Reference Class Forecasting for Capital Projects Beyond the Spreadsheet
A project forecast can look precise and still be wrong.
In capital projects, the spreadsheet often has thousands of lines. The Level 4 schedule has logic links, calendars, productivity assumptions, procurement dates, risk allowances, and carefully coded work packages. The cost estimate may have quotes, norms, crew builds, indirects, escalation, contingency, and management reserve. On paper, it feels controlled.
Then the project enters execution. The brownfield tie-in takes longer than planned. Civil quantities grow. Fabrication productivity falls. Long-lead equipment slips. Commissioning uncovers interface gaps. The project team explains why this project is different.
Sometimes it is different. Often, it is different in the same way the last five projects were different.
That is the central value of reference class forecasting. It asks a disciplined question before leaders approve a baseline, sanction capital, or tell stakeholders the date is achievable:
What happened on a credible set of comparable projects, and why should this project perform materially better?
For oil and gas, EPC, construction, energy, mining, utilities, marine, infrastructure, and other complex capital projects, that question can protect decision quality. It does not replace detailed estimating or scheduling. It challenges them with evidence from actual outcomes.
Detailed Models Can Still Create False Confidence
Capital project organizations tend to trust the inside view. That means they build the forecast from the project outward. Teams break the scope into packages, estimate quantities, apply norms, sequence activities, assess risks, and defend assumptions.
That work matters. A project cannot be governed from a generic benchmark alone. The issue is that inside-view forecasts have predictable weaknesses.
Experienced teams recognize the pattern:
The estimate assumes mature scope while design is still moving.
The schedule assumes access, permits, engineering deliverables, vendor data, and site productivity will align.
Procurement dates use supplier promises without enough allowance for review cycles, logistics, customs, expediting, and late technical clarification.
Construction productivity reflects ideal crew output rather than what similar sites achieved under comparable constraints.
Contingency is shaped by affordability rather than risk exposure.
Commissioning and start-up receive optimistic durations because the mechanical completion date dominates the story.
None of this requires bad faith. PMI, APM, AACE International, RICS, CII, ICE, the World Bank, OECD, and public infrastructure bodies have all, in different ways, emphasized the same broad lesson: project outcomes depend not only on planning detail, but also on uncertainty, front-end quality, governance, risk, capability, market conditions, and behavioral bias.
The more detailed a model becomes, the easier it is to mistake precision for accuracy. A 22,000-line schedule can still be built on weak productivity assumptions. A Class 3-style estimate can still carry untested allowances. A Monte Carlo model can still produce misleading results if the inputs are narrow, correlated risks are ignored, or political pressure has already compressed the baseline.
This is where experienced executives should be cautious. Detail is useful when it reflects reality. Detail is dangerous when it hides fragile assumptions behind professional formatting.
Why The Inside View Dominates
Project teams rely heavily on internal models for understandable reasons.
They know the scope. They understand the contract strategy. They have walked the site, negotiated packages, met suppliers, and debated constructability. They can explain why the proposed offshore platform, transmission corridor, refinery turnaround, mine expansion, desalination plant, port upgrade, or metro package has special features.
The inside view also aligns with organizational incentives. A project seeking sanction must be fundable. A contractor bidding a competitive EPC job must be attractive. A business unit wants its opportunity approved. A public agency may face pressure to announce benefits and delivery dates. A weak-looking business case may not survive.
The result is often subtle optimism rather than deliberate manipulation. Schedule risk gets treated as manageable. Productivity losses get framed as recoverable. Interface uncertainty gets pushed into later phases. Contingency becomes a negotiation target.
Reference class forecasting introduces a counterweight. It brings the organization back to observed outcomes, not just team belief.
What Reference Class Forecasting Changes
Reference class forecasting was developed in the decision sciences and is strongly associated with work by Daniel Kahneman and Amos Tversky on the planning fallacy, and later with Bent Flyvbjerg’s work on major project forecasting. The core idea is simple: when predicting the outcome of a planned action, compare it with outcomes from a relevant class of similar past actions.
In project terms, it means looking outside the current estimate and schedule to answer questions such as:
What did comparable projects actually cost at completion?
How long did they actually take from sanction to mechanical completion, energization, first production, or beneficial use?
How far did productivity vary from baseline norms?
How often did procurement packages arrive late, need rework, or trigger claims?
What contingency was consumed?
Which phases suffered the most growth?
What happened between mechanical completion, commissioning, start-up, and stable operations?
This is the outside view. It does not ask, “What do we believe will happen if our plan works?” It asks, “What happened when other capable teams made similar plans under similar constraints?”
That shift matters because capital projects repeat patterns.
A brownfield refinery modification may involve unique tie-ins, but similar shutdown windows expose similar access, isolation, permit, and discovery risks. A utility substation program may vary by site, but design approvals, outage planning, equipment delivery, and civil works often show recurring behavior. A mining expansion may have site-specific geology, but earthworks, camp logistics, plant tie-ins, and pre-commissioning risks can be compared with prior expansions. An offshore module campaign may use a different yard, but engineering maturity, vendor data, weight growth, marine logistics, and hook-up productivity have histories.
Reference class forecasting does not claim every project is the same. It claims that the distribution of outcomes from comparable projects is often more informative than a single internally generated number.
The Practical Output Is A Range, Not A Slogan
Good reference class work does not produce a vague warning that “projects always overrun.” It produces a structured comparison.
For example, a project team may present a cost estimate with a base cost, contingency, and management reserve. The reference class review may show that comparable projects had a wide range of final outturn costs due to late design growth, ground conditions, industrial relations, marine spreads, or commissioning defects. The leadership discussion then changes.
Instead of asking whether the spreadsheet adds up, executives can ask:
Why is this project forecast near the favorable end of the historical range?
What conditions must be true for that outcome?
Are those conditions already secured, or merely planned?
Has the estimate included risks that affected prior projects?
If the project wants P80 confidence, what level of contingency or schedule allowance does the historical range imply?
The same logic applies to schedule. A deterministic program may show first gas, energization, passenger service, or ore throughput by a fixed date. A reference class can show whether similar projects reached that milestone within that timeframe, and what caused delays when they didn't.
This improves executive challenge because it makes optimism visible.
Building Reference Classes That Matter
The weakest criticism of reference class forecasting is also the most common: “Our project is unique.”
Complex projects are always unique in some respects. The question is not whether differences exist. The question is whether those differences matter to the forecast.
A meaningful reference class groups projects that share enough outcome drivers to support a useful comparison. The class does not need to be perfect. It needs to be more reliable than relying only on internal assumptions.
Select Comparators By Outcome Drivers
A useful reference class for a capital project usually considers several dimensions.
Comparator factor | Why it matters |
Asset type | LNG, offshore, transmission, rail, tunneling, mining process plants, water treatment, ports, and utilities have different risk profiles. |
Delivery model | EPC lump sum, reimbursable EPCM, alliance, design-build, owner-managed packages, and multiple prime contracts allocate risk differently. |
Project phase and maturity | A forecast at concept select should not be compared blindly with forecasts made after detailed design. |
Scope character | Brownfield work, live operations, tie-ins, shutdowns, utilities diversions, marine works, and underground conditions drive execution risk. |
Geography and market | Labor availability, logistics, permitting, customs, weather, local supply chain, and industrial relations affect cost and time. |
Scale and complexity | A small repeatable package and a megaproject with high interface density do not behave the same way. |
Technology and novelty | First-of-a-kind equipment, new process integration, or unfamiliar execution methods require different treatment. |
Contracting and procurement timing | Early commitments, long-lead items, vendor data quality, and package fragmentation influence delivery certainty. |
The goal is not to find a clone. It is to find a set of projects that faced similar forces.
For example, a brownfield offshore compression project should not benchmark only against generic topsides tonnage. It should consider hook-up density, SIMOPS restrictions, bed space, lifting windows, shutdown dependencies, workpack quality, and the extent of offshore commissioning. A metro station box should not be benchmarked only by structural volume. It should also consider utility diversions, ground risk, property interfaces, third-party approvals, and systems integration.
Use Historical Data Without Pretending It Is Pure
Historical project data is rarely clean. Cost breakdown structures differ. Schedules use different milestone definitions. Contingency may be reported inconsistently. Claims and scope changes may be treated differently. Productivity records may mix direct and indirect labor. Commissioning durations may be hidden inside broader completion periods.
That does not make the data useless. It means the organization needs data discipline.
A practical approach is to document:
What is included and excluded in each historical outcome
The maturity of the baseline at the point of comparison
Whether scope changes were owner-driven, market-driven, or a correction of underestimated scope
The milestone definitions used for schedule comparison
The normalization method for location, escalation, currency, and market conditions
Known data quality limits
The reasons for including or excluding each comparator
This is where cost engineers, planners, project controls teams, commercial managers, estimators, and delivery leads should work together. Reference class forecasting is not a data science exercise separated from project reality. It needs judgment, but the judgment must be explicit.
Compare Distributions, Not Anecdotes
One failed project is not a reference class. One success story is not proof of deliverability.
A good reference class looks at a set of outcomes and presents a distribution. That may include cost growth, schedule growth, productivity variance, contingency drawdown, procurement delay, or ramp-up duration. The team can then discuss P50 and P80 outcomes in commercial terms.
P50 is commonly understood as a median-style confidence point, subject to the method used. P80 indicates a higher confidence level, again depending on the analysis and inputs. The exact method matters, but the governance principle is straightforward: a project should not claim high confidence while using assumptions that sit near the optimistic edge of relevant history.
AACE International’s recommended practices, RICS guidance, CII research, professional planning bodies, and infrastructure assurance frameworks all support more disciplined treatment of estimate maturity, risk, contingency, benchmarking, and assurance. They may use different language and methods, but they align on one point: a credible forecast needs more than internal conviction.
Applying Outside-view Evidence Across The Delivery Chain
Reference class forecasting is often discussed in relation to capex approval, but its value is wider. It can challenge many parts of the project promise.
Capital Expenditure And Contingency
Cost estimates usually grow because of many forces acting together: immature scope, quantity growth, design development, claims, productivity loss, procurement escalation, late changes, access constraints, rework, and commissioning fixes.
A reference class helps leaders test whether contingency is sized against real exposure. If comparable EPC projects consumed most contingency before mechanical completion, the new project should explain why its allowance is lower. If prior infrastructure schemes saw repeated utility diversion growth, the estimate should not treat diversions as a minor provisional item. If past mining projects underestimated bulk earthworks, camp costs, or plant tie-ins, those areas deserve direct challenge.
This does not mean copying another project’s contingency percentage. It means using observed outcomes to test whether the project’s risk allowance is credible.
Schedule Duration And Float
Schedules often fail because they assume clean handoffs. Engineering releases design. Procurement places orders. Vendors submit data. Civil works start on time. Mechanical erection follows. Electrical and instrumentation teams get access. Pre-commissioning begins with completed systems.
Actual projects are messier.
Reference class evidence can test total duration and phase durations. It can show whether similar utility projects achieved planned energization dates, whether offshore campaigns met hook-up durations, whether tunneling projects recovered from early delays, or whether industrial plants reached stable operations soon after mechanical completion.
This matters because schedules can look logical but be commercially unrealistic. A critical path can be technically correct and still depend on productivity, approvals, access, and vendor performance that history does not support.
Construction Productivity
Productivity is one of the most sensitive assumptions in EPC and construction forecasting. Small errors in direct labor productivity can drive high cost and schedule effects.
Reference classes can compare planned to actual performance for:
Piping installation in congested modules
Cable pulling and terminations in late-stage industrial projects
Concrete placement in remote locations
Structural steel erection in constrained sites
Earthworks in mining or infrastructure corridors
Marine works exposed to weather windows
Brownfield demolition and reinstatement
The key is to compare the conditions, not just the work type. A piping inch-dia norm from a clean fabrication environment will not predict offshore hook-up performance. A trenching rate in open ground will not predict urban utilities work with unknown services and traffic constraints.
Procurement And Supply Chain Performance
Project teams often treat procurement as a schedule input rather than a forecast risk. Yet long-lead equipment, vendor engineering, inspection, expediting, shipping, customs, preservation, and site receipt regularly shape outcomes.
A reference class can track actual durations for transformers, compressors, subsea equipment, switchgear, turbines, rail systems, control systems, valves, and bulk materials. It can also compare vendor document cycle times, factory acceptance test issues, and late technical queries.
In a tight market, historical data must be adjusted. Old data from a loose supply market may understate current risk. Still, prior outcomes remain useful because they identify where procurement optimism tends to enter the baseline.
Commissioning, Start-up, And Operational Readiness
Commissioning is where many weak assumptions become visible.
Mechanical completion does not mean the asset is ready to operate. Systems need energization, flushing, drying, loop checks, functional testing, safety validation, control logic, operator training, spares, maintenance systems, procedures, regulatory acceptance, and punch-list closeout. Interdependent systems rarely complete in the neat sequence shown in early schedules.
Reference classes can compare the time between:
Mechanical completion and ready for start-up
Energization and beneficial use
First product and stable production
Handover and full operating capability
Substantial completion and defects closeout
This is especially important for utilities, process plants, offshore facilities, rail systems, ports, data centers, mining plants, and industrial assets where integration risk sits late in the project. A business case that treats start-up as a short tail may overstate revenue timing and understate working capital exposure.
Operational Readiness
Operational readiness is often forecast through checklists. The outside view asks whether similar assets actually entered service with trained staff, approved procedures, spare parts, maintenance plans, permits, digital systems, and emergency arrangements in place.
That can change the sanction conversation. If past projects reached physical completion before the operating organization was ready, the new project should include readiness as a critical workstream, not an assumption.
Make Reference Class Forecasting Part Of Sanction And Assurance
Reference class forecasting has the most value when it influences decisions before commitments become hard to reverse. That means using it during business-case development, option selection, front-end loading, project sanction, tender evaluation, baseline approval, and major reforecast events.
It should not appear as a late-stage audit paragraph after the preferred answer has already been chosen.
Reconcile The Inside View And The Outside View
The practical question is not which view wins automatically. The best governance reconciles both.
Inside-view forecast | Outside-view challenge |
Built from project-specific scope, quantities, schedule logic, and risk registers | Built from actual outcomes on comparable projects |
Shows how the team plans to deliver | Shows how similar plans performed in reality |
Can reflect current contracts, site conditions, and design maturity | Can expose optimism bias and missing uncertainty |
May become too narrow under pressure | May be too broad if comparators are poorly selected |
A strong assurance process asks the project to explain the gap. If the internal estimate is more optimistic than the reference class, the team should show credible reasons. Those reasons might include advanced design maturity, secured permits, completed enabling works, proven modular strategy, firm vendor slots, lower interface density, or a repeatable program delivery model.
Unsupported confidence should not be enough.
Use Independent Assurance At Key Gates
Independent assurance helps because project teams often cannot see their own optimism. An internal project controls group, external assurance team, owner’s engineer, cost consultant, peer review panel, or stage-gate board can test the forecast against comparable outcomes.
The review should examine:
Estimate maturity and basis of estimate
Schedule realism and critical path fragility
Quantity growth exposure
Productivity assumptions
Market and procurement assumptions
Contingency and management reserve
Interface and integration risks
Commissioning and start-up duration
Operational readiness assumptions
Differences between the proposed baseline and reference class outcomes
Government infrastructure bodies and multilateral lenders have increasingly emphasized business-case quality, appraisal, risk management, governance, and benefits realism. Professional bodies such as PMI, APM, ICE, RICS, AACE International, and CII also provide frameworks, practices, and research that support stronger front-end definition, controls, benchmarking, and risk-informed decision-making.
The language differs by sector. The commercial need is the same.
Build A Comparable-project Database That People Trust
Many organizations already have the raw material for reference class forecasting, but it is scattered across bid files, closeout reports, claims records, cost systems, planning tools, procurement logs, and commissioning databases.
A useful comparable-project database does not need to start as a perfect system. It can begin with a curated set of completed projects and a clear data model.
At minimum, capture:
Original sanctioned cost and final outturn cost
Original key milestones and actual dates
Basis and maturity of the approved baseline
Scope changes and their causes
Contracting model and risk allocation
Major procurement package performance
Productivity assumptions and actuals where available
Contingency drawdown
Claims and commercial settlements
Commissioning and start-up performance
Lessons that explain the numbers
The strongest databases include narrative context. Numbers without explanation can mislead. A cost growth figure means little unless the organization knows whether it came from owner scope change, underestimated quantities, poor productivity, late design, ground conditions, regulatory change, or contractor failure.
Use Ranges In Executive Decisions
Executives do not need false precision. They need decision-useful ranges.
A sanction paper that shows only one cost and one date invites weak challenge. A better paper shows the internal forecast, the reference class range, the confidence level, and the reasons for divergence.
For example, leadership might see:
The project team’s base estimate and schedule.
P50 and P80 cost and schedule outcomes based on quantitative risk analysis.
A benchmark range from comparable projects.
The main assumptions required to achieve the proposed target.
Specific mitigations already completed, not merely planned.
A clear statement of residual exposure.
This helps prevent strategic pressure from compressing the baseline. It also helps leaders decide whether to approve, defer, redesign, re-scope, change contracting strategy, add contingency, or demand more front-end work.
Good governance does not punish teams for showing uncertainty. It punishes the organization later if uncertainty is hidden.
Avoid Common Traps
Reference class forecasting fails when organizations use it superficially.
One trap is cherry-picking comparators that support the preferred answer. Another is mixing unrelated projects into one pool and calling it a benchmark. A third is treating historical performance as destiny, which can discourage improvement. The point is not to say, “Past projects overran, so this one must overrun.” The point is to ask, “What must be different, and can we prove it?”
Another trap is normalizing away every uncomfortable data point. Teams sometimes exclude troubled projects because they were “badly managed,” then keep the successful projects because they reflect what good management can achieve. That logic creates a biased sample. If comparable organizations repeatedly faced similar problems, those problems belong in the discussion.
A more disciplined approach separates the data into categories. For example, a utility program might compare greenfield substations, brownfield extensions, and live network cutovers separately. A mining company might distinguish process plant expansions from sustaining capital and tailings infrastructure. An oil and gas operator might separate offshore hook-up, subsea installation, onshore gas processing, and brownfield shutdown scopes.
Granularity improves the forecast when it reflects real performance drivers.
The Commercial Value Is Better Challenged
The strongest reason to use reference class forecasting is not academic. It is commercial.
Better forecasts improve capital allocation. They help owners decide which projects deserve funding, which need more definition, and which carry unpriced risk. They help contractors avoid bids that look profitable only because productivity and risk allowances are unrealistic. They help lenders, boards, partners, and public stakeholders understand exposure before they lock in commitments.
They also improve conversations between functions. Estimators, planners, engineers, procurement teams, construction managers, commissioning leads, finance teams, and executives can debate evidence rather than preferences.
That debate may be uncomfortable. It should be. A forecast that can't withstand comparison with relevant history isn't ready to carry major commitments.
A Practical Takeaway For Capital Project Leaders
Reference-class forecasting for capital projects beyond the spreadsheet isn't a rejection of detailed planning. It tests whether detailed planning has stayed connected to reality.
The spreadsheet still matters. The schedule still matters. The estimate basis, risk model, procurement plan, construction methodology, and commissioning strategy all matter. Yet you shouldn't accept any of them just because they are detailed, internally consistent, or presented with confidence.
A credible forecast should answer three questions before sanction:
What does the project team believe will happen?
What happened on comparable projects?
Why is any difference between the two justified by evidence already in hand?
That discipline helps expose optimism bias, unrealistic baselines, weak assumptions, understated contingency, and strategic pressure while decision-makers can still act. The last similar project may not predict the future perfectly, but it often knows more than the current spreadsheet admits.
Credible forecasting should be tested against what actually happened on comparable projects, not only what the current project team believes will happen.
