Across the cost reference classes, the median project ends about 1.1 times its estimate
The median roadway ends at 1.13 times its estimate, and the median IT project lands on target at 1.00. The tails carry the risk. The distribution has a heavy right tail, and a few projects run past three and a half times.
Source: the reference classes in Hubbard, Budzier & Leed (2025), How to Measure Anything in Project Management, Wiley.
Does not show: a single overrun rate for all projects. Medians differ by project type and by the stage at which the estimate is made.
UK rail projects need a 60% cost uplift to be 80% confident of no overrun
At the Full Business Case stage, the recommended uplift on rail capital costs is 60% at 80% confidence and 19% at the median. The analysis behind the guidance draws on a database including 2,522 new-build rail, road, bridge and tunnel projects.
Source: Budzier & Leed (2021), Updating the Evidence Behind the Optimism Bias Uplifts for Transport Appraisals, Oxford Global Projects for the UK Department for Transport.
Does not show: one uplift for everything. Figures vary by project type and rise sharply at earlier stages, to 121% for rail at Strategic Outline.
Clean-energy projects run 84% over budget on average
Across a dataset of more than 1,100 clean-energy and decarbonisation projects, the mean capital-cost overrun is 84% and schedules run 51% long. The cost figures rest on the 482 projects in the dataset with cost data. For projects completed since 2000 the mean cost overrun is 36%.
Source: Greig, Techau, Leed & Budzier (2025), Maintaining capital discipline in the rush to net-zero energy systems, iScience 28(4), 112241.
Does not show: that each technology performs alike. Excluding nuclear, the full-sample cost overrun falls to 17%, and solar PV lands closest to estimate.
More than 200 capital projects monitored by Hong Kong's AI early-warning system
The Project Supervision System, which I helped build, flags public-works projects drifting towards cost overrun or delay early enough to act. It was trained on 849 completed projects spanning HKD 460 billion of outturn cost.
Source: Flyvbjerg, Budzier, Agard & Leed (2022), AI in Action, Hong Kong Development Bureau (PSGO). Portfolio size per the Hong Kong Development Bureau.
80% of the Hong Kong system's cost forecasts land within ±10%
The current models reach 91% accuracy on cost overruns and 88% on schedule overruns, using how a project is spending in flight rather than plans alone.
Source: AI-PSS-3, Oxford Global Projects for the Hong Kong Development Bureau, presented at Machine Learning Week Europe 2024.
Does not show: general forecasting accuracy. The figures are cross-validated performance on Hong Kong's own portfolio.
Selection can account for roughly 25 to 50% of observed cost overruns
Across most infrastructure categories, ranking projects on noisy estimates and funding the best-looking ones can produce overruns in the funded projects even when no single estimate is biased. The central reading is near a third.
Source: Leed (2026), Noise, Selection, and the Illusion of Bias, SSRN Working Paper No. 6429358. The version on SSRN is an earlier draft.
Does not show: that optimism bias is gone. Selection explains only part of observed overruns. For dams, the category with the largest overruns, the central estimate of the share explained by selection is about 12%. Where funding decisions are not purely benefit-cost driven the effect is smaller.
One bioscore point of nearby ecological quality is associated with about 2.8 percent higher house prices
The bioscore, from the national Biodiversitetskort map, is an index of ecological quality scored 0 to 19 and averaged within 1 kilometre of each home. When homes of the same type within the same parish are compared across around 850,000 Danish sales, one additional point is associated with about 2.8 percent higher prices, or DKK 57,500 at the median sale price of DKK 2 million. Homes of the same type in the same parish typically differ by about half a point. One standard deviation of that within-parish variation, 0.56 points, is associated with about 1.6 percent higher prices, or DKK 32,000. Without any adjustment, one point is associated with about 10 percent higher prices across the country. Around seventy percent of that raw association disappears once other housing and location characteristics are held constant and homes are compared in this way. I read that reduction mostly as a matter of who chooses to live near nature.
Source: My current research at Aarhus University, a working paper in progress titled The Price of Biodiversity: Hedonic Evidence from Danish Housing Markets.
Does not show: a causal effect or a price for any particular place. These are preliminary associations from an unpublished paper and may change before release.
Nuclear waste storage needed a 202% cost uplift to be 80% sure
A reference-class forecast over 216 comparable past projects found that the lowest official Swiss contingency gave its estimate barely 30% certainty. At the median the required uplift is 67%.
Source: Budzier, Flyvbjerg, Garavaglia & Leed (2018), Quantitative Cost and Schedule Risk Analysis of Nuclear Waste Storage, SSRN Working Paper No. 3303410.
Does not show: a forecast for other countries' programmes. It is a Swiss case, and the reference class pools nuclear and underground mining projects.
Cost overruns analysed across 88 stadium projects
The analysis behind Danish coverage of stadium economics, including the planned Aarhus stadium.
Source: coverage in Jyllands-Posten (November 2024) and TV2 Østjylland (December 2024), both linked from the Speaking page.
Does not show: a headline overrun percentage. Ask me for the underlying figures before quoting a magnitude.
