or thousands of iterations, a distribution of outcomes is formed. For cost, it shows different possible values and their cumulative probabilities. For schedule, it shows different completion dates and the confidence level associated with each one.

The method is particularly useful when several sources of uncertainty interact and a simple analytical solution does not adequately represent the system.

In projects, Monte Carlo can be used for:

schedule risk analysis;
cost and contingency analysis;
integrated cost and schedule assessment;
comparison of alternatives;
assumption sensitivity analysis;
identification of risk drivers;
assessment of confidence in contractual milestones.

Why a deterministic date can be misleading

Traditional schedules use one duration for each activity. That duration may be a best estimate, an average, a target, or a negotiated value. The problem is that none of these choices eliminates actual variability.

An activity estimated at 10 days may finish in 8, 10, 14, or 20 days depending on productivity, approvals, interfaces, resource availability, and risk events.

When hundreds of activities carry uncertainty, project completion also becomes uncertain. Adding deterministic durations does not reveal the distribution of possible dates.

Simulation makes it possible to observe how variability propagates through the schedule logic network.

A deterministic date may hide a commitment with a low probability of success. Monte Carlo transforms schedule uncertainty into a distribution and reveals the actual confidence level associated with the milestone.

Structure quantitative analyses within risk management →

P50, P80, and other percentiles

Percentiles are points on the cumulative distribution.

If the P50 date is September 30, this means that approximately 50% of the simulated scenarios finished by that date and 50% finished later. If the P80 date is October 20, about 80% of the scenarios finished by that point.

For cost, the reasoning is similar. A P80 of $120 million represents a value that was sufficient in approximately 80% of the model scenarios.

P80 does not mean “80% contingency.” It also does not mean an 80% guarantee. It represents the confidence level produced by the model, conditional on the assumptions and distributions used.

The choice of percentile should reflect risk appetite, commitment criticality, governance, and the consequence of noncompliance.

S-curve and cumulative distribution

Monte Carlo results are often presented as an S-curve, or cumulative distribution.

The horizontal axis shows cost or date. The vertical axis shows cumulative probability. The curve can answer questions such as:

what is the probability of meeting the current budget?
what value corresponds to P80?
what is the difference between P50 and P90?
how much contingency is required to reach a given confidence level?
what date represents a more defensible commitment?

The curve also helps compare alternatives. Two solutions may have the same mean value but very different distributions.

Monte Carlo for schedule analysis

In schedule risk analysis, activities receive duration or uncertainty distributions. Discrete risks can be modeled with a probability of occurrence and an impact on specific activities.

In each iteration, the network is recalculated with new durations and events. The result is a completion date. Repeating the process produces the probabilistic schedule distribution.

This approach is more informative than simply adding a buffer at the end because it considers project logic and interaction among paths.

Schedule quality before simulation

Monte Carlo does not fix a technically poor schedule.

Before simulating, it is necessary to verify:

complete predecessor and successor logic;
absence of unnecessary artificial constraints;
correct treatment of calendars;
consistent durations;
identifiable critical path;
summary activities not used as logic;
defined milestones;
realistic updates;
progress correctly recorded;
near-critical and converging paths.

If the logic network is broken, uncertainty propagation will also be broken.

Quantitative analysis should begin with a schedule audit.

Deterministic critical path vs. probabilistic criticality

The critical path shown in the baseline schedule results from the current deterministic durations. In the simulation, different paths may become critical in different iterations.

Therefore, an activity may have a high criticality index even if it is not on the current deterministic critical path.

This indicator shows in how many scenarios the activity participated in the path that determined completion. It helps uncover hidden risks on near-critical paths.

A schedule with several paths converging on one milestone may be more vulnerable than a simple critical-path reading suggests.

Duration distributions: triangular, beta, normal, or another?

The choice of distribution should represent the phenomenon and the quality of the information.

A triangular distribution can use minimum, most likely, and maximum. It is simple and intuitive, but strongly depends on expert judgment.

Beta or PERT distributions can produce smoother shapes when there is a dominant central estimate. Normal distributions may be inappropriate when negative values are impossible or when asymmetry is relevant.

There is no universally correct distribution. The analyst should justify why a given shape represents the variable.

More important than choosing a sophisticated function is avoiding arbitrary ranges applied equally to all activities.

How to estimate minimum, most likely, and maximum

The three estimates must reflect plausible conditions, not wishes.

One approach is to ask:

plausible minimum: duration achievable under favorable conditions without depending on extraordinary events;
most likely: duration consistent with expected productivity and conditions;
plausible maximum: duration under reasonably conceivable adverse conditions, excluding catastrophes treated as discrete risks.

Historical data are preferable when comparable. When they do not exist, structured interviews with experts should reduce optimism and anchoring biases.

It is advisable to document the origin of each range or rule by activity class.

Variability vs. discrete risk

Not every uncertainty should be modeled in the same way.

Variability is present even when the process occurs normally: productivity, review duration, installation time, and crew output.

Discrete risk is an event that may or may not occur: supplier failure, licensing delay, equipment failure during FAT, or extreme rain interrupting construction.

Modeling everything as a duration range can hide causality. Modeling all variability as discrete risks can generate hundreds of artificial events.

A good analysis separates the two sources.

How to model discrete risks

A risk can have a probability of occurrence and an impact distribution.

In each iteration, the model draws whether the event occurs. If it does, the impact is applied to the activity, cost, or set of related elements.

Example: there is a 30% probability of delay in approval. If it occurs, the impact may range from 10 to 30 days.

This structure preserves the difference between chance of occurrence and magnitude of consequence.

The risk register should identify which events enter the model and how they were parameterized.

Monte Carlo for cost analysis

For cost, estimate components can receive distributions related to quantities, prices, productivity, rates, exchange rates, logistics, or engineering uncertainty.

Discrete risks add impacts when they materialize in an iteration.

The model sums the components to produce a total cost for each scenario. The resulting distribution makes it possible to calculate percentiles, mean, deviation, ranges, and the contingency required for the desired confidence level.

The WBS structure is useful for organizing the elements and identifying where uncertainty is concentrated.

Base estimate and risk analysis

The analysis must distinguish what is already incorporated into the base estimate from what represents additional uncertainty.

If an item has already been estimated with conservative productivity and then receives another broad distribution based on the same risk, double counting occurs.

The model should document:

base value;
nature of uncertainty;
distribution applied;
additional discrete risks;
correlations;
excluded items.

Without this architecture, the result may appear sophisticated and still be economically inconsistent.

Correlation: one of the most critical points

Project variables may move together.

Exchange-rate increases affect several imported items. Low productivity can affect several work fronts. Design delays can shift multiple procurement packages.

If the model assumes complete independence, it may underestimate extremes because simultaneous adverse scenarios appear less frequently than they do in reality.

On the other hand, applying high correlation to everything can artificially inflate dispersion.

Correlations should be justified by a technical mechanism, data, or structured judgment.

Causal dependence is not just statistical correlation

Two risks may be related because one causes the other.

Late approval may delay manufacturing, which reduces the assembly window and compresses commissioning. This is a causal chain, not merely two correlated variables.

Whenever possible, the model should represent causal logic directly rather than using a correlation coefficient as a substitute.

This distinction improves interpretation and allows more effective responses to be defined.

How many iterations are necessary?

The number depends on model complexity and result stability. Thousands of iterations are common because computational cost is low.

The objective is not to reach a magic number but to verify convergence. Relevant percentiles and statistics should remain stable as the number of iterations increases.

Models with rare events may require more iterations to adequately capture the distribution tail.

Running more iterations does not fix poor data.

More iterations do not fix poor assumptions. Simulation robustness depends on the quality of the schedule, distributions, discrete risks, and modeled dependencies — not on the volume of random draws.

See how to relate P50/P80 to contingency reserve →

Sensitivity analysis

After understanding the distribution, the next question is: what most influences the result?

Sensitivity analyses can show which activities, variables, or risks contribute most to cost or schedule.

Tornado charts, correlation with the result, and criticality indices are examples.

This information has management value because it directs effort. If 70% of schedule variability comes from three interfaces, it may be more efficient to treat those interfaces than to add general contingency.

Monte Carlo should support decisions about where to act, not merely produce a P80.

Schedule drivers

In schedules, drivers may be activities with high probabilistic criticality, strong correlation with the completion date, or a wide duration range.

They may also be discrete risks that affect important milestones.

Identifying drivers makes it possible to revise sequencing, bring approvals forward, create supply alternatives, increase resources, review the construction strategy, or protect critical windows.

The analysis transforms an abstract distribution into an action plan.

Cost drivers

For cost, drivers may be linked to high-value items, high uncertainty, exchange rates, quantities, productivity, or high-consequence events.

A low-cost item with high variability may be less relevant than an expensive item with moderate variation.

Sensitivity analysis helps prioritize value engineering, negotiation, hedging, quantity review, contracting, or contingency strategy.

How to relate Monte Carlo to contingency reserve

The distribution makes it possible to choose a confidence level and calculate the difference relative to the base value or another reference point.

If the base cost is $100 million and P80 is $115 million, the $15 million difference can inform the reserve required to reach that confidence level, provided the baseline and reserve methodology is consistent.

The same applies to schedule: the difference between the deterministic date and P80 can guide the commitment margin.

This approach is more transparent than applying an arbitrary percentage, but it depends on model quality.

P50 is not necessarily the best target

P50 represents a balance between scenarios above and below, not a universal recommendation.

For exploratory internal decisions, P50 may be appropriate. For a critical contractual commitment, the organization may choose P70, P80, or another level.

The choice should consider the cost of additional protection and the consequence of noncompliance.

Very high confidence levels can make the project economically unviable; very low levels can produce systematically optimistic commitments.

The decision is one of governance, not software.

Schedule Risk Analysis — SRA

Schedule Risk Analysis applies quantitative techniques to the schedule to assess the probability of meeting milestones and identify schedule drivers.

The process typically involves:

validate the quality of the logic network;
define duration uncertainty;
model discrete risks;
establish correlations when applicable;
run the simulation;
analyze the completion curve;
identify probabilistically critical activities;
test mitigation scenarios.

The result may reveal that the contractual date is far below P50, indicating an aggressive commitment.

Should cost and schedule be analyzed together?

In many projects, cost and schedule are interdependent.

Delay increases mobilization, site administration, equipment rental, and indirect costs. Acceleration can reduce duration and increase cost. Technical failures can produce both.

Integrated cost and schedule models seek to represent these relationships. They are more complex, but may be necessary in large programs.

When cost and schedule are modeled separately, the team should acknowledge the limitations and avoid interpreting the two distributions as independent.

Joint Confidence Level

Some methodologies simultaneously assess the probability of meeting cost and schedule. The Joint Confidence Level (JCL) represents this combined view.

It can be useful when financing and schedule decisions need to consider dependence between the two dimensions.

It is not necessary for every project. Its use requires data maturity and an integrated model.

How to test mitigation scenarios

One of the most useful applications of Monte Carlo is comparing the model before and after a response.

Example: qualifying an alternative supplier can reduce the probability of delay from 40% to 15%. The simulation shows how much this action shifts P80 and reduces contingency.

Another example: bringing an approval forward may remove a near-critical path.

This comparison makes it possible to quantify the benefit of mitigation and support investment decisions in responses.

“What-if” scenarios

The analysis can test alternatives such as:

execute packages in parallel;
contract a secondary supplier;
increase the installation shift;
bring procurement forward;
defer a given function;
change the commissioning sequence;
increase contingency;
reduce scope;
accelerate approval.

Each scenario should keep assumptions documented to allow a fair comparison.

Optimism bias

Experts tend to underestimate duration and impact, especially when commercial or political targets have already been announced.

Collection of distributions should seek to reduce this bias through historical data, independent interviews, reference analysis, and comparison with previous projects.

If all “plausible maximums” are only a few points above the base value, the distribution may be artificially narrow.

Quantitative analysis must challenge the estimate, not merely formalize it.

Tails and extreme risks

Means and central percentiles can hide severe events.

Low-probability, high-impact risks can produce long tails. In safety, operational continuity, or major financial losses, the organization may need to analyze extreme scenarios separately.

Monte Carlo is a tool, but it does not replace disaster scenario analysis, HAZOP, Bow Tie, or other specialized techniques when the nature of the risk requires them.

The mistake of removing risks “because they are unlikely”

Excluding all low-probability risks can artificially reduce the distribution tail.

The inclusion criterion should consider materiality, not frequency alone.

At the same time, adding dozens of remote events without a basis can inflate the result.

The portfolio must be curated using technical criteria and record why each risk was included or excluded.

Historical data and calibration

Models improve when they are compared with actual outcomes.

After project closeout, it is possible to verify:

where final cost fell within the predicted distribution;
whether actual completion was within the simulated range;
which distributions were optimistic;
which risks occurred;
which correlations were inadequate;
whether P80 was systematically conservative or insufficient.

This feedback makes it possible to calibrate future models.

Monte Carlo in early phases

Even with limited definition, the technique can be useful if uncertainties are represented honestly.

In the conceptual phase, broader distributions and parametric scenarios can reflect the lack of definition. The mistake is producing a narrow result to create an appearance of precision.

As the project matures, distributions can be refined with engineering, supplier, and field data.

Model evolution should follow project maturity.

Monte Carlo in procurement

Critical supplies can be modeled considering lead time, drawing approval, manufacturing, FAT, logistics, customs clearance, and installation.

Single-source supplier, manufacturing capacity, and import risks can be included as discrete events.

The analysis helps determine whether the schedule should bring contracting forward, require alternatives, or protect milestones with additional contingency.

Monte Carlo in construction and implementation

Productivity, weather, work-front availability, access, interferences, and rework create significant variability.

Distributions can be based on historical productivity, field data, or expert ranges.

It is important to respect dependencies among activities. A crew’s productivity cannot be sampled independently across dozens of activities if they all share the same field condition.

Monte Carlo in commissioning

Commissioning concentrates integration and readiness risks.

FAT failure, documentation delay, subsystem unavailability, and the need for retesting can shift acceptance.

Modeling these events may show that the time between installation and operation is insufficient for the expected confidence level.

This information makes it possible to extend the testing window, bring pre-commissioning forward, or prioritize critical systems.

How to document a Monte Carlo analysis

A defensible report should record:

baseline version;
data date;
model scope;
activities or costs included;
distributions used;
source of estimates;
discrete risks;
correlations;
number of iterations;
software or method;
P10/P50/P80/P90 results as applicable;
sensitivity drivers;
limitations;
mitigation scenarios;
decision made.

Without documentation, the result is not reproducible.

Monte Carlo analysis flow applied to engineering project schedule or cost

What the software does not decide

Tools automate random sampling and calculations, but they do not decide:

whether the schedule is realistic;
which risk should be included;
which distribution is defensible;
which correlation makes sense;
which percentile the organization should adopt;
which response is economically appropriate;
whether the baseline is contaminated by hidden margins.

These decisions require engineering, management, and governance.

Common mistakes in Monte Carlo models

Among the most frequent mistakes are:

applying the same distribution to every activity;
ignoring relevant correlations;
using a schedule with defective logic;
duplicating risks already embedded in durations;
excluding near-critical paths;
using percentiles without explaining the confidence level;
interpreting P80 as a guarantee;
relying excessively on the mean;
failing to document parameter sources;
running thousands of iterations without validating assumptions;
presenting sophisticated charts without a link to decision-making.

Monte Carlo adds value only when the distribution changes a decision. The objective is not to produce a sophisticated chart, but to decide contingency, commitment, mitigation, prioritization, or strategy based on explicit confidence.

Support critical decisions with Technical Engineering Consulting →

When Monte Carlo is not worthwhile

Not every project needs probabilistic analysis.

If the decision is simple, the exposure is small, and a conservative scenario already provides sufficient information, the modeling cost may not be justified.

The technique adds more value when:

the schedule or cost commitment is material;
there are several sources of uncertainty;
contingency is significant;
there is a need to justify a confidence level;
alternatives must be compared;
the schedule has many competing paths;
correlated risks may change the result.

Sophistication should be proportional to the decision.

Relationship with the risk matrix

The qualitative matrix helps prioritize which risks deserve deeper analysis. Monte Carlo quantifies part of the uncertainty in terms of an outcome distribution.

The two techniques do not compete. The matrix may identify a high supply risk; the simulation can show how much it shifts schedule P80 and how much a mitigation reduces contingency.

The combination improves the transition between qualitative analysis and quantitative decision-making.

Relationship with the risk register

The risk register provides probability, impact, owner, response, and history for discrete events.

Quantitative modeling should maintain links to the register identifiers. Thus, when a risk is mitigated or closed, the model can be updated traceably.

This link prevents the simulation from becoming a parallel file disconnected from project management.

Final considerations

Monte Carlo makes it possible to transform uncertainty into a distribution of outcomes and associate cost and schedule commitments with explicit confidence levels. In engineering, this is particularly useful for contingency, critical schedules, CAPEX, procurement, and decisions in which a single deterministic estimate hides material exposure.

The value of the technique is not in the number of iterations. It lies in the quality of the baseline, distributions, risks, dependencies, and interpretation.

P50, P80, and S-curves are useful only when the organization understands what they represent and uses the information to decide: adjust contingency, review the commitment, treat drivers, test alternatives, or consciously accept a given level of exposure.

A good quantitative analysis does not replace engineering. It makes uncertainty more visible so that technical and management decisions are more defensible.

Technical references

[1] U.S. GOVERNMENT ACCOUNTABILITY OFFICE. Schedule Assessment Guide: Best Practices for Project Schedules. Washington, DC: GAO, 2015. Available at: https://www.gao.gov/products/gao-16-89g

[2] NATIONAL AERONAUTICS AND SPACE ADMINISTRATION. NASA Cost Estimating Handbook — Appendix G: Cost Risk and Uncertainty Methodologies. Washington, DC: NASA. Available at: https://www.nasa.gov/ocfo/ppc-corner/nasa-cost-estimating-handbook-ceh/

[3] PROJECT MANAGEMENT INSTITUTE. Risk Management in Portfolios, Programs, and Projects: A Practice Guide. Newtown Square: PMI, 2024. Available at: https://www.pmi.org/standards/risk-management-in-portfolios

[4] U.S. DEPARTMENT OF ENERGY. Curating the Inputs for a Contingency Reserve Calculation. Washington, DC: DOE, 2022. Available at: https://www.energy.gov/sites/default/files/2023-03/Curating%20the%20Inputs%20for%20a%20Contingency%20Reserve%20Calculation.pdf

Frequently asked questions

What is Monte Carlo simulation in projects?

It is a probabilistic technique that runs many scenarios based on uncertainty and risk distributions to generate a distribution of cost or schedule outcomes.

What does P50 mean in Monte Carlo?

It is the percentile at which approximately 50% of simulated scenarios result in a value or date equal to or lower. Half of the scenarios are above it.

What does P80 mean?

It is the value or date that approximately 80% of simulated scenarios do not exceed, conditional on the model assumptions.

Does P80 mean 80% contingency?

No. P80 is a confidence level on the cumulative distribution, not a contingency percentage.

How many iterations does a simulation need?

There is no universal number. Thousands are common; the important criterion is the stability of relevant percentiles and statistics and adequate capture of rare events.

Can Monte Carlo be used for schedules?

Yes. Schedule Risk Analysis varies durations and risks in the logic network to estimate probabilities of meeting milestones and identify schedule drivers.

Does Monte Carlo replace the risk matrix?

No. The matrix prioritizes risks qualitatively; Monte Carlo quantifies effects on cost or schedule distributions. The techniques are complementary.

When is Monte Carlo not necessary?

When the decision is simple, exposure is small, or scenario analyses already provide sufficient information. The complexity of the technique should be proportional to the decision.

Complementary technical materials