The Outside View: Reference-Class Forecasting for Large Capital Projects

1. When the Inside View Becomes the Only Story

On large capital projects, the forecasting room rarely feels empty. Schedules are built task by task. Cost models accumulate quantities, rates and contingencies. Risk registers lengthen. Sensitivity tables multiply. The pack looks thorough because it is full of the project’s own detail — objectives, interfaces, resources, obstacles, and a narrative of how the work will unfold. That is the inside view: a case-focused forecast constructed from the particulars of this investment.

The difficulty is not that detail is useless. Detail is necessary for design, procurement and control. The difficulty is that, left alone, the same detail systematically underweights what similar finished projects actually cost, took and delivered. Teams treat the venture as a unique one-off. Class outcomes from comparable past projects are ignored or waved away. Even cautious inside estimates tend to be optimistic. By the time overrun and delay are undeniable, identity and sunk spend have already fused around the early number.

This essay treats that pattern as a forecasting stance problem, not only as a modelling problem. Large capital projects need a disciplined outside view — reference-class forecasting — that places the present case on an empirical distribution of similar past outcomes before commitment hardens. The method does not replace bottom-up estimating. It disciplines the anchor against which bottom-up numbers are judged.

2. Roadmap

This comprehensive blog will explore how sponsors, investment committees, owner estimators, PMC leads and steering members can apply the outside view and reference-class forecasting on large capital projects. It will first separate the inside view from the outside view and show why uniqueness bias drops class data. It will then set out the three-step reference-class procedure that became the public core of the method, distinguish a later six-step elaboration used inside a wider governing sequence, and show how base-rate tables and fat-tail patterns discipline anchors without turning generic uplift percentages into a substitute for real class work. It will place optimism where it belongs — after a realistic commit decision — and mark carefully how industrial front-end loading and owner-shaping practice sit beside, not inside, the outside-view school. Application of executive-initiative outside-view evidence to capital-project gates is labelled transfer where the underlying studies were not themselves capital databases. Brief bridges keep this post distinct from companion ProTrain pieces on framing the decision question, smart teams, and bias across the project life.

3. Brief Bridges — What This Post Is Not

Three neighbours sit close enough to confuse. Keeping them distinct protects each method.

Decide the Question First. Framing Before Choosing owns and structures the live decision question — what is being decided, by whom, against which criteria — before options are ranked. Outside-view forecasting assumes a usable question exists. If the room still disagrees on what choice is live, class data will be pasted onto the wrong problem. The two stack; they do not replace each other. See Decide the Question First.

Smart Teams. Team hygiene, psychological safety and delivery-organism habits matter once a programme is under way. They do not substitute for an outside-view gate on cost, schedule and benefit promises. Estimate realism and team quality reinforce each other; merging them into one vague “culture” recipe blurs both toolkits.

Optimism and misrepresentation across gates. A companion bias strand separates honest delusion (planning fallacy, inside view) from strategic misrepresentation (rosy numbers that win the slot). Reference-class forecasting is one forecasting control aimed especially at honest inside-view error and at making deceptive anchors harder to defend. It is not the full lifecycle bias audit from business case through opening, and it does not by itself rewrite incentives in multi-firm chains.

Front-end loading and shaping. Industrial definition quality before authorization is a neighbouring practice from a different empiric tradition. It is treated later as adjacency, not as a rebrand of reference-class forecasting.

4. Inside View Versus Outside View

Two distinct forecasting modes can be applied to the same investment and produce very different answers.

The inside view is the default. Attention narrows to the case at hand: objectives, resources brought to the work, obstacles imagined, mental scenarios of progress, extrapolation of current trends. Consultants and owner teams often stay inside unless they are required to supply comparative data. Even when the tone is conservative, the structure of the exercise — weaving a story of this project — tends toward overconfidence about control, underweighting of combined failure paths, and neglect of how competitors and host environments have already shaped outcomes for similar ventures.

The outside view, also known as reference-class forecasting, changes the object of attention. Case detail is set aside long enough to examine a class of similar completed projects, to lay out a rough distribution of their outcomes, and to position the present project in that distribution. Competitor effects, regulatory friction and ordinary execution messiness are already embedded in class results. The outside view can miss extremes with no historical precedent, but for most large capital projects it is more likely to yield a realistic estimate when applied with equal skill.

A classic teaching vignette makes the gap concrete. A curriculum team’s inside estimates clustered between roughly a year and a half and two and a half years. When prompted for comparable programmes, the outside picture was stark: a large share never finished; those that finished took many years; the present team was not favourably distinguished. The team ignored the outside view, finished on a long outside-view timescale, and saw little uptake. The point is not the sector. The point is the stance: inside scenarios felt more informative than class base rates, and the class was right.

Figure 1: Inside view versus outside view — case scenarios and bottom-up detail on one side; reference-class distribution and placement of the present project on the other.

Organizational amplifiers make the inside view sticky. Optimistic proposal plans become anchors; contingency is treated as enough. Internal competition for scarce capital rewards rosy forecasts. Stretch goals and compensation targets punish pessimism as disloyalty. Group reinforcement seals the story. These are reasons the outside view must be institutionalized, not left to individual courage in a heated gate.

On large capital projects the inside view is reinforced by artefacts that look like rigor. A 400-line schedule is still a story about this sequence. A Monte Carlo model that samples only the team’s own uncertainty bands still treats the project as a closed world. A benefits register that extrapolates the sponsor’s preferred demand path still ignores how often similar benefits failed to appear. Outside-view discipline asks a prior question: among ventures that looked like this at decision, what actually happened?

The advantage of the outside view is strongest for de novo efforts — new technology plants, entirely new market entries, first-of-a-kind configurations — precisely where inside-view pressure feels like a moral imperative and where optimism biases are greatest. That is an uncomfortable finding for capital programmes that advertise uniqueness as a virtue in the funding narrative.

5. Uniqueness Bias and the Dropped Class

Uniqueness bias is the belief that this project is so singular that past projects do not apply. On large capital projects the belief is easy to defend: every alignment, geology package, offtake regime and political coalition really is particular. The forecasting error is to treat particularity as a licence to ignore distributional information.

The counter is blunt: the project is still one of those — one of a class (rail lines, process plants, data centres, hospitals, stadiums, IT transformations). Unknown unknowns that will later surprise the inside view are already visible, in aggregate, in the outcomes of completed peers. If the team believes the venture is unique, the outside view is dropped, and the planning fallacy returns under a more sophisticated spreadsheet.

Class design is itself a judgment. The class must be broad enough to be statistically meaningful and narrow enough to be comparable. When precedents are unclear, multiple analogous sets can be examined — for example the owner’s own prior new-technology investments and other owners’ similar-technology projects — rather than inventing a single perfect peer that does not exist. That judgment is still outside-view work. It is not a return to pure storytelling.

Figure 2: Uniqueness bias drops the reference class; the “one of those” move restores class data before commitment.

6. Reference-Class Forecasting — The Three-Step Core

The public three-step procedure for reference-class forecasting is straightforward enough to put in a gate checklist.

Step 1 — Identify a relevant reference class of past similar projects, broad enough to support statistics and narrow enough to compare.

Step 2 — Establish a probability distribution of outcomes for that class from credible empirical data — not from the team’s preferred scenario alone.

Step 3 — Position the specific project on that distribution to arrive at the most likely outcome, and to read risk of overrun or shortfall at chosen confidence levels.

Planning associations have publicly endorsed using reference-class forecasting in addition to traditional methods, and have warned against relying solely on conventional civil-engineering forecasts for non-routine projects. The method is presented as a way of bypassing both optimism bias and strategic misrepresentation by insisting on actual performance in a comparable class.

Worked transport uplifts published with that procedure illustrate the logic without becoming house constants for every asset class. In one reported transport set, willingness to accept a fifty percent risk of cost overrun implied a fifteen percent uplift for roads and a forty percent uplift for rail at decision-to-build, constant prices; willingness to accept only a ten percent risk implied much larger uplifts. An eighty-percent percentile example for a motorway business case used a thirty-two percent uplift. These figures are transport reference-class results from that body of work. They are not universal percentages for airports, IT programmes or process plants, and they are not the later multi-type base-rate appendix discussed below.

Reference-class forecasting does not forbid bottom-up work. It changes the order of authority. The class distribution is the anchor; the bottom-up model is evidence for or against a justified adjustment. Teams that reverse the order — build the detailed model first, then hunt for a class that flatters it — are performing the inside view with comparative decoration.

Risk appetite must be stated explicitly when uplifts or percentiles are used. A board willing to accept a one-in-two chance of overrun chooses a different point on the class distribution than a board that will accept only a one-in-ten chance. Hiding that choice inside a single “P50” label without showing the class is how optimism re-enters through vocabulary.

Figure 3: Three-step reference-class forecasting — define the class, build the empirical distribution, place this project.

7. Six Steps, Similarity Weighting, and the FOX Forecasting Stage

A later governance paper expands forecasting into a six-move reference-class sequence and situates it inside a three-stage governing process: Forecasting, Organizing and Executing. The six moves include defining a class of similar successes and failures, assembling the full historical distribution, making an intuitive placement, assessing the predictability of that placement, adjusting toward the class mean as predictability falls, and multiplying as needed for cost, time and benefits. Similarity-based forecasting is discussed as a sibling that weights past cases by resemblance rather than treating class members equally — with the important limit that humans still choose what enters the class.

This essay does not silently replace the three-step core with the six-step list. Both are in the knowledge base; they serve different levels of detail. For most board packs, the three-step core remains the operable checklist. The six-step elaboration is useful when a specialist forecasting cell must document placement judgment and predictability explicitly.

The wider governing sequence matters as adjacency, not as the spine of this post. Forecasting attacks honest inside-view error. Organizing attacks deception and agency problems in multi-tier principal–agent chains — incentives, risk sharing, independent review, premortems, collective forecasting routines. Executing attacks one-off uniqueness through modularity and experience. Those later stages deserve their own treatment. Here the claim is narrower: without an outside-view forecasting stage, organizing and executing inherit a fiction.

Worked comparisons in that governance paper — a tram programme whose claimed confidence interval was breached, a hydropower budget that moved sharply between appraisal and completion — are case facts, not ProTrain house constants. Currency, year and scope belong to the original cases.

8. Base Rates, Fat Tails, and Anchor Discipline

Large-project research later published multi-type base-rate tables for cost risk: many project types, many thousands of projects, with columns for mean cost overrun in real terms, the share of projects in a severe upper tail, and the mean overrun inside that tail. Illustrative extremes in that table structure include Olympic Games with very high mean overrun and a large upper-tail share; nuclear storage among the worst tails; and modular energy types such as solar and wind among the best-behaved, alongside roads and some thermal and transmission classes that are not fat-tailed in the same way. Information-technology tails can be extreme. The teaching use of such a table is not to frighten a board with a single horror number. It is to show that cost-risk shape differs by type, and that many types are fat-tailed: rare disasters pull the mean.

Process discipline follows. Anchor on the class mean of real outcomes. Adjust only with compelling, data-supported reasons. When in doubt, skip adjustment — the class mean is the forecast. That is the opposite of the usual habit: anchoring on the optimistic proposal and adjusting slightly toward realism.

Two separation rules protect integrity. First, do not merge an earlier eight-type academic base-rate table with the later twenty-five-type appendix; they are different publications with different samples. Second, do not merge Flyvbjerg-school base rates with Merrow / IPA outcome statistics. Both traditions are empirically serious; they are not interchangeable spreadsheets. Partial transcription and OCR limits in secondary notes mean mid-row figures should be verified against primary tables before they are put in a live investment paper.

Independent studies have claimed large accuracy gains for reference-class methods versus conventional forecasts. Those claims should be read as motivation to institutionalize class work, not as a promise that any percentage pasted into a cell is “RCF.”

Fat tails change portfolio logic as well as single-project logic. If a type is fat-tailed, a handful of disasters dominate average performance; “typical” experience understates what the next commitment is exposed to. Modular types that behave better in the tables invite a design question — what is the repeatable block? — without turning this essay into a modularity treatise. The forecasting point stands on its own: know whether the class you are joining is thin-tailed or fat-tailed before you treat the mean as a mild contingency problem.

Base-rate tables are evidence for cost risk in board papers, not scaremongering. Used well, they shorten arguments that otherwise recycle anecdotes. Used badly, they become a single cherry-picked horror slide. Show the row for the type, the sample idea, and where this project is placed — then stop theatrical selection of the worst cell on the page.

9. Optimism in Its Place — Goals Versus Forecasts

Optimism is not a character flaw to be purged from project life. Enthusiasm, resilience and motivation matter once a programme must be delivered. The error is to use the same optimistic story both to win the commit decision and to run the work.

The outside-view prescription separates the two. Use realistic, class-informed forecasts to decide whether to commit and on what terms. After commitment, aggressive goals can motivate implementers — provided the board and lenders still hold the outside-view number as the accountability baseline. Decision and support functions need realism; pure implementer morale may even be harmed by constant re-display of bleak base rates. At the top — the sponsor, unit head or project champion — both stances are required: optimism to lead, and outside-view realism when choosing among competing initiatives.

On large capital projects this separation is often inverted. Public promises and internal stretch targets are set from the inside view before class data are faced. Later “re-baselines” are treated as surprises rather than as the delayed arrival of the outside view. Think-slow sequencing helps: outside-view realism belongs in the slow gate; delivery energy belongs after the gate, not instead of it.

Figure 4: Optimism in its place — outside-view realism to decide whether to commit; goal-driven energy after commitment, with the class-informed baseline retained for accountability.

10. Institutionalizing the Outside View (and What Is Not RCF)

Individual analysts rarely defeat capital-slot competition alone. Reference-class work has to appear in the institution: mandatory-style requirements for major programmes before referendum or board lock-in; independent review of the class and the placement; board papers that show the distribution beside the bottom-up number; kill criteria set before identity and sunk spend fuse.

Government appraisal guidance that requires an explicit, empirically based optimism adjustment is a related but distinct instrument. Where better class data are absent, generic upper and lower bounds appear as a fallback — for example broad ranges for non-standard civil engineering capital. Those upper bounds are starting points, not a named reference-class forecast. Decorating a guess with a generic percentage and calling it reference-class forecasting is exactly the hygiene failure the method exists to prevent. If there is no class, say so. Then either build one, buy one, or admit that the estimate is not outside-view disciplined.

A practical bias audit across gates can ask, at each review: is the dominant failure mode delusion, deception, or not knowable yet? Put one outside-view comparison beside the bottom-up number. Treat repeated waiver of class requirements as a candidate normalization of deviance, not as a new standard.

Institutional design details that support honesty without pretending incentives disappear include independent double-blind peer review of forecasts, structured lists of reasons the forecast could be wrong (premortem discipline), shifts from purely individual to collective forecasting rounds, and publication of forecasts so reputation attaches to accuracy over time. None of these replaces a reference class; all of them make it harder to defend an undefended inside number.

For owner organisations that run repeated capital programmes, a standing reference-class library — curated classes, vintages, exclusion rules — beats one-off consultant packs reinvented at every gate. The library must be allowed to embarrass favourite projects. If class data are only commissioned when they are expected to comfort the room, the institution has not adopted the outside view; it has adopted outside-view theatre.

11. Front-End Neighbours — Merrow / IPA Adjacency (Do Not Merge Schools)

Industrial megaproject research in the Independent Project Analysis tradition emphasizes front-end loading: the completeness of scope and definition work before full-funds authorization, typically phased with gates that can proceed, recycle or stop. Best-practical definition is treated as a scored practice, not a slogan. Reported investment in that front end is material in capital and cycle time, and projects that miss adequate definition quality show sharply worse value outcomes in that database tradition. Owner shaping, integrated owner teams and resistance to pure schedule-driven authorization are part of the same neighbourhood.

This is complementary to outside-view forecasting, not identical with it. Both oppose premature lock-in. Front-end loading asks whether the project is defined enough to authorize. Reference-class forecasting asks whether the promised cost, time and benefits are realistic against class outcomes. A beautifully defined package can still carry an inside-view fantasy number; a class-corrected number can still sit on a poorly shaped project. Stack them. Do not rename one as the other, and do not paste IPA indices into Flyvbjerg base-rate tables or the reverse.

Construction Industry Institute language about front-end planning and team alignment during early phases appears in the wider corpus as related vocabulary for objective convergence. It is not treated here as a source of reference-class procedure. Where CII ideas appear in owner systems, they remain front-end practice neighbours — another reason to keep schools labelled.

12. Practice on Large Capital Projects

What should a capital-project gate pack contain if the outside view is real rather than decorative?

Name the class. State the reference class in plain language (for example “urban light-rail lines opened in the last twenty years with similar length and governance,” or “brownfield petrochemical expansions of comparable capacity”). If multiple classes were tested, show them.

Show the distribution. At least mean outcome, a sense of upper-tail risk, and the data vintage. Prefer primary tables over secondary paraphrases when numbers will drive money.

Place this project. Explain why the team believes the venture sits at the mean, better than the mean, or worse — with evidence, not brand aspiration. If predictability of that judgment is low, move toward the class mean.

State adjustments or skip them. Every adjustment away from the class mean needs a data-supported reason. “Our team is excellent” is not, by itself, such a reason.

Separate goals from forecasts. Record the outside-view baseline used for commit. Record delivery targets separately if they are more aggressive. Do not let the aggressive target overwrite the baseline in the lender pack.

Time the work before lock-in. Rail, dams, airports, hospitals, IT programmes and similar one-off public or quasi-public ventures need class work before referendum language, press dates or board irrevocability. After political commitment, escalating estimates are often theatre around a decision already made.

Where the underlying outside-view literature studied corporate initiatives — plant closures, market entry, mergers — and the move here is into large capital project gates, that application is transfer: reasoned extension of a shared forecasting stance, not a claim that every cited failure rate was measured on megaprojects. Where megaproject databases and transport class studies are the source, say so and keep their numbers in their lanes.

A short steering-committee script helps. First: name whether the live fight is about the decision question, about class realism, or about delivery incentives. Second: demand the class slide before the detailed WBS walkthrough. Third: ask what would have to be true for this project to sit in the best quartile of the class — and whether those conditions are evidenced or hoped. Fourth: record the outside-view baseline in the minutes with the same formality as the approved budget line.

Transfer labelling matters for intellectual honesty. Executive-initiative evidence on manufacturing plants, mergers and market entry informs the stance (inside versus outside) and the organisational amplifiers of optimism. Megaproject and transport databases inform type-specific cost-risk shapes. Blending them into one undifferentiated “industry average overrun” is how careful empirics become folklore. Keep the lanes, label the transfer, and let the board see which claim rests on which kind of evidence.

Figure 5: Stack map — owned decision question; outside-view / reference-class forecasting; bias and incentive hygiene across gates; FEL / shaping as front-end neighbour. Stack the layers; do not merge the schools.

13. Closing — One Outside-View Comparison Beside the Bottom-Up Number

Large capital projects fail forecasts for many reasons. Some are political. Some are deceptive. Many begin as ordinary inside-view optimism dressed as professionalism: a thorough case story that never consulted the class.

The remedy advocated here is deliberately narrow and deliberately institutional. Switch forecasting stance before commitment. Build or buy a reference class. Show the distribution. Place the project. Adjust only with evidence. Keep optimism for delivery energy after the gate, not as the gate’s arithmetic. Treat generic uplift tables as fallbacks when class data are absent, not as a costume for RCF. Keep industrial front-end loading and owner shaping as neighbours that improve definition readiness — complementary, separately labelled, never merged into one spreadsheet religion.

Put one outside-view comparison beside the bottom-up number. If there is no class, say so. That single hygiene move, repeated at every gate until it becomes boring, is how the outside view stops being a seminar topic and becomes how large capital projects are approved.

References

1. Bent Flyvbjerg with Dan Gardner, How Big Things Get Done: The Surprising Factors Behind Every Successful Project, from Home Renovations to Space Exploration (Currency / Random House, 2023).

2. Dan Lovallo and Daniel Kahneman, “Delusions of Success: How Optimism Undermines Executives’ Decisions,” Harvard Business Review, July 2003, reprint R0307D.

3. Bent Flyvbjerg, “From Nobel Prize to Project Management: Getting Risks Right,” Project Management Journal 37, no. 3 (2006): 5–15. Open text arXiv:1302.3642.

4. Dan Lovallo, Matteo Cristofaro, and Bent Flyvbjerg, “Governing Large Projects: A Three-Stage Process to Get It Right,” Academy of Management Perspectives 37, no. 2 (2023): 1–19, https://doi.org/10.5465/amp.2021.0129; SSRN abstract 4526068.

5. HM Treasury, Green Book supplementary guidance on optimism bias (published 21 April 2013.

6. Edward W. Merrow, Industrial Megaprojects: Concepts, Strategies, and Practices for Success, 2nd ed. (Wiley, copyright 2024).

7. Bent Flyvbjerg, “Top Ten Behavioral Biases in Project Management: An Overview,” Project Management Journal (2021).

8. Bent Flyvbjerg, Massimo Garbuio, and Dan Lovallo, “Delusion and Deception in Large Infrastructure Projects: Two Models for Explaining and Preventing Executive Disaster,” California Management Review (2009).