Identify the right reference class for your situation

Find a well-defined set of past situations that are structurally similar to yours and collect their outcome data.

Why it works

Inside-view planning treats your situation as unique, so historical failure rates feel irrelevant. Identifying the reference class forces acknowledgment that your project belongs to a category with a known distribution of outcomes — and that distribution is the best prior for your forecast before any project-specific information is added. The mechanism is Bayesian: the class distribution is the prior; inside-view details are updating evidence, not the starting point.

How to do it

1. Define what type of project or decision you are making (e.g., "software project over six months with a new team," not "my project").
2. Search for historical data on completed projects of that type: duration, cost overruns, success rates.
3. Choose the reference class that is specific enough to be meaningful and broad enough to have statistical power.
4. Record the median and the range (10th–90th percentile) of outcomes for this class before opening your project plan.

Evidence

Flyvbjerg and colleagues analyzed hundreds of large infrastructure projects and found systematic cost overruns averaging 28% for roads and much higher for rail and tunnels. Reference class forecasting has since been formally recommended by the UK Treasury and the American Planning Association as a planning tool. (observational)

The method requires a usable reference class; for genuinely novel situations (first-ever technology, unprecedented events) the class is too thin or nonexistent, and base-rate anchoring provides false precision.

Sources

• Flyvbjerg, Holm & Buhl (2002), "Underestimating costs in public works projects," Journal of the American Planning Association
• Kahneman & Tversky (1979), "Intuitive prediction: Biases and corrective procedures"

Common mistake

Defining the reference class too narrowly ("projects exactly like mine") to preserve optimism — if you can only find two past examples, you do not have a reference class, you have anecdotes.