Reference Class Forecasting: How to Correct Optimism Bias on Capital Projects
Reference class forecasting corrects optimism bias in capital cost estimates using outside-view data. How to build a defensible class and apply the uplift.
This article covers why bottom-up estimating systematically under-forecasts, what reference class forecasting does mechanically and how it differs from benchmarking, how to define and normalize a class you can defend at a stage gate, how North American funding agencies have already institutionalized the outside view, and the four ways the method fails in practice. The payoff is a second, independently derived number you can put beside your base estimate — and a documented argument for the gap between them.
Why the Inside View Under-Forecasts
A bottom-up estimate is an inside view by construction. It decomposes scope, prices each element, and sums. That method is excellent at what it can see and structurally blind to what it cannot: scope that has not been imagined yet, interfaces that have not been drawn, and the accumulated drag of a hundred small productivity assumptions that each look reasonable in isolation and are collectively optimistic.
The blindness is not a competence problem. It is a property of decomposition. When a cost engineer prices a concrete pour, the estimate reflects the pour as designed at that moment. It cannot reflect the redesign that permitting will force in eighteen months, because that redesign is not yet scope. Add a risk register and a contingency analysis and the picture improves, but only for risks somebody thought to write down. The residual — the risks nobody listed — is exactly the category that historical overrun data captures and the inside view does not.
AACE International concedes the point in its own structure. RP 119R-21, Cost Estimate Accuracy Range and Contingency Determination Using Tables Derived from Parametric Risk Models (rev. November 28, 2022), derives accuracy ranges and contingency from parametric risk models in which the dominant drivers are the level of scope definition — expressed through the AACE estimate classes — and a technology and complexity rating. Neither driver is a line item. Both are outside-view attributes of the project as a whole, quantified from empirical data across many projects. The most widely used contingency reference in the discipline is, in its mechanics, already a reference class method.
What Reference Class Forecasting Actually Does
The method has three steps and one non-negotiable rule.
Identify the class. Assemble a set of completed projects genuinely comparable to yours — same asset type, similar scale, similar delivery model, similar market.
Establish the distribution of forecast error. For each project in the class, compute the ratio of final outturn cost to the estimate approved at the same decision gate you are now at. The output is a distribution of how wrong estimates of this type were, not a distribution of what projects of this type cost.
Locate your project and apply the uplift. Choose the confidence level the funding decision requires, read the corresponding percentile off the class distribution, and apply it to your base estimate.
The non-negotiable rule is in step two. A reference class is a distribution of forecast error, not a distribution of cost. This is what separates reference class forecasting from benchmarking, and it is the distinction practitioners most often collapse. Normalizing historical cost data to compare $/lane-mile or $/installed kW tells you whether your estimate is plausible against peers. It says nothing about whether estimates of this type are biased, because the peer costs and your estimate are both point values. Only the ratio of outturn to gate-stage estimate exposes the bias.
A reference class is a distribution of forecast error, not a distribution of cost — the moment you forget that, you are benchmarking again and calling it something else.
The gate anchoring matters as much as the class definition. An estimate approved at Class 5 and an estimate approved at Class 3 have entirely different error distributions, and mixing them produces an uplift that is wrong for every project it is applied to. If the class data does not record which gate each estimate was struck at, the class is not usable.
Building a Reference Class You Can Defend
Class definition is where the method is won or lost, and it is governed by a genuine tension: a tighter class is more relevant and has fewer members; a looser class has statistical body and less relevance. There is no formula. There are defensible rules.
Match the decision gate, always. Non-negotiable, as above.
Match the delivery model. Design-bid-build, progressive design-build and alliance contracts do not share a forecast error distribution. Contract strategy changes who absorbs scope growth, which changes what shows up as owner cost overrun.
Match the market and the regime. A class assembled from projects sanctioned before 2021 encodes an escalation environment that no longer exists. Normalize every historical outturn to a common date using Statistics Canada’s Building Construction Price Index for Canadian work or the relevant BLS producer price indexes for US work, and state the base period explicitly.
Be honest about sample size. Below roughly twenty projects the tail of the distribution is unknowable, and P80 read off eight data points is a number with a false precision problem. Say so in the estimate basis rather than reporting the percentile as though it were measured.
The data itself is more available in North America than most teams assume. Your own portfolio closeouts are the first source and the most relevant, if capital closeout data was ever captured in a comparable structure — most organizations discover at this point that it was not, which is itself a finding worth reporting. Beyond that: FTA Before-and-After studies for transit, state DOT bid tabulations and engineer’s estimate records for highway work, and, for Canadian federal projects, the costing and project-brief structures set out in the Treasury Board of Canada Secretariat’s Guide to Costing and the Mandatory Procedures for Project Complexity and Risk Assessments under the Directive on the Management of Projects and Programmes.
Where Reference Class Forecasting Breaks Down
The method has four failure modes, and three of them are self-inflicted.
Double-counting the uplift. This is the most damaging and the most common. Reference class forecasting quantifies systemic risk — the residual, unlistable uncertainty that historical error distributions capture. So does a parametric contingency derived from scope definition and complexity. Applying an RCF uplift on top of a 119R-21 contingency counts the same uncertainty twice and produces a number the project will never spend. The uplift substitutes for the systemic component; it stacks only on top of quantified project-specific risk, and only if that risk is genuinely specific.
The flattering class. Whoever defines the class controls the answer. Widen it to include the well-run projects and the uplift falls; the choice is defensible either way and almost never documented. Class definition must be recorded in the estimate basis and reviewed by someone with no stake in the sanction decision.
Regime change. A class is a claim that the past is informative about the future. Supply-chain conditions, labor availability and escalation regimes shift, and a class assembled across a boundary averages two different worlds. Date-normalization helps with prices; it does not help with structural change in how projects get delivered.
Strategic misrepresentation. Where estimates are optimistic because optimism gets projects approved, reference class forecasting constrains the behavior without curing it — the class selection simply becomes the new place to be optimistic. This is a governance problem with a governance fix, not a modeling problem. The academic literature has stayed active on exactly this point, with critical reassessments of the method’s promises and limits appearing in Production Planning & Control through 2025 and 2026.
Running It Alongside the Bottom-Up Estimate
Reference class forecasting is not a replacement for a detailed estimate and it does not produce a number you can build a control budget from. It produces a challenge number. Use it that way.
At each stage gate, run the class distribution against the current base estimate and put the two results side by side in the cold-eyes review. If the bottom-up estimate carries 12% contingency and the class distribution puts P65 at 27% above the stripped base, that gap is the deliverable. It is not a number to quietly adopt, and it is not a number to argue away. It is a documented question the project has to answer: which specific features of this project justify performing better than the class, and is that argument written down anywhere a decision-maker can test it?
Record the class definition, the sample size, the normalization basis, the gate the historical estimates were struck at, and the reconciliation to the base estimate in the estimate basis memorandum. A reference class forecast that lives in a spreadsheet on one analyst’s drive is not governance. A reference class forecast that a sanction committee has seen, argued with, and either adopted or explicitly overruled is.
Key Takeaways
- Bottom-up estimating is an inside view and is structurally blind to unlisted scope and accumulated optimism, which is precisely the uncertainty that historical overrun data measures.
- Reference class forecasting builds a distribution of forecast error — outturn cost divided by the estimate approved at the same decision gate — and is not the same thing as benchmarking unit costs against peers.
- A defensible class matches the decision gate, the delivery model and the market regime, normalizes every outturn to a common base period, and states its sample size honestly rather than reporting a false-precision percentile.
- North American practice has already institutionalized the outside view: FTA Oversight Procedure 40 strips contingency, applies beta range factors from historic FTA projects, and funds at P65.
- The method’s dominant failure mode is double-counting — an RCF uplift substitutes for parametric systemic contingency rather than stacking on top of it.
- Treat the result as a challenge number that forces a written argument at the gate, not as a control budget.