Skip to content
Find your tool
Guides · 6 min read

Correlation: the setting that moves your P90

Leave correlation out and the S-curve still looks fine. It is just too narrow. In our example, the P90 reserve rises by about 70% once items move together.

By Mike Bouman, Quantil · November 4, 2026

Leave correlation out of a cost risk model and it will still produce a clean S-curve. It will just be too narrow, and the P80 and P90 will come out too low. This guide shows how large that effect is, and how to set correlation without pretending to precision you don't have.

Why independence makes the curve too narrow

In a simulation without correlation, every item draws its own luck. In one run the earthworks come in high and the steel comes in low, and the two cancel out. Across many items, those offsets add up to a total that barely moves. It's the same effect that makes a diversified investment portfolio more stable than a single share.

Real projects don't diversify like that. A rise in the steel price hits every steel-heavy item at once. Poor ground conditions hit the foundations, the excavation and the schedule together. When items move together, the bad runs stay bad, and the tail of the distribution gets longer.

The effect in numbers

To show it, we simulated a simple fictional estimate: eight cost items of $1.5M each, a base of $12.0M. Each item has the same range, from 10% below to 30% above its base. We ran it three times, each with 100,000 iterations and seed 42, changing only the correlation between the items.

Correlation between itemsMeanP50P80P90Chance that base + 10% holds
None (independent)$12.40M$12.39M$12.65M$12.79M99%
0.3, moderate$12.40M$12.37M$12.85M$13.11M92%
0.6, strong$12.40M$12.35M$12.99M$13.34M87%

Look at what changes, and what doesn’t.

The example is simplified on purpose: equal items and no risk events. On a real estimate the exact numbers differ, but the direction is always the same. In the fictional tunnel case of our QRCIM whitepaper, adding correlation barely moved the center of the distribution and clearly widened the tail.

How to set correlation without inventing precision

Nobody knows the correlation between two cost items to two decimals. You don't need to. What matters is the difference between "none" and "substantial". A coarse matrix with a handful of reasoned values is much closer to reality than an empty one. These are the guide values QRCIM shows on its correlation sheet:

SituationGuide value
Same material and same supplier0.7 to 0.9
Same material, different supplier0.5 to 0.7
Same weather or market exposure0.4 to 0.6
Same discipline with related scope0.3 to 0.5
Shared indirect costs0.2 to 0.4
Unrelated disciplines0.0 to 0.2

Work by group rather than cell by cell: give all items within one block, such as civil works, the same value, and use a lower value between blocks for general market exposure. Record the reason for each non-zero value in your assumptions log.

Don't overcorrect. Correlations that are too high inflate the tail just as an empty matrix shrinks it. If you're unsure about a pair, leave it at zero and write down why. And if a curve looks suspiciously narrow, check two things first: are the ranges wide enough, and has correlation been recorded?
A practical tip. In risk sessions, correlation is often the setting people most want to skip, because nobody can put a number on it. A simpler question helps: if this item goes badly, which others go badly with it? That conversation usually produces three or four groups, such as everything that depends on steel and cable prices, or on ground conditions. A moderate correlation within those groups is often enough to get a realistic tail.

Correlation in the Quantil tools

QRCIM lets you fill in a correlation matrix across up to fifty cost items and applies it with a Cholesky decomposition, the standard technique for drawing correlated values. The Cost & Schedule Risk Analyzer supports correlation between risks as well. The simulator on our homepage leaves it out on purpose, to keep the example simple. That's also why it's labeled simplified.