What is reference class forecasting?

Reference class forecasting is an estimation method that replaces the forecast built bottom-up from a single project with the distribution of actual outcomes across a class of comparable finished projects. Instead of asking what this project will cost by adding up its activities, it asks what the most similar projects actually cost, and uses that distribution as the starting point to be adjusted. The theoretical origin lies in Daniel Kahneman and Amos Tversky's work on the difference between the inside view of a problem, held by the people inside it, and the outside view, which treats it as one case among many. Bent Flyvbjerg turned it into an operating procedure and put it on governments' agendas by documenting the systematic gap between estimates and final costs. Kahneman called it the single most important piece of advice for increasing the accuracy of a forecast, and the United Kingdom was the first European government to make it mandatory for transport project estimates.

Why a bottom-up estimate inherits the optimism of whoever makes it

An estimate built activity by activity looks more rigorous precisely because it is detailed, and instead it is fragile by construction: each line is estimated by whoever will carry it out, for the case where everything goes as planned, and the things that go wrong have no budget line because they have no name. The detail does not correct the distortion, it multiplies it. A reference class, by contrast, automatically incorporates everything that went wrong in past projects, including the causes nobody had anticipated, because it looks at outcomes rather than intentions. The study by Flyvbjerg and colleagues published in the Project Management Journal in 2026, covering 11,011 projects across 23 categories, measures a mean cost overrun of 73% for IT projects with a median that matches the estimate. The two figures together say what neither would say alone: the typical project lands on plan, and the mean is dragged by a minority of cases that overrun catastrophically. The same study reaches a stronger conclusion, namely that IT is the only one of the 23 categories with a tail fat enough to make mean and variance formally infinite. If that holds, "a mean overrun of 73%" is a statistic of the sample rather than a stable property of the phenomenon: the risk is not a 73% overrun, it is the tail.

An enterprise example

A manufacturing company has to replace its ERP. The vendor estimates nine months and the project lead confirms it, because the phased plan does add up. The reference class version asks a different question: how long did the last ten ERP replacements take at companies of comparable size and complexity. If the answer is that two took ten months, six landed around sixteen, and two were abandoned, the honest estimate is not nine months: it is a distribution with a tail to be managed up front, for instance by deciding in advance which modules stay out of the first release. The phased plan was not wrong, it just answered a different question from the one that mattered.

Why it matters for decision makers

For a decision maker the value is that it moves the argument from indefensible ground to verifiable ground. An internal estimate can only be challenged with another opinion, and in a meeting the more senior opinion wins; a reference class is challenged by bringing data, that is, by explaining why this project does not belong to that class, which is a legitimate and arguable objection. It is also the most direct antidote to the planning fallacy, and the precondition for any sensible Monte Carlo simulation: a simulation started from invented distributions produces invented confidence intervals. The adoption cost is low, because you do not need a formal method to begin, you need to stop throwing away the history of your own finished projects. The resistance it meets is not technical but political: a reference class makes it visible that the estimate a tender was won with was never a forecast.