Reference class forecasting answers a single question: when projects like this one were attempted, what actually happened? Instead of building a bottom-up estimate from tasks and assumptions — the inside view — you find a class of similar past projects and use the distribution of their outcomes as your forecast. The method was developed by Daniel Kahneman and Amos Tversky and operationalised by Bent Flyvbjerg for megaprojects. The core claim: the inside view is systematically optimistic; the outside view corrects by anchoring to base rates.
A software team estimates a new platform will ship in six months. They list features, assign days per feature, add buffer. That is the inside view. Reference class forecasting asks: what fraction of similar platform projects at similar companies shipped within 2× their initial estimate? The answer is usually a majority. The forecast should therefore be twelve months or the distribution of that reference class, not six. The same logic applies to cost. Infrastructure projects, product launches, and M&A integrations have documented track records. Those track records are the reference class.
Flyvbjerg's work on transport and construction showed that cost overruns and schedule slips are the norm, not the exception. Reference class forecasting does not eliminate overruns; it produces forecasts that reflect how often overruns occur. The decision-maker gets a distribution: 50th percentile, 80th percentile, worst case. Planning then uses the percentile that matches the organisation's risk tolerance. Governments and enterprises that adopt reference class forecasting often set budgets at the 80th percentile of the reference class. Projects still overrun, but less often than when forecasts are anchored to the inside view.
The practical requirement is a well-defined reference class. "Similar" must be specified: same sector, same type of project, same scale band, same technology maturity. A first-of-its-kind project has a thin or contested reference class; the forecast carries more uncertainty. The discipline is explicit: name the reference class, cite the data, report the distribution. When no reference class exists, the forecast should say so. That transparency is the point — not false precision, but an honest outside-view anchor.
Section 2
How to See It
Reference class forecasting appears wherever forecasts are explicitly or implicitly anchored to the distribution of past similar cases. Look for use of base rates, benchmarks from comparable projects, or formal reference-class databases. The absence of it is also a signal: bottom-up plans with no comparison to historical outcomes are relying on the inside view.
Business
You're seeing Reference Class Forecasting when a PMO mandates that every capital project must show its estimate against a reference class of similar projects (sector, size, type). The budget is set at the 80th percentile of that distribution. Proposals that cannot cite a reference class must flag the gap. The organisation has shifted from "we think it will cost X" to "projects like this typically cost between A and B; we are planning for the B end."
Technology
You're seeing Reference Class Forecasting when an engineering lead refuses to give a single ship date and instead says: "In our reference class of similar platform builds, 50% shipped within 1.5× initial estimate and 80% within 2×. Our inside-view estimate is Q2; the reference class suggests we should plan for Q3 and have a contingency for Q4." The forecast is the distribution, not the point estimate.
Investing
You're seeing Reference Class Forecasting when an investor evaluates a founder's revenue forecast by comparing it to the distribution of outcomes for similar companies at the same stage and in the same category. "Companies in this reference class had a median time to $10M ARR of 4 years; you're projecting 2. The burden is on you to show why you're an outlier." The outside view disciplines the inside view.
Markets
You're seeing Reference Class Forecasting when a policy unit models infrastructure cost using Flyvbjerg-style reference classes (e.g. rail, roads, IT) and publishes confidence intervals. The headline number is not the engineer's estimate but the median or 80th percentile of the reference class. Political and commercial decisions then use a range that reflects historical reality.
Section 3
How to Use It
Decision filter
"Before locking a forecast for cost, schedule, or outcome, identify a reference class of similar past cases. What was the distribution of results? Anchor your forecast to that distribution — at least the median, ideally a chosen percentile — and treat the inside view as a source of variance, not the primary estimate."
As a founder
Use reference classes when you set roadmaps, raise rounds, or commit to customers. Your team's inside view will underestimate time and overestimate certainty. Find a reference class: similar products, similar stage, similar team size. Use the distribution (e.g. "similar B2B SaaS at our stage hit $1M ARR in 18–36 months") to set internal and external expectations. When investors ask for a single number, give the range from the reference class and explain the anchor. The discipline reduces surprise and builds credibility when you beat the median.
As an investor
Test founder forecasts against reference classes. When a founder says "we'll hit $10M ARR in 24 months," ask: what fraction of comparable companies did that? If the reference class says 20%, the forecast is a tail outcome. That does not make it wrong, but it shifts the burden of proof. Use reference classes to set milestones and to spot plans that are systematically optimistic. The best founders already use an outside view; the rest need the discipline imposed.
As a decision-maker
Before approving a budget or a timeline, require a reference class. What similar initiatives existed? What were their outcomes? If no reference class is available, the forecast is high uncertainty by definition. Plan for that: wider bands, earlier checkpoints, exit criteria. Avoid the trap of approving the inside view because it is detailed. Detail does not fix optimism; the outside view does.
Common misapplication: Using a reference class that is not similar enough. "Other software projects" is too broad; "other greenfield B2B platforms at seed stage" is closer. The class must be defined so that the cases are exchangeable for forecasting purposes. Wrong reference classes produce misleading base rates.
Second misapplication: Dismissing the reference class because "we're different." Sometimes the difference is material; often it is overconfidence. The burden of proof should be on the party claiming outlier status. Document why this case is not in the reference class before discarding the base rate.
Bezos insisted on thinking in terms of what was possible at scale over long time horizons, but Amazon also used reference-class discipline for execution. The company studied comparable e-commerce and infra buildouts to set expectations for fulfilment centres, AWS capacity, and new verticals. Bezos's "disagree and commit" culture included requiring that big bets show how they compared to similar past initiatives — an institutional outside view.
Hastings has emphasised learning from the distribution of outcomes in content and product bets. Netflix uses data on how similar shows and features performed to set expectations and kill underperformers. The reference class is explicit: "shows in this genre, this budget band, this audience" — and forecasts for spend and engagement are anchored to that class rather than to the creative team's inside view alone.
Section 6
Visual Explanation
Reference class forecasting: the inside view (your plan) usually sits at the optimistic left of the distribution. The outside view anchors the forecast to the reference class — e.g. median or 80th percentile — reducing systematic optimism.
Section 7
Connected Models
Reference class forecasting sits between probabilistic thinking, planning bias correction, and scenario-based decision-making. These models either explain why the inside view fails, provide the statistical frame, or extend the approach to full scenario sets.
Reinforces
Planning Fallacy
The planning fallacy is the tendency to underestimate time, cost, and risk for one's own projects. Reference class forecasting is the corrective: use the distribution of similar past projects as the anchor. The two are paired — the fallacy describes the bias; reference class forecasting is the method to counter it.
Reinforces
Regression to the Mean
Regression to the mean says extreme outcomes tend to be followed by outcomes closer to the average. Reference class forecasting uses the mean (and full distribution) of a reference class as that average. Both push forecasts toward base rates and away from overconfidence in exceptionalism.
Leads-to
Scenario Analysis
Scenario analysis explores multiple futures (best, base, worst). Reference class forecasting provides the base and the range: the percentiles of the reference class are natural scenario anchors. Combine both: use the reference class for the distribution, then scenario analysis for narrative and drivers.
Leads-to
Pre-Mortem Analysis
Pre-mortems ask why a project might fail. Reference class forecasting tells you how often similar projects overrun or fail. Use the reference class to set the baseline failure rate and the pre-mortem to identify project-specific causes. The pre-mortem enriches the outside view with inside-view mechanisms.
Section 8
One Key Quote
"Overestimation of benefits and underestimation of costs are the rule rather than the exception. Reference class forecasting is a method to correct for this by basing forecasts on the actual performance of a reference class of comparable projects."
— Bent Flyvbjerg
The inside view is the default; it feels more relevant because it is specific to your project. The outside view feels abstract. The correction is to make the reference class explicit and to anchor the forecast to it. The quote captures the empirical reality — systematic optimism — and the fix: anchor to comparable cases.
Section 9
Analyst's Take
Faster Than Normal — Editorial View
Most forecasts are inside-view point estimates. Teams list tasks, add buffer, and call it a plan. Reference class forecasting forces a different question: what happened when others did similar things? That question is uncomfortable because the answer often implies longer timelines and higher costs. The payoff is fewer surprises and better resource allocation.
Define the reference class explicitly. "Similar" must be operational: same industry, same type of project, same scale band. Vague reference classes produce vague base rates. When the class is thin, say so and widen the uncertainty band. When the class is strong, use it as the primary anchor and treat the inside view as a sensitivity.
Use percentiles, not just the median. The median of the reference class is a fair default. For high-stakes or one-shot projects, plan at the 80th percentile of cost or time. The extra margin is the price of avoiding the left tail of the distribution. Organisations that adopt this discipline report fewer overruns and less last-minute rescoping.
Investors should demand reference classes. When a founder gives a single-number forecast, ask for the reference class. What fraction of comparable companies hit that target? If the answer is "we're different," ask for evidence. The best founders already use an outside view; the rest learn when investors require it.
Section 10
Test Yourself
Is this mental model at work here?
Scenario 1
A product team estimates a launch in 6 months. The PMO requires them to compare against 20 similar product launches; the median was 11 months and 80th percentile 16 months. The team revises the plan to target 12 months with a 16-month contingency.
Scenario 2
A founder tells an investor they will reach $5M ARR in 18 months. The investor checks a dataset of similar B2B SaaS companies and finds the median time to $5M was 36 months. The investor asks the founder to justify why they are an outlier.
Scenario 3
A construction firm bids on a bridge using its internal cost model. It does not compare the estimate to other bridge projects of similar scope and location.
Scenario 4
A government requires that all IT projects above $50M must show their estimate against the distribution of 50+ similar past IT projects and set the budget at the 80th percentile of that distribution.
Section 11
Further Reading
The idea comes from Kahneman and Tversky; Flyvbjerg turned it into a practical method for megaprojects. These sources cover the psychology, the evidence, and how to implement it.
Kahneman explains the planning fallacy and the value of the outside view. Reference class forecasting is the operationalisation of that insight for project forecasting.
Tetlock's work on forecasting emphasises base rates and updating from evidence — consistent with reference class thinking for predictions.
Summary: Reference class forecasting replaces optimistic inside-view estimates with forecasts anchored to the distribution of similar past cases. Define the reference class, use its distribution (e.g. median or 80th percentile), and treat the inside view as a source of adjustment, not the primary anchor. The method reduces systematic overruns and improves decision quality when a relevant reference class exists.
Tension
Probabilistic Thinking
Probabilistic thinking is the habit of reasoning in distributions. Reference class forecasting is one implementation: the distribution comes from the reference class. The tension: when no good reference class exists, probabilistic thinking still demands a distribution, but you must construct it from other evidence and acknowledge higher uncertainty.
Tension
Optimism Bias
Optimism bias is the tendency to expect better outcomes than base rates justify. Reference class forecasting corrects by anchoring to those base rates. The tension: people resist the outside view because it feels pessimistic. Adopting reference class forecasting means accepting that "we're different" needs proof, not default.
