Planning fallacy

The weekend repaint

“Painting the bathroom is a Saturday-morning job.” The last two rooms took the whole weekend each, but “that was because of the old wallpaper, and the hallway had all that trim.”

The estimate comes from picturing the job going smoothly. The track record that would correct it is known, but each past overrun is explained as a one-off. The explanations may all be true; the problem is that a different one-off keeps turning up every time. This is the pattern Roger Buehler and colleagues described: past delays put down to specific, temporary causes, so they don’t count toward the next forecast.

A “worst case” that is still a plan

Asked for a worst-case estimate for a report, a student thinks through the steps: “Research a week, write a week, and if the draft needs a lot of revision, another week. Three weeks, tops.” Their last two reports of that length took four weeks and five.

Adding a pessimistic step doesn’t change where the number came from. The worst case is still a version of the plan, with one imagined problem added, not a look at how long such reports have actually taken. In the study described under Evidence, students’ worst-case estimates for their theses were, on average, still shorter than the time they took.

Someone else’s track record

A couple gets a contractor’s quote for a kitchen renovation and sets their budget at exactly that figure. Three friends who renovated kitchens in the past two years all went well over their quotes.

This version ignores a base rate (how often something happens across comparable cases) that belongs to other people rather than to you. The quote describes this job as planned; the friends’ experience is evidence about how jobs like it tend to go. Treating your case as unique, when it’s one of many similar ones, is what Kahneman calls taking the inside view.

• Time and cost. The research began with time predictions; the same pattern is described for budgets, especially in large public projects.
• Completion time versus working time. “When will it be done?” and “How many hours will it take?” are different forecasts. Torleif Halkjelsvik and Magne Jørgensen note that many studies don’t separate them.
• Self versus others. In Buehler’s studies, observers predicting someone else’s completion time erred in the pessimistic direction, not the optimistic one.

• When there’s no comparable record. For a genuinely new kind of task, a plan-based estimate may be the best available. The error is ignoring a track record that exists, not failing to have one.
• When this case really does differ, in a way you can name in advance. If past delays had a specific cause and that cause has actually been removed (last time you had to borrow the equipment; this time you own it), expecting to be faster is reasonable. The difference from explaining away is that the change is identified before the forecast and addresses what actually caused the earlier delays.
• When a best case is presented as a best case. A stretch target or an “if everything goes right” date is fine as long as nobody treats it as the expected date.

The test: is this number what I expect, given how long similar things have actually taken, or what the plan says if it goes right?

A short estimate from your own record

Friends expect a room repaint to take a weekend. A painter who does it often says: “About five hours. My last ten rooms this size took between four and six.”

The estimate is shorter than most people’s, and that can sound like overconfidence. But it’s built from the painter’s own track record on the same kind of job, which is exactly the information the fallacy leaves out.

A sound estimate that still ran late

An office move was scheduled for three months, based on eight similar moves the firm had managed, which took between two and a half and three and a half months. A furniture supplier went out of business mid-project, and the move took five months.

Running late doesn’t show the estimate was flawed. It came from a reference class of similar cases, and one unusual event pushed the result outside that range. Concluding afterward that “the plan was obviously too optimistic” would be Hindsight bias: judging the forecast by an outcome that wasn’t knowable when it was made.

Kahneman and Tversky’s account, which gave the error its name, contrasts two ways of forecasting, which Kahneman later called the inside view (working out how this task will unfold) and the outside view (asking how tasks like it have gone). People forecasting their own work default to the inside view. In Buehler’s 1994 think-aloud study, about 70% of what students said while predicting described their plans for the task, and about 1% mentioned past problems.

Kahneman, describing a textbook project that took far longer than his team predicted, puts it in terms of “what you see is all there is”: the team was “forecasting based on the information in front of us”, its progress on the first chapters, which were probably easier than the rest and written while enthusiasm was at its peak.

Buehler’s group also found that people explain their own past delays as more temporary and more specific to the situation, which makes those delays easy to set aside. Observers, who have no plan to picture, lean more on the track record.

Other explanations compete with or add to this one:

• Memory. Michael Roy and colleagues argue that people may use their past experience correctly but remember past tasks as having taken less time than they did.
• Incentives. For public works, Bent Flyvbjerg and colleagues concluded that low cost estimates are best explained by strategic misrepresentation (promoters understating costs to win approval) rather than honest error. Flyvbjerg’s later work addresses optimism bias and strategic misrepresentation as distinct problems.
• Anchors and question format. Halkjelsvik and Jørgensen’s review lists various anchoring effects, incentives and how the question is asked among the influences on time predictions.

• Connect the past to this plan, not just recall it. In Buehler’s fourth study, simply remembering past assignments didn’t change predictions. Students who were also asked to spell out how their past experience applied to the new task made predictions that were no longer biased on average (60% finished on time, against 29% in the control group). Their predictions were, however, no more accurate in absolute terms: the errors just stopped leaning one way.
• Break the task into parts. Justin Kruger and Matt Evans found that asking people to list a task’s components led to longer, and in some cases less biased, time estimates.
• Use a reference class. Reference class forecasting bases the estimate on the actual outcomes of similar past projects. Jordy Batselier and Mario Vanhoucke tested it on real project data and found it forecast both cost and time better than the common alternatives they compared it with (simulation-based and earned-value methods). In everyday terms: before estimating, ask how long the last few comparable jobs actually took.
• Ask someone outside the task. Observers in Buehler’s studies were less optimistic, but not more accurate; an outside estimate is a useful check, not a replacement.

Status: contested. Overruns are well documented, and early studies of students’ real tasks found clear underestimation. But no preregistered or multi-lab replication of the laboratory effect could be found, nor a meta-analysis corrected for publication bias, and a major review disputes how general the pattern is. Under this site’s criteria that falls short of robust, and the dispute over generality and cause fits contested.

• The term comes from Kahneman and Tversky (1979), who described forecasts that neglect information about similar cases.
• Buehler, Griffin and Ross (1994) asked psychology students how long their honors theses would take. The average best estimate was 33.9 days; the theses took 55.5 days, and fewer than a third of students finished by their estimate. Even their estimates for “if everything went as poorly as it possibly could” (48.6 days) came in short on average. Across four studies of academic and everyday tasks, fewer than half of participants finished in the time they predicted. The estimates still tracked reality in one respect: students who predicted longer times did take longer.
• Buehler, Griffin and Peetz (2010) reviewed the research program that followed, covering cognitive, motivational and social influences on the bias.
• Field evidence on overruns. Flyvbjerg, Holm and Buhl (2002) studied 258 transportation projects and found actual costs exceeded estimates in 86% of them, by 28% on average (rail about 45%, bridges and tunnels about 34%, roads about 20%). This shows forecasts for large projects are biased, but the authors attributed it to deliberate misrepresentation, not a reasoning error, so it isn’t direct evidence for the psychological claim. Peter Love and Dominic Ahiaga-Dagbui (2018) criticized the study’s methods and called its split between “error” and “lie” a false dichotomy.
• Generality is disputed. Halkjelsvik and Jørgensen’s (2012) review of time-prediction studies found that underestimation was reported more often than overestimation in engineering and management studies, but not in studies from psychology. They also argue that the common finding “small tasks are overestimated, large tasks underestimated” may be partly a statistical artifact of random error.
• The explanation is disputed. Roy and colleagues (2005) reviewed evidence that underestimation may come from biased memories of past durations rather than from neglecting them.

What seems secure is that forecasts for large projects overrun far more often than not, and that estimates built from a plan leave out delays a track record would include. What remains uncertain is how widespread systematic underestimation is in ordinary personal tasks, and how much of it is optimism, faulty memory, incentives or the way questions are asked.