Monte Carlo schedule risk analysis (and what a P80 date is)

A deterministic schedule assigns one duration to every task and reports one finish date. But every task duration is really a range, and the project finish is the sum of dependent ranges along the longest path. Because delays accumulate down a chain while early finishes rarely get passed forward, the deterministic date is almost always optimistic — it behaves like a best case, not a most-likely case.

Worse, parallel paths create merge bias: when two or more chains feed one milestone, the milestone waits for the slowest, so the more paths you merge, the more likely at least one is late. A deterministic plan ignores this entirely. Monte Carlo simulation captures both effects, which is why its P50 is frequently later than the deterministic date.

Monte Carlo needs a duration range for each task, not a point. The standard input is a three-point estimate: an optimistic (best-case) duration, a most-likely duration, and a pessimistic (worst-case) duration. These three numbers define a distribution the simulation samples from.

The most common choice is the PERT-Beta distribution. It is bounded by the optimistic and pessimistic values, peaks at the most-likely value, and is skewed — which matches how real tasks behave, since they can run far over but only modestly under. Triangular and uniform distributions are simpler alternatives; PERT-Beta is preferred because its smoother shape and tunable skew better reflect estimation reality.

Task-duration uncertainty is only half the picture. The other half is discrete risk events: a CMC vendor slips, an assay fails and has to be repeated, a long-lead part is back-ordered, a gate review demands extra work. Each such risk has a probability of occurring and an impact (extra days, or a multiplier on affected tasks).

On every iteration the simulation rolls each risk: with probability p the event fires and its impact is applied to the relevant tasks; otherwise it does not. Across thousands of iterations a 30%-likely, 20-day risk shows up in roughly 30% of runs, lengthening the tail of the distribution exactly as much as it should. This is what separates true schedule risk analysis from simply sampling task ranges.

Each iteration is one complete, internally consistent version of the project: every task gets a sampled duration, every risk is resolved, and the critical path is recomputed for that specific scenario. The result is a single finish date for that run. Repeat thousands of times and you accumulate a histogram of finish dates — the project's outcome distribution.

The histogram is usually summarized as an S-curve (cumulative probability). The S-curve reads directly: pick any date on the x-axis and the curve tells you the probability of finishing on or before it. Pick any confidence level on the y-axis and it tells you the date that delivers it.

A P-value is a percentile of the finish-date distribution. The P50 is the date by which the project finishes in 50% of simulations — a true coin-flip median, and a far more honest internal target than the deterministic date. The P80 is the date hit in 80% of runs: late enough to absorb most realistic variation, which is why it is the common choice for external commitments and board dates. The P90 is more conservative still, reserved for high-stakes regulatory or contractual deadlines.

The gap between P50 and P80 is the cost of confidence — the extra calendar time you are buying to raise your odds from a coin flip to a near-certainty. That gap is also the rigorous basis for sizing a project buffer: rather than guessing, you set the buffer to the P50-to-P80 spread the simulation actually produced.

In a deterministic plan the critical path is fixed. In simulation it is not: because durations vary run to run, different paths become critical in different iterations. The criticality index of a task is the percentage of iterations in which it lands on the critical path. A task that is critical in 95% of runs deserves your attention far more than one that is critical in 10%.

This is one of the most useful outputs of an SRA. A deterministic Gantt can hide a near-critical path that is critical most of the time once uncertainty is modeled — and the criticality index surfaces exactly those tasks the static plan underrates.

A tornado chart ranks tasks and risk events by how strongly each one drives the variation in the finish date, measured by the correlation between an input's sampled duration and the project outcome across all iterations. Bars are sorted longest-to-shortest, producing the tornado shape.

The value is focus. Of hundreds of tasks, typically a handful dominate the spread of outcomes. Tightening estimates, adding mitigation, or buying down risk on those few items moves the P80 more than touching everything else combined — so the tornado chart tells you where management attention pays off.

Monte Carlo output feeds directly into Critical Chain buffer sizing. Instead of the rule-of-thumb 50%-of-removed-safety method, you size the project buffer from the simulated distribution: the buffer equals the gap between your aggressive median (P50) plan and the confidence level you are committing to (commonly P80). Feeding buffers are sized the same way from the distributions of their supporting chains.

Buffers derived this way are defensible. They reflect the actual modeled uncertainty of your specific program — its merge points, its risk register, its task ranges — rather than a generic percentage, which is what AACE RP 132R-23 Level 4 risk-driven scheduling expects.

CritPath AI runs 10,000-plus vectorized iterations on a NumPy/SciPy engine, so a full schedule simulation completes in seconds rather than the minutes a row-by-row desktop tool takes. It samples PERT-Beta (or triangular) task durations, injects discrete risk events by probability and impact, optionally applies duration correlation between related tasks, and runs a forward pass across the network each iteration to derive each task's criticality index — how often it finishes at and drives the project end (a simplified, finish-to-start-focused determination rather than a full per-iteration CPM honoring every dependency type).

The outputs are first-class objects in the product: the P50/P80/P90 S-curve, per-task criticality index, a tornado sensitivity chart, and buffer sizes fed straight into Critical Chain and Drum-Buffer-Rope. A Claude + Gemini copilot, grounded in your real dependency graph, explains in plain language which tasks dominate the tornado and what a change does to your P80. It is $10 per user per month, with AI usage billed separately by metered usage — versus legacy schedule-risk tools at roughly $10K-plus per seat or quote-only.