Your simulation assumed the bad days don't cluster
Track 1 built the machine. Week 15 of Schedule ran the project twenty thousand times, produced a distribution of finish dates, and showed you how to read a contingency off the curve instead of guessing at one.
It was correct. It also had an assumption inside it that nobody mentioned, and that assumption is the largest single error in most quantitative risk analysis anywhere.
When the machine sampled the gang rate for March, and then sampled crane availability for March, and then sampled the weather for March, it drew each one from its own curve, separately, as though the three had never met.
They meet every day. They meet in the same February.
One February, five register lines
Think about what three wet weeks in February actually do to this job.
The gang output falls, because you cannot tie steel efficiently in standing water. The crane stands idle, because you are not lifting in that wind. The raft pour gets postponed, because the ready-mix will not come out to a flooded formation. The groundwater rises, which is the risk we already have on the register from the August standpipe reading. And every follow-on trade slips, because the substructure is late.
That is five lines on the register. It is one event.
A simulation that samples those five independently is not modelling this project. It is modelling five different Februaries happening to five different sites, and then adding up the bills.
The formal name for it is a common cause. In practice it is the reason experienced people distrust risk models without being able to say why — the output feels too tidy, and it is too tidy, in a specific and measurable way.
What independence does to the answer
Here is the mechanism, and it is worth understanding because it is not intuitive.
When you add up a lot of independent things, the extremes cancel. For the total to come out very high, nearly every single item has to land high at once, and if they are independent that is astronomically unlikely. So the sum clusters tightly around its middle. Add fourteen independent risks and the total is far more predictable than any one of them.
That is a real mathematical result and it is why insurance works. It is also completely wrong for a construction project, because a construction project is not fourteen unrelated bets. It is one job, run by one team, in one winter, buying from one market.
So let's run it both ways and look.
Take the fourteen risks with the ranges from last week. Sample them independently, sixty thousand times, and the portfolio comes out at a median of $142,677, a P80 of $182,158 and a P95 of $224,325.
Now change one thing. Let a single driver — call it how the job is going — influence all fourteen, so that a bad month is bad across the board and a good one is good across the board. Same ranges. Same probabilities. Same everything else.
Median $135,643. P80 $196,451. P95 $269,563.
The mean does not move
Look at those two sets of numbers again, because the interesting thing is not where they differ. It is where they don't.
The mean is $145,183 in the independent run and $144,902 in the correlated one. Correlation did not change the expected value by a dollar, and it never does. If your risk report shows a single number and that number is an expected value, correlation is completely invisible to you.
The median actually went down, from $142,677 to $135,643. That is not a mistake. When everything moves together you get more months where the whole job runs well, as well as more where it doesn't. The middle hollows out.
What moves is the far end. P80 up eight percent. P95 up twenty percent. The standard deviation up by nearly half.
And the far end is the only part anybody actually buys. Contingency is not held to cover the average outcome — by definition you would run out half the time. It is held to cover the eightieth or ninetieth percentile. Correlation does not change what you expect. It changes what you should be afraid of, and that is precisely the part you are paying for.
One more number makes it concrete. Take the five largest risks and ask how often all five land together. Independent: 4.5% of runs. Correlated: 9.3%. The nightmare month is twice as likely as the model said.
The other error, in the other direction
There is an opposite mistake, and it is just as common because it looks conservative.
Work out the P80 of every risk on its own, then add up the column. On this register that gives $246,123.
That number is not cautious. It is wrong. Adding individual P80s assumes perfect correlation — that on the day the rock is at its eightieth percentile, the steel market and the client's wayleave and the crane and the supplier's solvency are all simultaneously at theirs. Nothing on earth is that badly organised.
The truth sits between the two, and the whole point of running a correlated simulation is that it puts you there rather than at either extreme.
Where this already happened to us
Cost & Cash Week 6 found $61,868 of overrun by month six and gave it two causes: the gang was slower than the rate assumed, and the steel came in dearer than the estimate.
Read that again with this week in mind. Those were not two problems.
The slow gang did not slow one activity. It slowed every activity it touched — pile caps, ground beams, the raft — and each of those carries its own cost code. So one cause arrived in the cost report as four or five separate small variances, none of them alarming on its own, spread across a page.
That is the same error running the other way. In the model, correlated causes get split into independent lines and the tail disappears. In the report, correlated effects get split across cost codes and the pattern disappears. Both times, one thing wearing several names.
Practical insight
You do not need software for this and you do not need to estimate a correlation matrix, which is a thing people are asked to do and cannot honestly do.
Take your register and write, next to each line, the one thing that would make it worse. Not the cause you already wrote in Week 4 — the background condition. Weather. Gang. Design flow. Market. Client responsiveness.
Now group the lines by that word.
Any group with three or more lines in it is a common cause, and those lines cannot be treated as independent. If you are running a simulation, correlate them. If you are not, do something simpler and more useful: take the largest group, assume every risk in it happens in the same month, add them up, and ask whether the project could absorb that.
That single sum is worth more than a correlation matrix, and you can do it on paper in twenty minutes.
Key takeaways
✔ Three wet weeks in February is one event that appears on the register as five separate lines.
✔ Independent sampling models five different Februaries happening to five different sites.
✔ Independent: median $142,677, P80 $182,158, P95 $224,325. Correlated: $135,643, $196,451, $269,563.
✔ The mean does not move at all, so any report built on expected values is blind to correlation.
✔ The median falls while the tail grows — more good months and more disasters, fewer ordinary ones.
✔ All five of the largest risks landing together goes from 4.5% of runs to 9.3%.
✔ Adding each risk's own P80 gives $246,123 and assumes perfect correlation. It is not conservative, it is wrong.
What's coming next
We now have fourteen risks, each with a shape, correlated by the things that actually drive them, producing a curve for the project as a whole.
And that curve is about cost, which is the thing Track 1 never did. Schedule Week 15 simulated the dates. Nobody has ever simulated the money on this job.
Next week we point the machine at the $827,008 we control, and get the first defensible answer to the question this whole track opened with: on a job priced at a million dollars, what is the number this project should actually be holding?
Enjoyed this lesson?
Join with Google to get each new lesson the moment it's published — and help me see which topics matter most to you. No spam, one email a week, unsubscribe anytime.
Already following on LinkedIn works too — this is just for the weekly email.