Planning fallacy
A Socratic walk-through of the planning fallacy — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why do experienced teams still underestimate how long their work will take?
Experience is supposed to fix estimation. A team that has been late on its last six projects has six data points, freely available, all pointing the same way. And yet the seventh estimate comes in optimistic again, from people who will readily agree, if asked in the abstract, that projects like theirs usually run late. That combination is the strange part — the knowledge is present and unused. What is the estimate being built from, if not from that?
Reasoning it through
REASONING #Ask what a person actually does when producing a schedule. They picture the work: these components, in this order, this one first, then the integration, then testing. They estimate each piece, and add up. This is a perfectly sensible procedure, and it has one structural property worth staring at. It can only include events that appear in the picture.
So what does not appear? Not the known risks — a careful planner adds contingency for those. What is missing is the whole category of things that have not been thought of: the supplier who goes under, the key person who leaves, the requirement that turns out to have been misunderstood, the dependency nobody knew existed. Any one of them is unlikely and unforeseeable. The proposition that several such things will happen is nearly certain, and it is exactly the proposition a step-by-step simulation cannot represent, because a simulation is made of named steps.
Notice the asymmetry that follows. Surprises can delay a project badly, but they very rarely accelerate it. A forecast built from the imaginable path is therefore not a central estimate of the outcome — it is close to a best case, and it gets reported as the expectation.
Kahneman and Lovallo gave this the name that sticks: the inside view, forecasting from the particulars of this plan, versus the outside view, forecasting from the record of similar past efforts. Kahneman's own example is the sharpest one I know. A team writing a school curriculum estimated, by looking at their outline and their progress, that they had about two years left. Then he asked a member who had seen other such teams what the record showed. The answer was that roughly forty per cent never finished at all, and that of those that did, none had taken less than seven years. The team heard this, and carried on with its own estimate. It took them eight.
Sit with that for a moment, because it is the striking part. They were not ignorant of the base rate — they had just been told it, by one of their own. Why did it not move them? Because the inside view is vivid and specific and theirs, while the base rate is a fact about other people, and every planner can list reasons why those other teams are not comparable. The bias survives knowing about the bias.
What is the fix, then? If the trouble is that the outside view is dismissed, the fix has to be procedural rather than exhortatory. That is what reference class forecasting does: define a class of genuinely comparable completed projects, get the actual distribution of their outcomes, place your project in that distribution, and adjust only for differences you can demonstrate rather than assert. Bent Flyvbjerg turned this into a working method for infrastructure, and the UK Treasury's guidance now requires estimates to carry empirically derived uplifts of this kind.
One more layer, and it complicates the story. Where estimates are competitive — tenders, funding rounds, internal bids for budget — the process selects for the lowest one. Even if every estimator were unbiased, the winner would be the one whose errors ran optimistic. And some underestimation is not error at all: Flyvbjerg argues that on large public projects strategic misrepresentation, forecasters shading numbers to get approval, does more work than cognition does. How the two divide is genuinely disputed, and probably varies by setting — the cognitive account fits a small team estimating its own sprint, the political one fits a bid for a tunnel.
The analogy
THE ANALOGY #Planning by the inside view is like estimating a train journey from the map: measure the distance, divide by the train's speed, and get an answer that is honest about geography and silent about everything else. The outside view is reading the operator's punctuality record for that route — which tells you nothing about why trains are late, and quite a lot about when you will arrive.
The map is simply less informative than the record, whereas a project plan is not wrong at all about its own steps — its blindness is confined to a category of event it cannot name; and unlike a railway, most organisations keep no honest archive of their past estimates, so the outside view is more often unavailable than refused.
Clarifying the model
THE MODEL #The planning fallacy is not general optimism, and that distinction does the real work. The same people who underestimate their own project will estimate someone else's accurately, and will agree that projects of this type usually overrun. The bias attaches to the specific plan you are inside, which is why more detail does not cure it — more detail makes the inside view more compelling, not more complete.
Two refinements worth carrying. First, adding a flat buffer is not the same as taking the outside view: a buffer scales the wrong estimate, and it is negotiated away by the same reasoning that produced it, whereas the outside view replaces the number with data whose provenance sits outside the argument. Second, the honest deliverable is a distribution rather than a date, because overruns are long-tailed — most projects somewhat late, a few catastrophically so — and a single day conceals that.
A picture of it
THE PICTURE #How to readEach bar is a reference class, and its height is how far the average completed project of that kind exceeded the cost forecast made when it was approved, from Flyvbjerg's study of transport projects across twenty nations. This is not a picture of the fallacy's mechanism but of its remedy: these three numbers are the outside view a planner of a new rail scheme should start from, before adding anything about their own plan. Read the gaps between the bars as a warning against a single generic uplift — the right adjustment depends on which class the project is actually in.
What became clearer
WHAT CLEARED #An estimate built from the plan can only contain what the planner imagined, and what defeats projects is mostly the residue nobody imagines — individually improbable, collectively near-certain, and asymmetric, since surprises delay and rarely accelerate. Experience does not fix it, because the record is a fact about other projects while the plan is vivid and yours. So the remedy is not to think harder about the schedule but to change where the number comes from: start from what comparable projects actually did, and let the specifics of this one adjust that rather than replace it.
Where to go next
ONWARD #- How to define a reference class narrow enough to be relevant and wide enough to have data.
- Why competitive bidding systematically selects the most optimistic forecast, and what procurement can do about it.
Key terms
TERMS #| Term | What it means |
|---|---|
| Inside view | a forecast built from the specifics of the plan at hand. |
| Outside view | a forecast built from the distribution of outcomes of comparable past efforts. |
| Reference class forecasting | the formal method of placing a project in a class of completed ones and forecasting from that record. |
| Strategic misrepresentation | deliberately optimistic forecasting to secure approval or funding, as distinct from cognitive bias. |
Every term the collection defines is gathered in the glossary.