Evidence · Cost risk
Did Norway really fix cost overruns?
Since 2000, every large Norwegian state project has been budgeted from a P50 and a P85. We scored the results the way you would score any forecast, and set Denmark beside them
What did Norway change in 2000?
Since 2000, every large state investment project in Norway has had its cost estimate checked by outside consultants before Parliament decides. The estimate is probabilistic. Parliament approves a budget at about the P85, the cost with an 85% chance of not being exceeded, and the agency is given the P50 as its target.
The Ministry of Finance set the scheme up after the overruns of the 1980s and 1990s: a working group it led found that 8 of 11 large state projects it examined had gone over budget. It is now called the State Project Model and covers projects estimated at more than NOK 1 billion, or NOK 300 million for digital projects. Two reviews by independent consultants sit at its core. QA1 tests the choice of concept before design starts. QA2 reviews the cost estimate and the plan for running the project before the funding decision, and it sets the two numbers.
From idea to a budget Parliament approves
QA1: which project, if any
Outside consultants test the case for the project and the concepts considered, including doing nothing, before the government chooses one.
QA2: the cost estimate
Before the funding decision, a second review checks the cost estimate and the plan. The estimate is a distribution of total cost, not a single figure.
Two numbers come out of it
The P50, as likely to be beaten as exceeded, becomes the agency’s target budget. The total budget for Parliament is normally the P85, less a list of possible savings agreed in advance.
The reserve sits above the project
The gap between the two, typically about 10% of the expected cost, is held at a higher level and released only with the ministry’s approval.
The results are counted
The Concept research programme at NTNU, the Norwegian University of Science and Technology, follows the projects to completion and publishes how they did.
That last step is the unusual one. Many owners quote a P50 or a P80. Very few go back and count how often projects came in under it. Norway has been counting for about 25 years, and its researchers publish the results in enough detail to score.
Did Norway’s P85 budgets work?
Mostly, on cost. In the largest recent study, 73% of 111 completed projects finished within their approved budget, on average 5% under it. That contradicts the common belief that big public projects always overrun. But a P85 should hold about 85 times in 100, and Norway’s held about 73. Schedules did worse: projects finished ten months late on average.
A P85 is a forecast: it says that 85 of every 100 projects will cost less. So the fair test is not whether most projects stayed inside their budget. It is whether the share that did is close to 85%. The chart scores the results that Morten Welde and Atle Engebø published in 2024 for projects approved between 2001 and 2017.
How Norway’s estimates score
Share of completed projects that finished at or under each number. The vertical line is where a calibrated estimate would land; the shaded band is the spread it would show by chance on 111 projects, nine times in ten.
73% of 111 projects finished within budget and 42% of 106 at or under their P50 (Welde and Engebø, 2024). Bands drawn for 111 projects.
The P85 shortfall is too large to be chance. A calibrated P85 would have kept at least 88 of the 111 projects inside, and 81 made it. The P50 result is within chance. Magne Jørgensen, Morten Welde and Torleif Halkjelsvik reached the same verdict with formal tests on 69 projects that started between 2001 and 2012: a hit rate of 75% at the P85, significantly short of 85%, and 42% at the P50, which was not significantly short of 50%.
Read as a forecast, the estimates’ ranges are too narrow. Jørgensen and his colleagues found the contingency from the most likely cost to the P85 averaged 13%, half the 26% their calibrated benchmark called for. Under a simple lognormal reading of the two shares, the same estimates would have had to quote their P94 to give a true 85% chance of staying within budget.
There is a fair objection. Because the approved budget is the P85 less a list of possible savings, Welde and Ole Jonny Klakegg argue that it sits nearer the P80, so up to 20% of projects could be expected to overrun. Against that standard 73% still falls short, but only just: a calibrated P80 would have kept 82 to 96 of the 111 inside, and 81 made it.
Time is the weaker half. Welde and Engebø found the average project finished 10 months after its planned date, though the median was 2 months: 53% finished late and 10% more than three years late. Only 34% finished both within budget and without delay.
Was the State Project Model the reason?
Partly. Projects that went through the scheme have mostly finished within budget, which points to it. But four other things push the same way: estimates grow before the decision, so success is measured against a later number; a budget invites spending up to it; only cost was policed; and the results have drifted the wrong way over time.
The evidence for the model. James Odeck, Morten Welde and Gro Holst Volden compared 40 large Norwegian road projects. The 18 estimated before the scheme overran their budgets by 13% on average; the 22 that went through outside review came in 10.8% under. Across all sectors, the projects Welde and Engebø studied finished 4% below their combined budgets.
But it is a before-and-after comparison. Odeck and his colleagues add two cautions themselves. The estimates did not get more accurate: the average miss in either direction only fell from 13% to 11%, and they conclude the review led to systematic overestimation. And road projects above NOK 100 million improved over the same years, including those too small to be reviewed, from 9.4% over budget to 3.8% under, while smaller projects did not. The authors read that as the review’s lessons spilling over. It also means there is no clean comparison group.
The goalposts moved. A project is judged against the budget set at QA2, close to the funding decision. By then much of its growth has already happened. Welde and Odeck followed 42 road projects opened between 2000 and 2014. On average their final cost was only 1% from the approved P50, but 39% above the estimate in the national transport plan that first put them on the list. A project can finish within budget and still cost far more than the figure that got it started.
Money allowed is money spent. Welde and Klakegg name the risk themselves: a project manager has few reasons to finish under budget, so without active ownership “the project may use all or close to all the funds available (Parkinson’s law)”. It also means small underruns on most projects do not make up for large overruns on a few: in Welde and Engebø’s study the ten worst projects averaged 32% over. Holding the reserve above the project is Norway’s answer.
Only cost was policed. The scheme sets a budget and publishes how projects did against it. Time has no equal check, and time is where the results are weakest. Discipline went where it was measured.
The results drifted. Welde and Klakegg split their projects by when they were approved, 32 in each period. Of those approved in 2001 to 2005, 13% overran their budget. For 2006 to 2009 it was 34%, and for 2010 to 2014 again 34%.
Share of projects that overran their budget, by year approved
32 projects in each period. A P85 budget should be overrun by about 15% of projects; a P80, by about 20%.
Source: Welde and Klakegg (2024), Table V. Periods are the year of the investment decision.
Early budgets were too generous and later ones too thin. Jørgensen and his colleagues saw the same drift: their P85 hit rate fell from close to 100% for the earliest projects to about 60% for those starting in 2010 to 2012. A model that had fixed estimating for good would not drift like that.
The fair reading is that the scheme made estimates honest enough to be useful and created the habit of counting the results. It did not make them calibrated, and it does not keep them calibrated by itself. The counting is what shows that, and it is the part most organisations skip.
What did Denmark do instead?
Denmark tried a simpler fix. From 2007, large road projects were budgeted as a base estimate plus a flat 30% reserve. Overruns stopped, but the money went unused. The national audit office found the projects used on average about 67% of the funds allocated to them, and the ministry said future road projects would get a 15% reserve.
Rigsrevisionen’s report 3/2019 covered 32 road projects. In 25 the cost had been overestimated, by DKK 3.9 billion in total, mainly because contract prices kept coming in below the estimates. On top of that, DKK 7.7 billion of reserves went unused. The audit office judged this not satisfactory: the money was tied up instead of funding the next projects, and there were no rules for handing it back.
A flat uplift moves the budget, not the estimate. It cannot tell a cautious estimate from a risky one, so the error simply changes sign: Denmark swapped overruns for overfunding. Jørgensen and his colleagues tell of a government agency head who reported every project under a P85-based budget for five years and took it as a sign of good management. They read a 100% hit rate as a strong sign of overestimation.
Norway’s percentiles came out a little low; Denmark’s percentage came out a lot high. Which is worse depends on what an overrun costs you against what idle money costs you, the question our guide to choosing a contingency percentile works through.
How do you back-test your own P50 and P85?
Count. For each finished project, note whether its final cost came in at or under the P50 and the P85 quoted when the budget was approved. A calibrated P50 catches about half of them and a P85 about 85 in 100. The two counts show whether your estimates sit too low and whether their ranges are too narrow.
Use the estimate made at the decision, not one revised later, or you repeat the moving-goalposts problem. The calculator starts with Norway’s shares as if they came from 100 projects, with the P85 10% above the P50, about the size of Norway’s reserve. Put in your own.
Back-test your P50 and P85
Enter how many projects have finished, and how many of them came in at or under each of the two numbers quoted when their budgets were approved. You see how your estimates behave, and what you would need to quote to be as sure as you meant to be.
A calibrated P50 would have 42 to 58 of 100 projects under it, 9 times in 10. You had 42, which is in line. A calibrated P85 would have 79 to 91 of 100 projects under it, 9 times in 10. You had 73, so your P85s are set too low. The gap between your P50 and P85 is about 79% of what your results show, so your ranges are too narrow. To be 85% sure, quote the P94 of the same estimates.
ModelRisk builds the cost estimate as a Monte Carlo simulation inside Excel, so the P50 and the P85 come from one model of the project’s own costs, risks and correlations, and you can see what drives the gap between them.
Start the free trialThe calculator treats outcomes as lognormal around your estimates and reads where your two numbers really sit from the shares you enter. The band uses the binomial spread for the number of projects. The extra percentage needs the gap between your two numbers, as a share of the lower one; leave it empty to skip it.
How many projects do you need? More than most owners have at first. With 20 projects, a calibrated P85 would put anywhere from 14 to 19 of them under it, so only a large miss shows. With 100 the band narrows to 79 to 91. Start counting now, keep the estimate that was approved, and judge over several years.
Would Norway’s model work in your organisation?
The method travels; the setting may not. Any owner can have the estimate reviewed by someone with no stake in it before committing, quote two numbers, one to manage to and one to approve, and count the results afterwards. Norway also has a finance ministry that can stop projects, a steady flow of large ones and stable institutions, which are harder to copy.
Three parts carry over to a company or a smaller public body without new law:
- A review before the money is committed, by people who will not run the project. Norway’s reviewers are consultants on framework contracts with the Ministry of Finance. A company can use another business unit or an outside estimator.
- Two numbers, held by different people. The team manages to the P50. The reserve up to the higher number sits with whoever funds the portfolio, so it is released for real risks and not absorbed by default.
- A scorecard. Record the approved P50 and P85 for every project and the final cost. After a few years, the calculator above tells you how far to trust your own estimates.
What Norway shows is the limit of all three. Its estimates improved a great deal, then drifted low, and its schedules were barely touched. The front end, before anyone commits, is where most of the cost growth happens. Jonas Eliasson found the same in Sweden: estimates for 134 transport projects grew by 42% on average between the first plan and the decision to build, and by about 4% after it. A probabilistic estimate is only as good as the model behind it.
That model is what ModelRisk builds for the cost estimate in Excel; our guide to calculating cost contingency takes one from base estimate to P80 step by step. For the schedule, the half Norway did not fix, Tamara does the same for a Primavera P6 or Microsoft Project plan.
Frequently asked questions
What is a P85 cost estimate?
The cost that the estimate gives an 85% chance of not being exceeded. Funding a project at its P85 accepts about a 15% chance of an overrun, if the estimate is calibrated. In Norway, the total budget Parliament approves for a large state project is set at about the P85.
What is the difference between a steering frame and a cost frame?
They are common translations of two Norwegian budget terms; the Ministry of Finance’s own translation calls them the target budget and the total budget. The steering frame (styringsramme) is what the agency is expected to deliver for, normally the P50. The cost frame (kostnadsramme) is the ceiling Parliament approves, normally the P85 less a list of possible savings. The reserve between them is held above the project.
How many finished projects do you need to back-test your estimates?
More than most owners have at first. With 20 projects, a calibrated P85 would put anywhere from 14 to 19 of them under it, nine times in ten, so only a large miss shows. With 100 projects the range narrows to 79 to 91. Keep the approved estimate for every project and judge over several years.
Sources
- Volden, G. H. (2016). The Norwegian governance scheme for major public investment projects. Presentation slides, Reykjavik. PDF
- Odeck, J., Welde, M. and Volden, G. H. (2015). The impact of external quality assurance of costs estimates on cost overruns: empirical evidence from the Norwegian road sector. European Journal of Transport and Infrastructure Research 15(3), 286–303. doi:10.18757/ejtir.2015.15.3.3079
- Welde, M. and Odeck, J. (2017). Cost escalations in the front-end of projects: empirical evidence from Norwegian road projects. Transport Reviews 37(5), 612–630. doi:10.1080/01441647.2016.1278285
- Welde, M. and Klakegg, O. J. (2024). Avoiding cost overrun through stochastic cost estimation and external quality assurance. IEEE Transactions on Engineering Management 71, 1984–1997. doi:10.1109/TEM.2022.3173175
- Jørgensen, M., Welde, M. and Halkjelsvik, T. (2023). Evaluation of probabilistic project cost estimates. IEEE Transactions on Engineering Management 70(10), 3481–3496. doi:10.1109/TEM.2021.3067050
- Welde, M. and Engebø, A. (2024). Cost and schedule performance in large government projects. IOP Conference Series: Earth and Environmental Science 1389, 012027. doi:10.1088/1755-1315/1389/1/012027
- Norwegian Ministry of Finance (2025). Circular R-108/25: The State Project Model (English translation). PDF
- Rigsrevisionen (2019). Report on the Danish Ministry of Transport and Housing’s budgeting of road projects, report 3/2019 (English extract). PDF
- Eliasson, J. (2025). Cost overruns of infrastructure projects: distributions, causes and remedies. Transportation Research Part A 198, 104532. doi:10.1016/j.tra.2025.104532
The scores in the chart and the calculator are our own calculations from the published shares. Treating the log of final cost as normal around each estimate, the two shares fix where the quoted P50 and P85 really sit, and from that the percentile that would give a true 85%. The bands are the 5th to 95th percentiles of the binomial distribution for the number of projects.
Articles
Keep reading
ModelRisk
Adding risk and uncertainty to your Excel model
Build the cost estimate as a simulation, read the P50 and the P85 from one model, and see which risks and line items open the gap between them.