M4.3 — Quantitative Risk Analysis

M4.3 — Quantitative Risk Analysis

1. Practical outcome

By the end of this section, the learner should be able to take a defined project objective, a credible current schedule or cost model, and a selected set of material uncertainties and turn them into a transparent quantitative view of possible outcomes. They should be able to distinguish deterministic dates and costs from probability distributions, decide which risks and sources of variability should be modelled, develop defensible ranges and event assumptions, understand the purpose of Monte Carlo simulation, interpret confidence levels and probability-of-success outputs, identify the uncertainties that drive the result, and use scenario analysis to support a management decision.

The objective is not to produce a sophisticated-looking probability curve. It is to answer practical questions such as: How likely are we to meet the opening date? What cost range is credible rather than merely possible? Which uncertainties are driving the exposure? Which response would materially improve the probability of success? How much confidence should management place in the result?

M4.3 builds directly on M4.1 and M4.2. M4.1 identifies and frames uncertainty. M4.2 prioritises and challenges relative significance. M4.3 quantifies selected uncertainty where the decision warrants it. It does not replace qualitative judgement, and it does not mean every risk in the register should be converted into a probability distribution.

A useful question throughout this Knowledge Area is: “If management is going to make a commitment based on this number, can we explain what uncertainty was modelled, what was not modelled, what evidence supports the inputs, and what would have to change for the result to improve?”

2. Core concepts

Quantitative analysis answers a different question from qualitative scoring

A qualitative matrix helps rank and discuss risks. A quantitative analysis estimates the range and likelihood of possible project outcomes. The two are related but not interchangeable. A risk with a high qualitative score may have limited influence on the final completion date if it sits on a path with substantial float. A medium-scored risk may become a major schedule driver if it affects a near-critical interface with little recovery opportunity.

The move from M4.2 to M4.3 should therefore be driven by the decision. Typical questions include:

·       What is the probability of completing by the fixed public opening date?

·       What date corresponds to a chosen confidence level such as P50 or P80?

·       What is the probability that final cost remains within the approved funding limit?

·       Which risks, activities or assumptions contribute most to the spread of possible outcomes?

·       How much does a proposed mitigation or recovery action improve the probability of success?

·       Does the project need a range, contingency or management action rather than a single deterministic promise?

Deterministic models show one outcome; probabilistic models show a distribution of outcomes

A deterministic schedule or cost forecast uses one duration, one cost and one set of assumptions for each element. It is essential for day-to-day control, but it does not show how uncertain the result is. A probabilistic model repeatedly samples from defined ranges and discrete risk events to generate many plausible outcomes. The result is a distribution rather than a single date or cost.

If a deterministic schedule says integrated testing finishes on 15 November, that does not mean there is a 100% probability of finishing on 15 November. It means that, under the current set of single-point assumptions, the network calculates to that date. M4.3 asks how the calculated outcome changes when credible uncertainty is introduced.

Discrete risk events and general variability are different kinds of uncertainty

A useful quantitative model distinguishes between event risks and continuous uncertainty or variability.

·       A discrete event risk is something that may or may not occur: for example, a specialist supplier may lose an agreed manufacturing slot. The model therefore needs both a probability of occurrence and a consequence if the event occurs.

·       Variability exists even if no named event occurs: for example, installation productivity may reasonably vary because room complexity and access conditions differ. The model therefore needs a credible range around duration, productivity or cost.

Both can matter. Modelling only named register risks can understate ordinary variability. Modelling only activity ranges can hide important event-driven exposure. The analyst should understand what each input represents and avoid representing the same uncertainty twice.

Expected value is useful, but it is not the same as a risk distribution

For a simple discrete cost risk, expected monetary value (EMV) can be calculated as probability multiplied by cost impact. For example, a 30% probability of a 1.0m CU cost consequence has an EMV of 0.30m CU. That is useful for comparing or aggregating exposure, but the project will not necessarily incur 0.30m CU. In a simple event model it may incur zero or 1.0m CU. The distribution of possible outcomes matters when setting commitments or testing confidence.

The same issue applies to schedule risk. A 30% probability of a ten-working-day delay does not mean the deterministic programme should automatically be extended by three days. The event may not happen, and if it does happen the network effect depends on where the delay lands, available float, concurrent paths and recovery options.

Three-point ranges are a way to express uncertainty, not a ritual

For durations, costs or productivity assumptions, practitioners often elicit an optimistic, most-likely and pessimistic value. These three points can be used to define a distribution such as triangular or beta-PERT. The choice of distribution matters less than whether the values are evidence-based, represent the intended uncertainty and are used consistently.

Record whether optimistic and pessimistic values are intended as hard bounds or specified percentiles. Do not enter elicited P10/P90 values as triangular minimum/maximum values without a justified conversion. The most-likely value is a mode, not automatically a mean or P50. Check whether the selected distribution and its tails represent the stated uncertainty.

If these values are merely “10% either side” of the deterministic duration without evidence, the model may look quantitative while containing no additional knowledge. Ranges should come from historical performance, reference classes, supplier information, productivity evidence, uncertainty workshops or explicit expert judgement that has been challenged for optimism and anchoring.

Probability distributions are modelling choices, not facts

A distribution is a mathematical representation of uncertainty. Common choices include triangular, beta-PERT, uniform, normal, lognormal and discrete distributions. The analyst should understand the behaviour being represented. A uniform distribution implies values within the range are equally plausible. A triangular or beta-PERT distribution places more emphasis around a central value. A lognormal distribution may be useful where adverse overrun has a longer tail than favourable underrun.

Do not choose a distribution because the software default looks familiar. Choose it because its shape is a reasonable representation of the uncertainty and because the input evidence is proportionate to the decision. Where evidence is weak, the model should expose that weakness rather than disguise it with more decimals.

Monte Carlo simulation is repeated conditional calculation

Monte Carlo simulation repeatedly samples values from the defined distributions and risk-event assumptions, recalculates the schedule or cost model, and stores the resulting completion date or final cost. After many iterations, the accumulated results provide an estimate of the distribution of outcomes under the modelled assumptions.

The simulation does not discover truth. It calculates the consequences of the assumptions supplied to it. A thousand or a hundred thousand iterations cannot compensate for a weak baseline, unrealistic ranges, incorrect risk mapping or missing correlations.

Schedule risk analysis depends on schedule quality

A schedule risk analysis (SRA) is only as credible as the network it analyses. Before introducing uncertainty, the schedule should have defensible logic, realistic remaining durations, controlled constraints and calendars, and a credible current status. If the schedule is full of hard constraints, open ends or unrealistic logic, the simulation may quantify defects in the model rather than uncertainty in the project.

The relationship with Module 2 is therefore direct. M2.2 provides logic and float discipline; M2.4 establishes a credible current position; M2.5 establishes the deterministic forecast and scenario logic. M4.3 then asks how uncertainty around that model changes the probability of different outcomes.

Cost quantitative risk analysis depends on cost-model traceability

A cost QRA similarly requires a cost model that can be reconciled to scope, current commitments, actuals, accruals and the remaining forecast. Uncertainty can be applied to remaining quantities, rates, productivity, escalation, commercial exposures and discrete risks, but the model must avoid mixing cost already incurred with uncertain future cost unless there is a specific reason.

Where a schedule risk drives cost - for example, a two-month delay creates extended project-management and facilities cost - that dependency should be represented once and transparently. Adding the same delay consequence as both a schedule-driven cost and a separate risk allowance creates double counting.

Define the cost boundary before sampling

At a common cut-off, express each trial final cost as reconciled incurred cost plus a sampled estimate to complete. Keep valid incurred cost fixed unless a separately stated accounting or valuation uncertainty is being analysed. Replace an uncertain remaining component with its sampled value; do not add that whole component again to an EAC that already contains it. Model only the incremental effect of a discrete event on top of the defined base.

Document which embedded allowances, commercial exposures, escalation and opportunities are already included in the deterministic ETC. Remove or reconcile the corresponding allowance before adding a separately modelled risk for the same consequence. When schedule drives extended attendance cost, calculate that effect from the sampled delay and applicable cost rate once. Compare total-cost outcomes with the authorised cap on the same scope and price basis.

Correlation and common-cause exposure can change the result materially

Project uncertainties are rarely fully independent. The same workforce constraint may affect several activities. The same design maturity problem may increase both rework and testing duration. The same market condition may affect several supplier prices. If the model treats every uncertainty as independent, favourable and adverse effects may unrealistically cancel one another.

Correlation should not be added mechanically to every input. It should be used where a defensible common relationship exists. Sometimes the clearer approach is to model the underlying common-cause event directly rather than apply an abstract correlation coefficient to many variables.

Opportunities should be modelled where they are credible

Quantitative risk analysis should not be threat-only. A staged handover, early access arrangement, productivity improvement or alternative supplier route may create a genuine favourable outcome. Opportunities need the same discipline as threats: probability, consequence, conditions for realisation, ownership and evidence. A speculative upside should not be used to neutralise a well-evidenced downside.

Confidence levels describe the modelled distribution, not certainty in the real world

A percentile is a point in the cumulative modelled distribution. P50 is a median; P80 is a date or cost met or bettered in at least 80% of modelled outcomes under the stated quantile convention. P80 cannot be earlier or lower than P50 for the same distribution, but they can coincide where outcomes are discrete. Percentiles are conditional model outputs, not guarantees.

A P80 date is therefore not an 80% guarantee. It is conditional on the quality and completeness of the model. Unknown scope, systemic change, excluded commercial events or poor input data can still invalidate the apparent confidence. Confidence in the model and confidence level within the model are different ideas.

Probability of meeting a target is often more useful than quoting a percentile

Where the project has a fixed opening date or funding cap, management often needs to know the probability of meeting that target. If the model says there is a 68% probability of opening by the committed date, that directly informs the decision. Saying the forecast is “P68” is less helpful unless the audience understands what that means and which assumptions sit behind it.

Sensitivity analysis helps explain what drives the distribution

A QRA should not finish with a histogram or cumulative curve. It should identify which inputs materially influence the outcome. Depending on the model and software, useful measures can include correlation-based sensitivity, tornado charts, criticality index, schedule sensitivity indices and scenario contribution analysis.

These outputs must be interpreted carefully. A variable can correlate strongly with the outcome without being a root cause. A risk may have a low probability but very high consequence. A schedule activity may frequently appear on the simulated critical path because it sits in a structurally sensitive area. The purpose is to direct management attention and challenge, not to turn sensitivity percentages into a new league table.

Quantitative outputs should lead to decisions and response testing

The most valuable use of QRA is often comparative. Run a credible “current exposure” model, then test a defined response: secure the supplier slot, add an assurance review, stage testing, resequence work, change resource levels, or remove a design uncertainty. If the action materially improves the probability of success, the model helps management judge whether the response is worth its cost and disruption.

This is where M4.3 connects forward to M4.4 Risk Response & Control. M4.3 quantifies the current and alternative exposure; M4.4 governs the actions that change it. It also connects to M4.5 Contingency & Uncertainty, where the organisation decides how quantified exposure should inform contingency and reserve decisions. M4.3 should not pre-empt those governance decisions.

Model transparency is more important than mathematical sophistication

A robust QRA should allow another competent practitioner to understand the decision question, source model, risks included, uncertainty ranges, distributions, correlations, exclusions, assumptions, scenarios and interpretation. If the result can only be defended by the person who built the simulation, the analysis is not mature enough for governance use.

3. Step-by-step method

1. Define the decision question and scope of analysis. State what management needs to know: probability of meeting a milestone, likely final cost range, confidence date, funding exposure, or effect of a response. Define the organisational and time boundary so the model does not accidentally answer a different question.

2. Confirm the deterministic source model is credible. For schedule QRA, validate status, logic, remaining durations, calendars, constraints and interfaces. For cost QRA, validate scope, current cost position, commitments, accruals, forecast structure and price basis. Do not use QRA to legitimise an unreliable source model.

3. Select the risks and uncertainty sources that are material to the decision. Use M4.1/M4.2 evidence, schedule and cost drivers, assumptions, historical variability and expert challenge. Not every risk belongs in the model. Include enough to represent material uncertainty without making the model untraceable.

4. Classify each input as event risk, range uncertainty, scenario assumption or excluded exposure. This forces clarity about what is being modelled. An event risk needs probability and consequence; a duration or cost range needs a distribution; a scenario may represent a management choice; an excluded exposure should be documented so decision-makers understand the model boundary.

5. Map each uncertainty to the part of the control model it can affect. Link schedule risks to activities, milestones, logic or calendars and cost risks to WBS/CBS elements or remaining forecast components. The mapping should reflect physical cause and consequence rather than simply attach every risk to project finish or total cost.

6. Develop input ranges and event consequences from evidence. Use actual productivity, reference-class data, supplier evidence, benchmark ranges and structured expert judgement. Define what optimistic and pessimistic values mean. Challenge anchoring to the deterministic forecast and challenge extreme values that are either implausibly narrow or catastrophe-driven.

7. Choose proportionate distributions and probabilities. Select distributions that represent the evidence and behaviour. Keep the level of mathematical sophistication proportionate to input quality. Record probability sources and the rationale for distribution shape.

8. Model dependencies, correlation and common causes. Identify where inputs are linked. Prefer explicit common-cause modelling where possible. Use correlation only when its meaning is understood and can be defended.

9. Include credible opportunities and response conditions. Model favourable events or productivity improvements only where they have clear enabling conditions, ownership and evidence. Do not add upside merely to balance downside.

10. Check for double counting and omissions. Test whether the same uncertainty appears in a duration range and a discrete event, in both schedule-driven and direct cost impacts, or in both estimate uncertainty and risk allowance. Also look for material uncertainties omitted because they sit outside the formal risk register.

11. Run deterministic and single-input reasonableness checks. Before full simulation, test that individual risk events and range changes move the model as expected. A ten-day delay should not create zero network effect unless float or parallel logic explains it. A 0.5m CU cost event should reconcile to the affected cost structure.

12. Run the simulation with enough iterations for stable outputs. Use an iteration count that produces reasonably stable percentiles and sensitivity results for the decision. Do not equate a larger iteration count with better input quality. Record the configuration so the analysis is reproducible.

13. Review output stability and data quality. Check whether rerunning the model produces materially different key percentiles or sensitivities. Investigate unstable results, implausible tails, date spikes caused by constraints, and distributions driven by a small number of questionable inputs.

14. Interpret the probability curve against real decision thresholds. Report probability of meeting committed dates or budgets, plus selected confidence points where useful. Explain the difference between deterministic forecast, target, P50, P80 and any governance threshold.

15. Identify the dominant drivers and test whether they make operational sense. Use sensitivity and criticality outputs to identify the main contributors. Challenge whether the model is revealing a genuine delivery driver or a modelling artefact such as an excessively wide input range.

16. Test defined mitigation, recovery and opportunity scenarios. Change only the assumptions affected by the response. Compare probability-of-success, confidence dates/costs and driver rankings. Avoid scenario comparisons built on inconsistent underlying assumptions.

17. Reconcile the quantitative result with the risk register and control forecasts. Update the understanding of material risks, ensure assumptions and triggers are consistent, and explain where QRA reveals exposure that was not obvious from qualitative scores. Do not silently overwrite the deterministic control forecast with a percentile date or cost.

18. Communicate the result, assumptions and decision clearly. Present the current exposure, key drivers, confidence, exclusions and effect of possible actions. Preserve the model and input basis so later analyses can distinguish genuine risk reduction from changed assumptions or model changes.

4. Worked Reference Project example — Community & Learning Centre

Context — Month 16 readiness to protect the fixed public opening

The Community & Learning Centre has a fixed public opening date. The current deterministic schedule has been updated and challenged through M2.4 and M2.5. It shows integrated testing, training and operational-readiness work completing eight working days before the opening date. Management wants to know whether those eight days represent genuine protection or simply a single-point forecast with significant uncertainty hidden behind it.

The team therefore defines the QRA question as: “What is the modelled probability of achieving the fixed opening date, what uncertainty is driving the outcome, and which near-term management actions most improve that probability?”

Step A — confirm the deterministic model

The schedule-control review confirms that the current schedule has credible status, no unexplained open ends on the readiness paths, controlled constraints, updated remaining durations and a clearly traceable fixed opening milestone. The deterministic finish is eight working days before opening. That becomes the reference point - not the assumed risk-adjusted answer.

Step B — select and classify material uncertainties

The team deliberately excludes a low-probability major building-loss event from this analysis because it would require a different resilience decision and would distort the operational question. The exclusion is recorded rather than silently ignored.

Assign the excluded building-loss scenario to a separate resilience assessment with an owner and decision route. Report the opening probabilities below as conditional on the stated model boundary, including that exclusion. They are not all-hazards probabilities of opening; recording an exclusion alone does not remove its relevance to the overall commitment.

Step C — challenge ranges and dependencies

The technology productivity range is based on actual installation and configuration performance from the first completed rooms, adjusted for the greater complexity of the remaining audio-visual areas. The readiness range is based on observed staff-release constraints. The fire-system and supplier events are based on current supplier and assurance evidence rather than historical register scores alone.

The review also identifies a common dependence: if fire-system acceptance slips, staged access for training is less likely to be realised in the affected areas. Rather than apply a broad mathematical correlation to every readiness activity, the model conditionally limits the staged-access opportunity when R2 occurs. This is more transparent because it reflects the physical dependency.

Step D — understand one risk before the full simulation

R2 has a 30% probability of occurring and, if it occurs, a 3-8 working-day consequence to the integrated testing interface. A simple expected delay calculation might suggest roughly 1.65 working days if a 5.5-day average consequence is assumed: 30% × 5.5 ≈ 1.65 days. That is useful as a sense check, but it is not the schedule answer. The event either occurs or does not occur in each simulation iteration, and its network effect depends on the sampled state of the other activities and available float.

Step E — run the current-exposure model

The following schedule outputs are illustrative teaching assumptions used to discuss decisions. They are not reported results from a supplied executable network or simulation. Reproducing them would require the full network, calendars, distributions, dependency rules and configuration.

This changes the management conversation. The schedule does not simply have “eight days contingency”. Under the defined uncertainty, the opening-date commitment has a modelled success probability of 72%, and an 80% confidence date falls after the commitment. The deterministic buffer is therefore not equivalent to an 80% confidence position.

Step F — identify the drivers

Sensitivity analysis indicates that the most influential contributors to opening-date variation are the fire-system acceptance event, technology configuration productivity, the specialist supplier window and operational-readiness duration. The staged-access opportunity has a favourable effect but does not fully offset the two largest threats.

The team does not interpret these as percentages of blame. They are used to ask where management intervention can most effectively change the distribution.

Step G — test a defined response scenario

Management considers three near-term actions: secure the specialist supplier support window contractually, complete a pre-witness fire-system assurance review before formal testing, and approve staged training access only for rooms that pass agreed readiness criteria.

The response scenario changes only the supported quantitative assumptions:

·       R1 probability reduces from 25% to 10% after written supplier-slot confirmation and named backup support are obtained.

·       R2 remains possible, but the consequence range reduces from 3-8 days to 1-4 days after the pre-witness review and defect-closeout plan.

·       The staged-access opportunity probability increases from 40% to 60% because the enabling criteria and room-release sequence are formally agreed.

The illustrative rerun increases the probability of meeting the fixed opening from 72% to 88%, moves P80 to three working days before opening, and reduces the sensitivity of the fire-system event. The precise numbers are less important than the decision logic: the model shows that targeted assurance and supplier actions materially improve the probability of success.

Step H — extend the same logic to cost

For learning purposes, assume a broader cost QRA on the current forecast. The deterministic total EAC is 49.2m CU against 50.0m CU authorisation. Remaining uncertainty includes supplier commercial exposure, productivity-driven labour cost, escalation on uncommitted items and a possible testing/rework event. The following outputs are illustrative teaching assumptions, not results reproduced from a supplied executable model.

The cost result is not automatically a request to increase the baseline or budget. It is information for management and for the contingency/governance treatment addressed in M4.5. M4.3 makes the exposure visible and testable; it does not decide the funding action.

Step I checking a small cost risk model by hand

Use a separate reduced Month-16 teaching model to verify the calculation mechanism. Reconciled incurred cost is fixed at 30.0m CU. Deterministic remaining cost is 19.2m CU: 18.0m CU other remaining work and 1.2m CU testing/support. The 49.2m CU base EAC excludes the two incremental events below. Assume no other variability or allowance, a constant price basis and independence solely for this demonstration.

Event A is a fire-test correction costing 0.4m CU with probability 30%. Event B is loss of a supplier support slot costing 0.6m CU with probability 20%. These are deliberately simplified teaching inputs, distinct from the wider schedule-event ranges and cost-output assumptions above. Trial EAC = 30.0 + 19.2 + 0.4 × A + 0.6 × B, where A and B are 0 or 1. Enumerate all four combinations rather than simulate.

The probabilities total 100%. Expected final cost is 49.2 + (0.30 × 0.4) + (0.20 × 0.6) = 49.44m CU. Using the lowest outcome whose cumulative probability reaches the requested level, P50 = 49.2, P80 = 49.6 and P95 = 50.2m CU. Probability of remaining within 50.0m CU is 94%. Thus the mean, percentile allowance and chance of meeting the cap answer different questions.

At P80, the incremental allowance above this reduced-model base is 0.4m CU. This is not the amount to add to the existing authorised budget or a replacement for the broader project QRA. The reduced model omits productivity variability, escalation, common causes and wider risks; its 94% outcome cannot be substituted for the separate illustrative 57% or 69% project results.

If evidenced responses reduce both event probabilities to 10%, the four probabilities become 81%, 9%, 9% and 1%; expected cost becomes 49.30m CU and chance of staying within the cap becomes 99%. P80 is then the 49.2m CU base under this discrete model. Test sensitivity to dependence before relying on the result: if both events share a cause, multiplying independent probabilities is invalid. Response costs must be included in the corresponding scenario before judging net benefit.

Management summary from M4.3:

The deterministic schedule retains eight working days between the current readiness forecast and the fixed opening date, but quantitative analysis indicates only a 72% modelled probability of achieving the opening under current uncertainty; the current P80 date falls two working days after the commitment. Fire-system acceptance, technology configuration productivity and the specialist supplier support window are the dominant drivers. The proposed pre-witness assurance review, secured supplier window and controlled staged-training arrangement improve the illustrative opening probability to 88% and move P80 to three working days before opening. Cost analysis similarly shows that the 49.2m CU deterministic EAC understates uncertainty around the 50.0m CU funding limit, with a current illustrative 57% probability of remaining within budget. Management should treat the QRA as decision evidence, not as a replacement for the controlled deterministic schedule or cost forecast.

5. Interpretation and professional judgement

The model is conditional, not predictive in an absolute sense

Every QRA is conditional on the model boundary, source data, input distributions and assumed dependencies. The practitioner should be able to say, “Under these assumptions, the model indicates…” rather than present a percentile as a guaranteed future state. This language is not weakness; it is accurate professional communication.

Do not let the simulation replace the delivery conversation

A risk model can show that technology productivity is influential, but it does not explain why productivity is uncertain. The analyst must return to the work: room complexity, access constraints, commissioning sequence, skills, supplier support and defect rates. Quantitative outputs are most useful when they sharpen operational questions.

Wide ranges can be an important finding, not a modelling inconvenience

Teams sometimes narrow inputs simply because a broad range produces an uncomfortable result. If the evidence genuinely supports a wide range, the uncertainty is real and should be managed. The correct response may be to obtain information, run a trial, secure a supplier decision or break the work into better-defined components, not to make the model look more confident.

A high percentile should not become a hidden deterministic target

If governance chooses to use a P80 date or cost for a particular commitment, the underlying deterministic control model should still remain visible and managed. Replacing all internal forecasts with a high percentile can reduce urgency, consume contingency invisibly and confuse risk allowance with planned work. The relationship between control forecast, commitment and confidence level should be explicit.

Do not confuse percentile confidence with management risk appetite

P50, P80 and P90 are statistical positions within the modelled distribution. They are not universal definitions of “acceptable”. The appropriate confidence threshold depends on the consequence of missing the objective, the organisation’s risk appetite, available contingency, contractual position, public commitment and ability to take corrective action. Governance chooses the threshold; the analyst explains the implications.

Correlations deserve challenge because they can create hidden model leverage

A small set of correlation assumptions can materially widen or narrow the distribution. Where possible, explain the physical common cause instead of relying on opaque coefficients. If coefficients are used, document the basis and test sensitivity to them. Correlation is not a cosmetic refinement.

The adverse tail deserves attention even when the median looks comfortable

A project can have a favourable P50 and still face a significant tail risk that matters to management. The question is not only “What is the median?” but also “What causes the adverse tail, how far can it extend, and what decisions reduce it?” This is especially important for fixed public dates, regulatory milestones and funding caps.

Quantification does not make weak risks objective

Inputs that originate in expert judgement remain judgement, even after they are placed into distributions. Good practice is to make the judgement explicit, challenge it with evidence and compare it with reference-class performance. Numbers improve structure; they do not eliminate bias.

Unknown and excluded uncertainties should remain visible

A QRA can only model what has been represented. Material exclusions, systemic uncertainties and emerging issues should be communicated alongside the probability outputs. A narrowly scoped model may be perfectly valid for a narrow decision, but it should not be represented as the total risk picture.

Quantitative analysis should be repeated when the decision or evidence changes

QRA is not a one-off gateway exercise. Rerun it when material risk responses are implemented, major assumptions change, the network changes significantly, the project enters a new phase, or new evidence alters important ranges. Preserve prior models so movement can be explained as real exposure change, changed assumptions or changed modelling method.

6. Practice diagnostic

Use this table to challenge the quality of the approach or deliverable. The rows consolidate the most important recurring red flags and checks; they are prompts for professional judgement rather than a substitute for the method above.

A good M4.3 output should make uncertainty more decision-useful, not merely more mathematical. If the analysis cannot explain what drives the result and which action changes it, the simulation has not yet become management insight.

7. Try-it-yourself practical exercise

Use the following simplified Reference Project position. The deterministic schedule meets the fixed opening date with six working days of margin. Total deterministic EAC is 49.4m CU against 50.0m CU authorisation; it is not an ETC. Assume 30.0m CU incurred cost and 19.4m CU deterministic ETC, including 0.80m CU for the uncommitted specialist-services component. Assume the source schedule and cost model have passed basic quality checks.

Complete the exercise in six parts:

·       1. Define the QRA decision question in one sentence and identify the two primary outputs management needs.

·       2. Classify each input as discrete event, range uncertainty or opportunity, and state where it should map into the schedule or cost model.

·       3. Identify at least two possible dependency or double-counting issues that should be checked before simulation.

For U-F, replace the 0.80m CU component already in ETC with the sampled specialist-services value; do not add the full sampled cost on top of 49.4m CU. Hold the 30.0m CU incurred cost fixed. State any separate risk allowances removed before event impacts are added, and price shared schedule-delay consequences once.

·       4. Build a simple probabilistic model using a spreadsheet, QRA tool or structured sampling approach. Report the probability of meeting the opening date and the probability of remaining within 50.0m CU. If software is unavailable, create at least 20 sampled scenarios and explain why the small sample is illustrative rather than statistically stable.

·       5. Identify the two or three dominant drivers and propose one evidence-based response for each. Rerun or reason through a post-response scenario using changed assumptions that are explicitly justified.

·       6. Write a management statement of no more than 150 words that distinguishes deterministic forecast, quantitative exposure, key drivers, model limitations and the recommended action.

Final challenge question: “What decision would change if the probability were 45%, 70% or 90% - and what evidence would justify changing the model inputs?”

8. Bridge to capability

The How-to teaches how structured uncertainty becomes quantitative decision evidence. At APPLIED and above, capability assessment should test whether the practitioner can select a proportionate method, prepare or challenge a realistic model and interpret its limits. Input bases, reproducible checks, confidence outputs and sensitivities should support the conclusion. Supporting work products are illustrative and should be assessed with the practitioner’s reasoning. Sample learning questions illustrate the judgement at each recognition level; they are discussion prompts rather than validated assessment items.

Learning question: Do you understand how a quantitative model expresses uncertainty?

Capability question for APPLIED and above: Can you demonstrate a reproducible analysis, defend its assumptions and use its outputs without false precision?