Simulation is usually explained in the abstract: distributions, iterations, a curve at the end. The practical question is narrower and harder — how do you range the resources and quantities in the estimate you have already priced, so that the numbers coming out mean something? That starts with the model, not the software that samples it.
A deterministic estimate produces one number and then adds one contingency percentage on top. Both are point judgements, and neither answers the question a funding body is actually asking: how confident can we be that this project will not exceed this budget? Monte Carlo quantitative risk analysis answers it directly by describing every uncertain input as a probability distribution rather than a single value, then running the cost model thousands of times, drawing a fresh value from each distribution on every pass.
After several thousand passes you have a complete picture of where total cost is likely to land, and contingency at P50, P75 or P90 is read off that picture rather than guessed. This is why the major Australian frameworks prefer it above their value thresholds. TMR calls probabilistic analysis the preferred method and a flat percentage “not recommended… potentially very inaccurate”; the Commonwealth’s DITRDCA guidance makes Monte Carlo mandatory above $25M out-turn; RES generally recommends simulation above $10M.
None of that machinery helps if the estimate underneath is a schedule of lump sums. The simulation samples the model you built, which means the discipline lives in the base estimate and the ranging, not in the run.
Resources, quantities, productivity and margins have to be separable before anything can be meaningfully ranged. TX1:Trinity builds estimates that way by default, and exports the whole model as a portable package for whoever runs the analysis.
Start a free 14-day trialThe method is conceptually simple and the simulation itself is almost trivial to run. What separates a model that survives review from one that does not is the work done before the button is pressed: a base estimate with real structure, and ranges that reflect what an experienced person actually believes rather than a reflexive plus or minus ten per cent.
One iteration is one plausible version of the project. Each pass draws a random value from every input distribution and totals the model, so the result is a single outcome that could genuinely have happened. Do that ten thousand times and the random noise cancels; what remains is the shape of the uncertainty you described.
A first-principles, bottom-up estimate of defined scope, built from quantities and rates, excluding contingency and escalation. This is the spine the whole analysis hangs on.
Single values become distributions. Inherent uncertainty is ranged on base lines or on the risk factors that drive them; discrete contingent risks come from the register as probability times impact.
Sample every distribution, total the model, repeat 5,000 to 10,000 times. Correlation is applied here, so inputs that share a cause move together rather than independently.
Collate every iteration into a histogram and a cumulative S-curve. Contingency is the distance between P50 and the chosen high-confidence percentile, usually P90.
An input distribution is just a statement about how a value might vary. The Australian frameworks converge on a small, practical set, and all of the continuous ones are defined by a three-point estimate: an optimistic value, a most-likely value and a pessimistic one. RES cautions that where objective data is absent, expert bounds are best treated as an 80% confidence interval — the 10th and 90th percentiles, adjusted for skew — because people anchor on the likely case and understate the extremes.
| Distribution | Defined by | Where it fits | Watch for |
|---|---|---|---|
| Triangular | Minimum, most likely, maximum. | The workhorse for continuous quantities and rates; intuitive for a subject-matter expert to populate. | Straight sides give it slightly heavier shoulders than reality. |
| PERT / BetaPert | Minimum, most likely, maximum, weighted towards the mode. | Where the most-likely value is genuinely well understood and deserves more weight. | RES recommends the “Alt” forms, because experts rarely capture true best and worst cases. |
| Trigen | Percentile bounds, typically P10 and P90, rather than absolutes. | Where ranges were elicited as percentiles; it truncates the triangular tails. | State which percentiles the bounds represent, or the truncation is meaningless. |
| Uniform | Minimum and maximum only, no mode. | A value known only to lie within a band, with no preferred figure inside it. | Rarely true in cost work; usually a sign the range was never really thought about. |
| Discrete — Bernoulli, Binomial | Probability of occurrence and a cost impact. | Contingent risk events from the register: a latent-condition claim, a heritage find, a regulatory change. | Model rare, high-impact events separately rather than burying them in general contingency. |
Iteration count buys precision, not accuracy. More runs make the simulation’s own output stable from one run to the next. They do nothing for ranges that are too tight or correlation that was never modelled — a model fed poor inputs returns a beautifully smooth S-curve that is wrong.
A simulation has no access to the project. It has access to your model, and it will faithfully sample whatever structure that model exposes. A line priced as a single lump sum offers exactly one thing to vary: the lump sum. You can put a triangular distribution on it, and the result will be a distribution shaped entirely by a number you made up, dressed in the authority of ten thousand iterations.
The same line built from resources offers something else entirely. Crew size, productivity, plant hours, material quantity and supply rate are each separately uncertain, each traceable to a different cause, and several of them are shared with other items in the estimate. That is what makes a risk-factor model possible: one driver — a labour market, a ground condition, a design maturity — can move every item that depends on it, and the correlation that dominates the answer arises from the structure instead of being asserted in a matrix afterwards.
So the honest sequence is: build the estimate so it has joints, then range the joints. An estimate assembled from priced resources is rangeable by anyone who understands the work. An estimate assembled from remembered totals is rangeable only by its author, and only in the sense that they can invent bounds for it.
If two costs move together — both driven by the same labour market, say — treating them as independent makes the total look far less variable than it really is. Within each iteration, independent highs and lows partly cancel, narrowing the total distribution and pulling the P90 down. Real cost items are rarely independent: a hot market lifts steel, concrete and labour at once, and wet weather that delays earthworks delays everything behind it.
RES warns that ignoring correlation can understate total-level variance considerably, and that its effect is more profound than the choice between different probability distributions. DITRDCA states the cardinal rule: no iteration may produce a combination of values that could not possibly occur. Correlation is modelled two ways. Functional correlation is built into the model’s own mathematics, which is what the risk-factor method gives you free — one driver multiplies many lines. Applied correlation is analyst-specified, with coefficients from −1 to +1 linking inputs that share a cause.
Line-item ranging places a distribution on each estimate line, which is intuitive and is what TMR describes for modelling inherent risk on base-estimate lines. The Commonwealth warns that structural difficulties inherent in line-item ranging make it difficult to arrive at realistic contingency assessments, because over-disaggregation washes out variance unless correlation is handled carefully, and it prefers the risk-factor method with models kept to roughly 20 to 40 inputs. In practice the two combine: range the factors that drive many items, and range individually only the handful of lines with genuinely idiosyncratic uncertainty.
Commonwealth guidance quotes a margin of error of about ±1.4% at 5,000 iterations, the practical floor for a credible model. TMR’s rule of thumb is 10,000, tightening the margin to roughly ±1.0%, and that is the comfortable default for funding-gate work. Beyond about 10,000 the curve barely changes; run more only if the tail is still visibly noisy, and spend the effort you saved on input quality instead.
The S-curve is the cumulative distribution, with probability-not-exceeded on the vertical and cost on the horizontal — read across from a confidence level to find the budget. The histogram is the probability density, the familiar right-skewed shape, showing where outcomes cluster and how far the upper tail runs. The tornado chart ranks inputs by how much each drives total-cost variance, longest bar at the top, and it is the most actionable output of the four because it converts one contingency figure into a list of things worth doing: more geotechnical investigation, a fixed-rate supply agreement, an earlier design freeze. Underneath them sit the headline percentiles — P50 for budget-setting, P90 for approval, P75 where a framework sets a tender-award delivery value.
Keep the two kinds of uncertainty apart. Inherent uncertainty is continuous and belongs on the estimate: the quantity might be higher, the rate might be dearer, the crew might be slower. Contingent risk is discrete and belongs in a dollarised register built from a facilitated workshop: an event that either happens or does not, with a probability and an impact. Blending them is how projects end up double-counting the same exposure, and how a reviewer ends up unable to tell which risks the contingency was actually sized for.
Be clear about the boundary first: TX1:Trinity does not ship a Monte Carlo engine. The simulation is run alongside it in a dedicated quantitative risk tool, and the Australian practice these pages draw on names @RISK as the usual one. Claiming otherwise would be the sort of overreach a reviewer notices immediately, and it would also miss the more interesting point, which is that the hard part of a defensible simulation is never the sampling.
What TX1:Trinity produces is a model that can be ranged. Rates are built from resources coded by type — labour, plant, material, subcontract, other, and composite groups — so the inputs a risk-factor model wants are already identified and already shared across the items that use them. Built-up resources recompute as the sum of their contributing rows, so a change to one component cascades to the package rate, to every item using it and to all allocations, which is functional correlation made visible: you can see, before any simulation, exactly which items a labour-rate movement touches. Quantities stay separate from rates and both are addressable through the formula engine’s project parametrics and line references, so the two kinds of uncertainty can be ranged independently rather than smeared together.
For what-if work there are allocation scenarios — named sets of pricing assumptions you can switch between and compare without disturbing the baseline estimate. They are useful for exactly the questions a tornado chart raises: what happens to the total if the plant fleet changes, if overheads are recovered differently, if the shift pattern moves. They are deterministic pricing sets, not a probabilistic engine, and it is worth saying so plainly — scenarios answer “what if this”, while a simulation answers “how likely is that”. When the analysis does need to leave the building, the portable, checksum-validated .tx1 package carries the whole project to whoever is running it, and back again with the numbers that came out. For how that handover fits the wider sequence, see the estimating process.
Between 5,000 and 10,000 for funding-gate work. Commonwealth guidance quotes a margin of error around plus or minus 1.4 per cent at 5,000 runs, which is the practical floor for a credible model. TMR's rule of thumb is 10,000, tightening the margin to roughly plus or minus 1.0 per cent. Past 10,000 the curve barely moves, so run more only if the tail of the S-curve is still visibly noisy.
Less than most people expect. Commonwealth guidance states that the choice between Beta PERT, triangular and Trigen usually has negligible impact on the result, and RES goes further, noting that the effect of excluding correlation is more profound than the choice between different probability distributions. Spend the effort on defining sensible ranges and modelling correlation properly, not on arguing about shape.
Correlation is the degree to which inputs move together. When inputs are treated as independent, their highs and lows partly cancel within each iteration, which narrows the total distribution and pulls the P90 down. Real cost items are rarely independent, because a hot market lifts steel, concrete and labour at once. DITRDCA puts the rule plainly: no iteration may produce a combination of values that could not possibly occur.
Line-item ranging is the more intuitive approach and the one TMR describes for modelling inherent risk directly on base-estimate lines. Commonwealth guidance prefers the risk-factor method, where a smaller set of drivers such as ground conditions or market rates multiplies the items each one affects, because correlation then arises structurally rather than being bolted on. It also recommends keeping models to roughly 20 to 40 inputs to avoid over-disaggregation washing out variance.
No. TX1:Trinity is a first-principles estimating application, and the simulation is run alongside it in a dedicated quantitative risk tool, with @RISK the one most commonly named in Australian practice. What TX1:Trinity contributes is the model the simulation needs: quantities separable from rates, rates decomposed into coded resources, margins held as distinct layers, and allocation scenarios for what-if pricing. Its allocation scenarios are named sets of pricing assumptions, not a probabilistic engine, and they should not be described as one.
Range resources, not recollections. Price from first principles, keep the joints visible, and hand the analyst a model instead of a total.