Estimates come from the record of finished work, not from the plan

An estimate is a forecast made under two pressures: the work is not done, and somebody is waiting for a number. The number is easiest to build from the plan for this job, which is vivid and specific. The record of comparable finished work is duller and less flattering. This fourth of ten units, after the third constructed the brief, argues that the vividness is the problem.

The claim has two halves of unequal weight. That people are optimistic about their own completion times is among the better replicated findings in the judgement literature. That a base rate from finished work repairs the error is a claim about practice, and researchers contest parts of it.

People trust the plan and discount their own history

Buehler, Griffin and Ross (1994) tested three hypotheses across five studies. People underestimated their own completion times but not other people's. Think-aloud records showed predictions built from imagined future scenarios rather than from past jobs. Participants dismissed their own past overruns as unrepresentative. In Study 4 the optimism disappeared when participants were told to connect relevant past experience to the prediction. The bias is not ignorance of the record.

Buehler, Griffin and Peetz (2010) review that work and extend the inside-outside model to cognitive, motivational and social processes. A current plan dominates attention. What the planner has an interest in presenting to a sponsor matters too. Exhortation will not repair that: the record has to be imposed on the forecast, not consulted beside it.

The size of the problem depends on whose numbers you read, and on which literature

Molokken and Jorgensen (2003) reviewed surveys of software effort estimation. Most projects overrun. Their figure is 60 to 80 per cent. Expert judgement was the most common method, and no evidence showed formal models more accurate. Average overruns, at 30 to 40 per cent, sat below the 89 per cent the Standish Chaos Report publicised. The magnitude quoted in a sales meeting usually comes from the least inspectable source.

Little (2006) analysed one company's schedule data. He found estimation accuracy following a lognormal distribution, with the uncertainty range nearly identical throughout the project. That conflicts with the cone of uncertainty read as a closing triangle. He cites the same Standish report for the claim that only 20 per cent of projects finish on time. If the range does not narrow, a confident mid-project date records the progress of a narrative, not the state of the work.

Halkjelsvik and Jorgensen (2012) brought together the psychology literature on task duration prediction and the engineering literature on expert effort estimation. Underestimation was reported more frequently in the engineering and management studies. The psychology studies did not report it more frequently. They challenge accepted results on complexity and experience. They argue that the finding, that small tasks are overestimated and large tasks underestimated, can be partly a statistical artefact of random error. The bias is not a constant to be added to every forecast. Results do not transfer cleanly between traditions.

The correction with evidence behind it: a base rate from comparable work

Flyvbjerg (2006) documents inaccurate forecasts of project cost and demand, and explains them through optimism bias and strategic misrepresentation. He sets out reference class forecasting, which prices a project from the actual performance of comparable finished work. He reports its first practical application, to transport cost forecasts. Flyvbjerg (2021) later ranks base-rate neglect as a primary reason projects underperform. He presents base rates drawn from 2,062 projects and names ten behavioural biases. He argues that bias is political as well as cognitive, because the person who produces the number is often rewarded for it.

The first application is a project a reader can follow. From August 2004 the United Kingdom's Department for Transport and HM Treasury required an uplift for optimism bias in the appraisal of large transport schemes. The uplift was derived from the cost records of comparable finished projects. The first recorded use came that October, in the planning of Edinburgh Tram Line 2. The base cost was 255 million pounds. The uplifts put the 50th percentile for total capital cost at 357 million pounds and the 80th at 400 million pounds. The second figure is the one at which the chance of staying inside budget is 80 per cent. The estimator was taking an outside view on a scheme with no comparable tram precedent of its own.

Buehler, Griffin and Peetz (2010) work inside the inside-outside framework, in which an optimistic view of the case is corrected by distributional information from outside it. Pricing from a class of finished projects makes that correction checkable. The class is named, so a client can disagree with the class rather than with the estimator's confidence.

An estimate is not a commitment, and it changes the work

Jorgensen and Sjoberg (2001) studied two software organisations and ran two experiments on what an estimate does after it is given. Projects prioritising cost with incomplete specifications trimmed the work to fit an optimistic estimate. Quality-critical work with complete specifications overran instead. Early anchor figures moved later estimates even when declared irrelevant. Optimistic estimates produced less effort and more errors than realistic ones.

McConnell (2006) supplies the distinction that keeps this honest. An estimate is a prediction. A target is a desirable business objective, and a commitment is a promise to deliver defined functionality at a quality level by a date. The request for an estimate is often a request for a commitment. Jorgensen (2004) adds what the estimator can control. Twelve best-practice guidelines carry supporting evidence, among them using documented data from previous tasks and stating the estimate's uncertainty.

McConnell (2006) opens his chapter with a sentence he attributes to Brooks (1975). The age of the argument is part of the point.

It is very difficult to make a vigorous, plausible, and job-risking defence of an estimate that is derived by no quantitative method, supported by little data, and certified chiefly by the hunches of the managers. (The published sentence spells that word in the American form.)

Contested question: is a base rate a correction or a bargaining position?

The literature agrees that plans are systematically flattering and disagrees about what follows. One position holds that the error is bias, partly cognitive and partly strategic. On that view an outside view of comparable projects bypasses both. Its chief statements are Flyvbjerg (2006), which sets out reference class forecasting, and Flyvbjerg (2021), which locates the failure in base-rate neglect across 2,062 projects.

A second position accepts the diagnosis and doubts the remedy's range. Baerenbold (2023) reviewed 41 papers and finds the effectiveness turning on how the class is defined, and best supported for large-scale projects. Cantarelli, Davis, Pinto and Turner (2026) set out a research agenda and record that validity, applicability and limits remain open as adoption grows. The method is strong where many comparable cases exist. It is weak for a team pricing a first-of-its-kind product, or one whose users work over a metered connection and on an older device.

A third position holds that the base rate is one lever among several. Halkjelsvik and Jorgensen (2012) show the bias is reported unevenly across literatures. The correction therefore cannot be calibrated with that confidence. Jorgensen and Sjoberg (2001) show the estimate reshaping the work, so an outside-view number can become an inside-view target on receipt. Jorgensen (2004) finds the estimator's process doing work that the choice of model does not.

What would settle the disagreement is evidence the literature does not contain. The needed study takes one set of projects and produces an inside-view forecast and a reference class forecast for each. It fixes the class definition in advance, records what was delivered, and measures each forecast's accuracy and its effect on the work.

On evidence the sides differ in kind. The bias finding rests on controlled experiments and a review; the remedy rests on one method's track record, which Baerenbold and Cantarelli find to turn on how the class is defined. While the dispute is open, a designer should give both numbers and mark which is measured and which is assumed. The strongest objection is that a base rate can become a bargaining position, because a forecast built to win work is as self-serving as an optimistic one.

Where the argument is heading

The direction of travel is towards documenting the provenance of an estimate rather than defending the number. Cantarelli, Davis, Pinto and Turner (2026) are building the agenda for the validity of reference class forecasting. Baerenbold (2023) asks for empirically based rules for forming a class. Flyvbjerg (2021) has assembled base rates from 2,062 projects. Jorgensen (2004) sets out when expert judgement performs. The professional form of the coming years is a written record. It names the class the work was priced against, the range, the owner of the commitment, and what was delivered.

The sources this unit rests on

• Baerenbold, R. (2023). Reducing risks in megaprojects: the potential of reference class forecasting. Project Leadership and Society. doi:10.1016/j.plas.2023.100103 Reviews 41 papers and finds its effectiveness depending mainly on how the reference class is defined, and best supported for large-scale projects. Calls for an empirically based framework for forming classes rather than a one-size-fits-all remedy.
• Brooks, F. P. (1975). The Mythical Man-Month. Proceedings of the international conference on Reliable software. doi:10.1145/800027.808439 The conference record of the book, not a measured study: it describes the book as mixing published software engineering data with personal assertion, offered to invite dispute. Cited for the age of the argument and for the warning McConnell (2006) attributes to Brooks.
• Buehler, R., Griffin, D. and Peetz, J. (2010). The Planning Fallacy: Cognitive, Motivational, and Social Origins. Advances in Experimental Social Psychology. doi:10.1016/s0065-2601(10)43001-4 Reviews the planning fallacy and develops an extended inside-outside model of its cognitive, motivational and social origins. Argues that the bias follows from how a current plan dominates attention and from what a planner has an interest in presenting.
• Buehler, R., Griffin, D. and Ross, M. (1994). Exploring the 'planning fallacy': why people underestimate their task completion times. Journal of Personality and Social Psychology. doi:10.1037/0022-3514.67.3.366 Five studies tested three hypotheses: people underestimate their own completion times but not others', predictions are built from plan-based scenarios, and the past is discounted as unrepresentative. In Study 4 the optimism was eliminated when participants were told to connect relevant past experience with the prediction.
• Cantarelli, C. C., Davis, K., Pinto, J. K. and Turner, N. (2026). Reference class forecasting: promises, problems, and a research agenda moving forward. Production Planning and Control. doi:10.1080/09537287.2025.2578708 Critically reviews the method, whose adoption has grown in policy, and finds its validity, applicability and limits still open. Sets out the research agenda, the unit's most recent evidence that the remedy is unsettled.
• Flyvbjerg, B. (2006). From Nobel Prize to Project Management: Getting Risks Right. Project Management Journal. doi:10.1177/875697280603700302 Documents inaccurate forecasts of cost and demand, explains them through optimism bias and strategic misrepresentation, and sets out reference class forecasting, which prices work from the actual performance of comparable projects. Reports its first practical application, to the cost forecasts of large transportation projects.
• Flyvbjerg, B. (2021). Top Ten Behavioural Biases in Project Management: An Overview. Project Management Journal. doi:10.1177/87569728211049046 (the published title spells the second word in the American form) Names the ten most consequential behavioural biases in project management, argues that bias is political as well as cognitive, and presents base rates from 2,062 projects. Identifies base-rate neglect as a primary reason projects underperform.
• Halkjelsvik, T. and Jorgensen, M. (2012). From origami to software development: a review of studies on judgement-based predictions of performance time. Psychological Bulletin. doi:10.1037/a0025996 (the published title spells that word in the American form) Integrates the psychology and engineering literatures on time prediction and finds underestimation reported more often in engineering and management studies than in psychology studies. Challenges earlier results on complexity and experience, and argues that the small-versus-large task pattern can be partly a statistical artefact.
• Jorgensen, M. (2004). A review of studies on expert estimation of software development effort. Journal of Systems and Software. doi:10.1016/s0164-1212(02)00156-5 Finds expert judgement the most frequently used strategy, no substantial evidence in favour of formal models, and fifteen studies comparing the two directly. Evaluates twelve best-practice guidelines, including documented data from previous tasks and the assessment of estimate uncertainty.
• Jorgensen, M. and Sjoberg, D. I. (2001). Impact of effort estimates on software project work. Information and Software Technology. doi:10.1016/s0950-5849(01)00203-8 Reports two case studies and two experiments on what an estimate does to the work that follows. Projects prioritising cost with incomplete specifications trimmed work to fit an optimistic estimate, while well-specified quality-critical work overran; early anchors moved later estimates even when declared irrelevant, and optimistic estimates produced more errors.
• Little, T. (2006). Schedule estimation and uncertainty surrounding the cone of uncertainty. IEEE Software. doi:10.1109/ms.2006.82 Analyses schedule data from a project portfolio and finds estimation accuracy following a lognormal distribution, with the uncertainty range nearly identical throughout the project, contradicting the cone of uncertainty read as a narrowing triangle. Cites the Standish report for the figure that only 20 per cent of projects finish on time.
• McConnell, S. (2006). Software Estimation: Demystifying the Black Art. Redmond, WA: Microsoft Press. https://www.microsoftpressstore.com/store/software-estimation-demystifying-the-black-art-9780735605350 The profession's standard practitioner reference on estimating. Its opening chapter distinguishes an estimate from a target and a commitment, argues that the three are routinely conflated when an executive asks for an estimate, and quotes the Brooks warning used above.
• Molokken, K. and Jorgensen, M. (2003). A review of software surveys on software effort estimation. Proceedings of the International Symposium on Empirical Software Engineering. doi:10.1109/isese.2003.1237981 Reviews software estimation surveys and finds that most projects overrun (60 to 80 per cent). Expert judgement is the most common method, with no evidence that formal models are more accurate, and average overruns of 30 to 40 per cent sit far below the 89 per cent publicised by the Standish Chaos Report.