Reference Class Forecasting
Reference Class Forecasting: Before predicting outcomes for your specific situation, look at the actual distribution of outcomes for comparable past situations. The base rate is almost always more accurate than your case-specific estimate.
What Is Reference Class Forecasting?
In the 1970s, Daniel Kahneman and Amos Tversky documented a systematic bias in human forecasting: people consistently underestimate how long projects will take, how much they will cost, and how likely they are to fail. This became known as the Planning Fallacy. The cause: when forecasting, people adopt an "inside view" — they reason about the specific case, its particular features, and the steps they plan to take. They neglect the "outside view" — what actually happened in comparable situations.
Reference Class Forecasting (RCF) is the systematic antidote. Developed and applied by Danish economist Bent Flyvbjerg (who documented massive systematic cost overruns in infrastructure projects worldwide), RCF requires forecasters to:
- Identify a reference class of comparable past projects or situations
- Establish the statistical distribution of outcomes for that class
- Use the actual base rate as the starting point for prediction
- Adjust for features that genuinely distinguish the current situation from the reference class
The key insight: the outside view almost always produces more accurate forecasts than the inside view, because it draws on actual outcomes rather than optimistic projections. A bridge construction project that is estimated to take 3 years should be immediately checked against the base rate for comparable bridge projects — and if that base rate shows a median of 4.5 years with 80% running over estimate, the forecast should reflect that distribution.
How It Works
Step 1: State your initial inside-view estimate
"This project will take 6 months and cost $500K."
Step 2: Define the reference class
What are comparable past situations?
— Same type of project (software platform builds)
— Same industry and scale
— Same level of complexity
Be specific but not so narrow that the class is too small.
Step 3: Find the actual outcome distribution
— What was the median outcome in the reference class?
— What was the 75th percentile? 90th?
— What fraction came in on time and on budget?
— What was the typical overrun ratio?
Step 4: Anchor your forecast on the base rate
"The reference class shows median timeline of 9 months
with 70% running over estimate. My base case should be
9 months, not 6."
Step 5: Adjust for genuine distinguishing factors
Are there specific features of your situation that make
it genuinely different from the reference class in
ways that affect the likely outcome?
Adjust conservatively — most inside-view features are
also features of the reference class cases.
Step 6: Express the forecast as a distribution
Not "9 months" but "50% probability it's done in 9 months,
25% probability it takes 12–14 months."
Real-World Examples
Example 1: Infrastructure Cost Overruns (Flyvbjerg's Research)
Bent Flyvbjerg analyzed 258 transportation infrastructure projects across 20 nations (railways, bridges, tunnels, roads) and found that 9 out of 10 ran over budget, with an average cost overrun of 28%. Rail projects overran by an average of 45%. For large projects, the overrun was often far larger: the Sydney Opera House overran by 1,400%; the Boston Big Dig by 220%.
In every case, the original forecasts were generated from inside views: project-specific analysis of costs, timelines, and risks. None was anchored on the base rate of what comparable projects actually cost.
Flyvbjerg's application of RCF: before accepting any infrastructure cost estimate, require forecasters to consult the reference class distribution and explain, specifically, why this project would be in the better-performing half of comparable projects. This shifts the default forecast from "this will be fine" to "this will probably overrun — how much?"
Example 2: Software Product Development Timelines
A software product team estimates a new feature integration will take 6 weeks. A senior manager runs a reference class check: of the last 12 significant integrations the team has completed, how many were done within the original estimate? Four of twelve (33%). What was the median actual timeline? 8.5 weeks. What was the 90th percentile? 14 weeks.
The reference class forecast: base case of 8.5 weeks (not 6), with a 30% probability of taking 12 weeks or more. The team communicates this to stakeholders. They build the product roadmap around the base case.
The integration actually takes 9 weeks. The forecast based on the reference class was much more accurate than the inside-view estimate.
Example 3: Startup Success Probability
A first-time founder is raising a seed round. She tells investors the probability of her startup succeeding is 60% — based on her assessment of her product's strength, her team's capabilities, and her target market's size.
An investor applies reference class thinking. The base rate for startups that raise seed funding and return 3x or more: approximately 10–15% by most studies. The base rate for first-time founders specifically is lower.
This doesn't mean the founder's confidence is necessarily wrong — she may have genuinely distinguishing factors. But it does mean that her 60% estimate requires a very specific and compelling explanation of why her situation differs from the base rate by such a large margin.
When to Use It
✅ For any project, timeline, or cost estimate where you're doing inside-view planning. Always check the base rate.
✅ When evaluating others' forecasts. Ask: what is the reference class for this prediction, and what does the base rate say?
✅ For investment due diligence. What is the actual success rate of comparable investments?
✅ For personal goals and commitments. How long have comparable goals actually taken people to achieve?
❌ When there is no meaningful reference class. Genuinely novel situations without comparable historical outcomes require different methods. RCF requires a reference class.
Model Combinations:
| Combine with | Effect |
|---|---|
| Bayesian Thinking | The reference class provides the prior probability; Bayesian updating adjusts for case-specific evidence |
| Margin of Safety | Use the reference class distribution to calibrate the appropriate margin — higher variance classes require larger margins |
| Planning Fallacy | RCF is the primary corrective for Planning Fallacy |
Common Misuses and Limitations
Misuse 1: Defining the reference class to get the desired answer. If you select a reference class narrow enough, you can find almost any base rate you want. The reference class should be defined by objective criteria (project type, scale, complexity) before the base rate is looked up.
Misuse 2: Ignoring genuine distinguishing factors. RCF anchors on the base rate and adjusts for distinguishing factors. Not every distinguishing factor is noise — some are genuinely predictive. The discipline is adjusting conservatively rather than using distinguishing factors to return to the inside-view estimate.
Limitation — reference class selection requires judgment: There is no algorithm for selecting the correct reference class. The choice of class significantly affects the base rate. This requires expertise in the relevant domain and intellectual honesty about which class is truly comparable.
Related Models
Planning Fallacy: The bias RCF is designed to correct — the systematic underestimation of time, cost, and risk.
Bayesian Thinking: RCF provides the prior; Bayesian updating handles the case-specific adjustment.
FAQ
Where do I find reference class data for my situation?
The best sources depend on the domain. For software projects: post-mortem reports, internal historical data if available, academic studies of software project completion rates. For infrastructure: Flyvbjerg's published research. For startups: CB Insights, Crunchbase, and academic venture capital studies. For personal goals: survey research on goal attainment rates. When published data isn't available, your own historical record (how long have similar past projects taken?) is often more accurate than an inside-view estimate.
Should I always use the median as my base case?
The median is a reasonable starting point, but the right anchor depends on your context. For decisions where the downside of being wrong is severe, anchor higher in the distribution (the 75th percentile). For decisions where you're making many similar bets and the average outcome is what matters, the median is appropriate. Also consider the skewness: if the distribution has a long right tail (a few projects take very much longer than the median), even the 75th percentile may understate the realistic upper bound.
What is the best resource for learning Reference Class Forecasting?
Bent Flyvbjerg's book How Big Things Get Done (2023, with Dan Gardner) is the most accessible treatment, covering infrastructure projects and broader applications. His academic paper 'Survival of the Unfittest' (2009) documents the systematic optimism in forecasting. Daniel Kahneman's Thinking, Fast and Slow (2011) covers the inside/outside view distinction in Chapters 22–23.
Apply This Model with AI
Describe the estimate you're making — project timeline, cost, success probability — and the type of situation in MindMax. The AI will help you identify an appropriate reference class, find or estimate the base rate, and anchor your forecast on the outside view.
🚀 Apply Reference Class Forecasting in MindMax →
Further Reading
- Bent Flyvbjerg and Dan Gardner, How Big Things Get Done (2023) — The most accessible and comprehensive treatment of reference class forecasting, with extensive examples.
- Daniel Kahneman, Thinking, Fast and Slow (2011) — Chapters 22–24 cover the inside vs. outside view and their implications for forecasting.
This page is part of the MindMax Mental Models Knowledge Base.