reference-class-forecasting

Reference-class forecasting (Kahneman and Tversky's inside view versus outside view, operationalized for planning by Bent Flyvbjerg to expose chronic cost and schedule overruns of large projects) treats your case not as a unique story but as one instance of a class of similar past cases whose actual outcomes are the most reliable anchor you have.

It is built to defeat the planning fallacy: forecasting from the inside view (the specific plan, the specific team, the reasons this time will go smoothly) systematically underestimates time and cost and overstates success, because it sees the plan but not the unknown unknowns that derailed every comparable effort. The outside view corrects this by starting from the base rate: how did projects like this actually turn out, before you adjust for anything special about yours.

When to use

The user must predict a quantity or outcome for a case that belongs to a recognizable class of similar efforts: how long a project will take, what it will cost, whether a launch will hit its target, what adoption or conversion to expect, the odds a venture succeeds. Triggers: "outside view", "base rate", "reference class", "planning fallacy", "how long will this really take", "are we being too optimistic", "how do things like this usually go".

Reach for it specifically when the inside-view estimate feels confident and you suspect that confidence is the problem: when there is a track record of similar cases to learn from, and a strong pull to believe this one is different. It is the outside-view complement to fermi-estimation: Fermi builds an estimate up from the specifics of this case (the inside view); reference-class anchors it on how the class actually performed. Run both and reconcile the gap. It also strengthens a pre-mortem (the base rate of failure is the reason to take the pre-mortem seriously) and feeds scenario-planning.

Do not use it when the case is genuinely without precedent (no class exists to anchor on, though be very skeptical of "unprecedented", which is usually optimism in disguise), nor for adjudicating which explanation is true (use analysis-of-competing-hypotheses).

Inputs

• PREDICTION (required): what is to be forecast, stated as a specific measurable outcome for a specific case: "how many months until this platform migration is live", "the final cost of this office build", "the first-year retention of this product". Name the target variable and its units.
• THE CASE (required): enough about the specific case to find its peers and judge its distinctiveness: what it is, its scale, its context, what makes the user think it is typical or atypical.
• INSIDE-VIEW ESTIMATE (optional): the estimate the user has already formed from the case's specifics, if any. Used to measure the planning-fallacy gap, not to anchor the forecast.