Nobody ever simulated the money.

The base estimate is not a fact

Almost every cost risk model works the same way. Take the estimate, treat it as the base, simulate the risk register on top, add the two together.

That is wrong at the first step, and it is wrong in a direction that always flatters.

The $827,008 is made of three numbers. Measured work at $710,000, priced off a bill of quantities against rates that assume a particular gang and a particular sequence. Preliminaries at $85,200, built from a programme that assumes a particular duration. Escalation at $31,808, from a market forecast that is a forecast.

None of those is a fact. They are three point estimates, and by Week 9 we know what a point estimate is — one value lifted out of a range.

Give them honest ranges. Measured work between $688,700 and $773,900. Preliminaries between $80,940 and $100,536, because prelims are time-related and stretch with the programme. Escalation between $19,085 and $50,893.

Run just those three, with nothing from the register at all, and the mean comes out at $836,950.

The project is carrying $827,008. Before a single risk on the register has occurred, the base estimate is already $9,942 light on average, purely because a mode is not a mean and construction ranges lean one way.

The double count nobody catches

Now the error that runs the other way, and it is worth more than the first one.

The register has a line on it called steel escalation, ranged from nothing to $84,000. The base estimate has a line in it called escalation, priced at $31,808.

Those are the same money.

If you simulate both, the model pays for market movement twice. And the fix is not to average them or split the difference. You have to decide which document owns that risk. If the estimator has priced escalation, the register line is not escalation — it is escalation beyond what was priced, and it should be re-based to say so. Otherwise take it out of the register entirely and let the base range carry it.

We have done the second. Thirteen risks go into the model, not fourteen, and the escalation range in the base does the work.

This is worth checking on your own job before anything else. Preliminaries and time-related risk are the usual pair. So are the estimator's waste allowance and the register's material wastage line. A model that double counts is not conservative — it is a model nobody will believe when it matters, and the first person to spot the duplication gets to dismiss the whole thing.

What the job actually costs

Sixty thousand runs. Base ranges plus thirteen risks, correlated at the level we settled on last week.

The median outturn is $946,796.

The eightieth percentile is $1,010,175.

Read that second number against the contract. The client is paying $1,000,000. At P80 — the level most companies say they want to sit at — this job costs more than it sells for.

Some more of the curve, because the shape matters more than any single point. P10 is $875,300. P90 is $1,050,850. P95 is $1,087,378. The mean is $956,440 and the standard deviation is $69,650.

And two probabilities that are worth more than the percentiles.

The chance this job costs less than $1,000,000 is 76.5%. Which is another way of saying that roughly one job in four, run exactly like this one, loses money outright — not misses its target, loses money.

The chance the contingency is enough is 9.7%. Cost & Cash Week 5 put that figure at 49.5% and called it a coin flip. It was a coin flip against a five-line register. Against a proper model it is one in ten.

Why the base matters as much as the risks

Run the same model again, but this time do what everybody does — hold the base estimate at $827,008 and simulate only the register.

P50 comes out at $936,258 instead of $946,796. P80 at $994,081 instead of $1,010,175. P95 at $1,064,026 instead of $1,087,378.

So treating the estimate as certain understates the answer by $16,094 at P80 and $23,351 at P95. That is not enormous, and it is not meant to be. The point is the direction and where it comes from.

It comes from nowhere. No risk occurred. Nothing went wrong. That gap is entirely the cost of pretending that a number built from rates, allowances and a forecast is a number rather than a range.

The model is still optimistic

Here is the part that should stop you, and it is the reason this article does not end with a triumphant number.

Week 7 pulled twenty comparable jobs and found the median final cost was 1.11 times the tender. On this contract that is $1,110,000.

Our model — fourteen risks properly written, ranged, correlated, with base uncertainty included, sixty thousand runs — says the probability of exceeding $1,110,000 is 3.1%. Its ninety-fifth percentile is $1,087,378, which is $22,622 below what history says the median outcome is.

So the most careful thing we know how to build says the typical outcome on jobs like this is a one-in-thirty event.

Both cannot be right. And the honest reading is not that the reference class is wrong — it is twenty real jobs with real final accounts. It is that a bottom-up model can only contain the risks somebody wrote down, and Week 2 is still true: the register only ever holds what the reading list produced. Every risk nobody identified is missing from this curve, and the historical ratio contains all of them.

Practical insight

Before you build any cost model, do the two-column check. It takes an hour and it is the difference between a model people act on and a model people argue about.

In one column, list what the estimator has already allowed for. Waste percentages, weather allowances, priced escalation, plant standing time, the prelims duration. In the other, list your register.

Draw a line between every pair that touches the same money, and for each pair decide which document owns it. Write the decision down, because somebody will ask.

Then, whatever else you do, put a range on the base. Even a crude one — minus three percent, plus nine — is closer to the truth than a single figure, and it takes ten minutes with the estimator who built it.

A model without base uncertainty is not a cost risk analysis. It is a risk register with more decimal places.

Key takeaways

✔ The base estimate is three ranges, not three numbers. Their mean is $836,950 against $827,008 carried.
✔ Escalation sat in both the estimate and the register. Decide which document owns a risk, or the model pays twice.
✔ Median outturn $946,796. P80 $1,010,175 — at the eightieth percentile this job costs more than the client is paying.
✔ The chance of coming in under the contract value is 76.5%. About one job in four loses money outright.
✔ The contingency is enough 9.7% of the time. Cost & Cash called it a coin flip against a five-line register; against a real model it is one in ten.
✔ Holding the base certain understates P80 by $16,094 and P95 by $23,351, without a single risk occurring.
✔ The model puts the historical median outcome at a 3.1% probability. A bottom-up curve can only contain the risks somebody wrote down.

What's coming next

We have a curve. It runs from about $875,000 at the tenth percentile to $1,087,000 at the ninety-fifth, and every point on it is a legitimate answer to the question of what this job costs.

Which means the analysis is finished and the decision has not started.

Somebody has to pick a point on that line. Fifty percent is the number you would choose if you were happy to be wrong half the time. Eighty is the number most companies say out loud. Ninety-five is the number that would price you out of every tender you enter.

Next week we ask what P80 actually costs on this job, why nobody in the project controls department has the authority to choose it, and what it means that a company can hold a written risk appetite and still bid at P50 every single time.