Planning fallacy

In Buehler, Griffin and Ross's founding study, honours students asked to predict as accurately as possible when they would hand in a thesis already under way averaged 33.9 days against an actual 55.5, and 29.7% finished inside the time they had named. Asked for the date “if everything went as poorly as it possibly could”, they averaged 48.6 days — still short of 55.5, and fewer than half (48.7%) met even that. The estimates were biased but not uninformative, correlating .77 with actual completion, and asking for a pessimistic figure moved the number without improving absolute accuracy: a worst-case field does not buy a calibrated estimate.

The mechanism is attention to the plan rather than to past experience. Think-aloud protocols in the same paper were dominated by future steps with almost no reference to how earlier assignments had gone, and observers given the same situation drew more on past cases and predicted longer. Study 4 is the debiasing result everything else leans on: students underestimated a computer assignment by 1.3 days in the control condition, 1.0 day when merely asked to recall similar past assignments, and 0.1 days when additionally walked through questions tying those past cases to this one, with 60.0% of that group finishing inside their prediction against 38.1% and 29.3%. Read the whole result: absolute error did not improve in any condition, and what moved actual completion was the deadline — 4.3 days under a one-week deadline against 9.1 under two weeks.

The headline generalisation does not survive the wider literature. Halkjelsvik and Jørgensen's Psychological Bulletin review of judgment-based performance-time predictions reports typical average unsigned errors of 20–30% and no general tendency towards either overoptimistic or overpessimistic predictions across the tasks and projects they could collect. Everyday tasks predicted in the laboratory came out on average unbiased, as did smaller professional software tasks, and where lateness carries a sharp penalty — telling a waiting customer when food will arrive — people are overpessimistic. What exists instead is a magnitude effect: where the experimenter fixes task size, larger tasks are underestimated and smaller ones overestimated. The field version of that effect is inflated: of 13 studies compared, all seven that measured task size by actual time or cost found larger tasks more underestimated, while those measuring size by predicted time or budget found little or none, because using the outcome as the size variable manufactures the correlation. Underestimation is therefore a claim about a region — long, open-ended, unfamiliar work the person is invested in — not about people, and treating a two-minute form like a Sydney Opera House is the first mistake.

That makes “break the task down” a direction, not a fix. In Forsyth and Burt's three experiments a single future task always received significantly less time than its summed subtasks, and marking estimates on a timeline ruled out rounding. Decomposition reliably raises the number, which helps only when the number was too low. In two office-task experiments, decomposed predictions ran 9% and 10% too high where holistic predictions of the same work ran 2% and 5% too high; in a setting that did skew low, decomposed predictions ran 0% and 8% too low against 13% and 26% for the whole. Connolly and Dean found decomposed estimates of software writing tasks higher but less accurate than holistic ones. Aggregation is arithmetically fragile as well: summing most-likely component times instead of means gives a total that is too low even when every component is honest.

Which components you name decides the direction. In Hadjichristidis, Summers and Thomas's document-formatting task, a packed description drew a mean estimate of 8.48 minutes, naming the fiddly step (phonetic characters) drew 12.73, and naming a quick step (italics and bold) drew 6.24. On getting ready for a date, packed drew 57.76 minutes, naming an early step 74.46, naming a late step 48.81, and naming typical components changed nothing. Kruger and Evans's enumeration result is the upward half of this pattern; the downward half is the part product teams need, because a step list naming the easy or final steps lowers the estimate. Concrete step-by-step planning has also been observed to increase overoptimism specifically in completion-date predictions, which is the prediction a progress tracker invites.

The intervention with the better institutional record is the outside view. Flyvbjerg, Holm and Buhl found an 86% chance that a randomly chosen project cost more than estimated across 258 transport infrastructure projects worth about $90 billion, with average overruns of 44.7% for rail, 33.8% for bridges and tunnels and 20.4% for roads, 27.6% overall; they conclude error cannot explain the pattern and strategic misrepresentation can — a cause no debiasing of the forecaster touches. Forecasting from the distribution of outcomes in a reference class of past projects is policy rather than hypothesis: HM Treasury's supplementary Green Book guidance has appraisers start from an upper-bound uplift drawn from historical projects and justify any reduction, with works-duration uplifts from 4% upper and 1% lower for standard buildings to 54% and 10% for equipment and development. That is an adopted, data-derived correction, not a controlled trial. No randomised evidence shows reference-class forecasting beats the inside view, and the one observational software comparison, by Furulund and Moløkken-Østvold — projects using estimation checklists averaging 10% too low against 38% too low for projects that did not — is not randomised either.

Almost none of this has been tested inside a product interface, and the one controlled strand runs against the easy advice. Galesic and Bosnjak randomly announced the same web questionnaire as taking 10, 20 or 30 minutes: the longer the stated length, the fewer people started and the fewer completed. An honest estimate at the top of a flow is a real cost, which is why it gets shaved, and nothing in the literature shows an accurate estimate recovers that cost later in retention or trust. A step tracker is no free substitute either: Villar, Callegaro and Yang meta-analysed 32 randomised progress-indicator experiments and found constant indicators did not reduce drop-off overall, fast-to-slow indicators reduced it, slow-to-fast indicators increased it, and a constant indicator raised drop-off where a small incentive had been promised.

Where a user types the estimate, the field sets the answer. In Jørgensen's experiment software professionals predicting in workdays gave the equivalent of 88 work hours on a task where those predicting in work hours gave 45, and 335 against 224 on a larger one. Halkjelsvik and Jørgensen close with the ways to extract an overoptimistic prediction — a low anchor, several bids with the lowest taken, a small task asked about first, rewards for speed, a fine-grained unit — which is a list of things estimate fields, bidding marketplaces and sprint tools do by accident. The overall picture is mixed: the original demonstration is solid for the narrow case it covers, the general claim is not supported, and both remedies the folk version recommends are either conditional on the direction of the existing bias or untested in an interface.