Anatomy of an Overrun

Step 1 of 5

The point estimate

Picture a road scheme on the day it is approved. Its business case carries one number for the final cost, and each downstream decision treats that number as settled. The figure shows the estimate as it appears in the plan, a single point at 1.0 on a scale of final cost as a multiple of the estimate.

Step 2 of 5

The reference class

No road is the first of its kind. The curve is the roadways cost class from the tables in Hubbard, Budzier & Leed (2025), How to Measure Anything in Project Management, Wiley, covering 1,263 road projects with known outcomes. A quarter of them finished at or under their approved figure, half ended at 1.13 times it or more, and one in twenty reached 1.95 times or beyond. Underruns are shallow and overruns are deep, which is why the curve leans right.

Try this yourself: Reference Class Forecasting in the Lab.

Step 3 of 5

Selection

When projects compete for funding, the ones that look cheapest tend to win. An estimate can look cheap because the project is cheap or because the estimate fell low. Because the approval process cannot tell those apart, funding the lowest numbers clips the favourable tail before construction starts. The funded portfolio then begins life tilted towards overrun even when each estimator was unbiased. My current research examines projects that are ranked on their estimated ratio of benefits to costs, where candidate projects differ in true quality as much as those on real shortlists do. For that case I find that random error in the estimates, combined with selection, accounts for roughly 25 to 50 percent of observed overruns in most infrastructure categories, with a central reading near a third. The range assumes estimates as accurate as is usual when funding is approved and a budget that covers about 30 percent of the estimated cost of all candidate projects.

Try this yourself: the Noise Audit in the Lab.

Step 4 of 5

The fat right tail

The right side of the curve is where budgets fail. In this class one road project in four ends beyond 1.32 times its estimate, and one in twenty ends beyond 1.95 times. A P90 budget is one that outcomes like these would have stayed inside nine times in ten, so choosing a percentile is choosing how often you are willing to go back for more money. That choice is invisible while the budget is a single number.

Try this yourself: the Monte Carlo Simulator in the Lab.

Step 5 of 5

What to do about it

The remedy begins with budgeting from the distribution rather than from the point. Reference class forecasting is one way to do that, since it reads the budget off a chosen percentile of comparable projects. Where funding was competitive, the observed class is itself a set of winners. Part of the uplift, the margin added to the original estimate to reach that percentile, therefore reflects selection among competing estimates. My current research suggests raising each estimate only by enough to correct its estimation bias, the systematic tendency of estimates to fall below final costs. The part of the uplift that reflects selection would then be held in a reserve for the funded portfolio as a whole. Which method improves estimating accuracy most is an empirical question that the same research examines. A further, less visible improvement is to measure, keeping records of estimates and outcomes so that next year's reference class includes your own projects.

Try this yourself: Selection-Adjusted RCF in the Lab.