Stop Asking for a Point Estimate: Forecast Software Work as a Calibrated Range
A single number handed back under uncertainty is not a plan; it is a commitment wearing a forecast's clothes. This article argues that point estimates invite systematic anchoring bias that confidence intervals do not fix, and that teams should predict effort from the distribution of comparable completed work, state a range, and then choose a risk level explicitly — while treating the reference-class method as a disciplined craft rather than a formula.
Series: developer-tools (8 articles)
- …
- 5Stop Asking for a Point Estimate: Forecast Software Work as a Calibrated Range This article
- The error is documented, not anecdotal
Software work has a long record of landing late. A survey of the field, summarizing Magne Jørgensen's work in IEEE Software, reports that cost and effort overruns average around 30% [1]. That average is a rounded summary across heterogeneous studies, so read it as an order of magnitude rather than a target anyone should reproduce. What matters is the direction: expert judgment, which is how most teams actually estimate, is biased and inconsistent even when the people doing it are experienced.
The mechanism behind the bias is not that estimators are careless. It is that judgment under uncertainty runs on heuristics, and one of them is demonstrably manipulable.
- The first number heard does the deciding
In anchoring and adjustment, a person starts from an initial value and moves away from it, but the adjustment is usually insufficient, so the final judgment stays pulled toward the anchor. A controlled software-estimation exercise put this directly to the test. Twenty-three participants estimated the same project from identical requirements; the only difference was a single sentence in the background material, where a manager who admitted no estimating experience guessed two months, guessed twenty months, or gave no number. The mean estimates came out at 6.8 months with the low anchor, 8.3 months with no anchor, and 17.4 months with the high anchor. The high-versus-low difference was strongly significant (t = 4.273, p < 0.001), and the high anchor differed from the no-anchor control (t = 3.137, p < 0.01) [2].
Two details keep this honest. The low anchor did not differ significantly from no anchor (p > 0.1), so the evidence supports a strong high-anchor pull rather than a symmetric effect, and the sample is small — 23 people, most of them graduate students rather than senior professionals [2]. A larger set of three experiments involving 381 software professionals reinforces the core finding and adds an uncomfortable result: the anchoring effect did not weaken when the anchor was less precise or came from a less credible source, which means even implausible numbers from untrusted sources can shift an estimate [3]. A separate randomized controlled field experiment by Jørgensen and Grimstad, published in IEEE Transactions on Software Engineering, found that irrelevant and misleading information changed effort estimates in a professional setting rather than only in the classroom [4].
The participants in these studies believed the outside number had not influenced them. That is the practical danger: the bias operates below the level where intent or diligence can catch it, and asking people to try harder does not switch it off.
- A confidence interval is not a rescue
A natural defense is to keep the point estimate but ask for a range around it. The same exercise examined whether that helps, comparing the worst-case figures from the low-anchor group against the best-case figures from the high-anchor group. The ranges did not overlap enough to compensate, and the difference remained significant (p < 0.05) [2]. Once the anchor has set the center, decorating that center with uncertainty inherits the same displacement. The fix has to change where the center comes from in the first place.
- Take the outside view
The alternative flips the order of operations. Instead of building the plan up task by task from the inside, you start from the outside: identify a reference class of comparable completed projects, collect the distribution of their actual outcomes, and then position your work inside that distribution before adjusting for its specifics. Reference class forecasting, developed by Bent Flyvbjerg from Kahneman and Tversky's work on the planning fallacy, bypasses optimism and strategic bias by anchoring on empirical results rather than on a negotiated scenario [5]. The method has three steps and was endorsed by the American Planning Association in 2005, which recommended never relying solely on conventional inside-view forecasts [5].
The uplift it implies is concrete. Flyvbjerg's rail-project data indicate that reaching a fifty-fifty chance of staying within budget requires adjusting the initial estimate upward by about 40%, and holding the chance of overrun to 10% requires an uplift near 68% [5]. Those figures come from transportation infrastructure rather than software, so import the logic and the discipline, not the percentages.

Figure by the author (illustrative).
- An estimation contract
Element | Point-estimate habit | Reference-class contract |
|---|---|---|
Starting point | First number in the room | Distribution of actual past outcomes |
Output | One precise figure | A range with stated probabilities |
Uncertainty | Added afterward as a buffer | Built into the distribution |
Commitment | Implicit in the number | A separate, explicit risk choice |
Review | Re-negotiated under pressure | Re-calibrated against new actuals |
Failure mode | Anchored and overconfident | Weak if the reference class is wrong |
The commitment decision belongs to the table's last row and should be made openly. Once the range exists, choosing to ship against its optimistic end is a risk decision a sponsor can own, rather than a fake certainty the estimator is forced to certify.
- Where the method genuinely breaks
A 2025 systematic review of reference class forecasting is blunt that the science is not settled. Critics argue that overruns in large projects often reflect complexity, pervasive uncertainty, and genuine error rather than a systematic optimism bias, and that pessimism bias and group estimation also occur [6]. The hardest practical problem is constructing the reference class itself. Make it too broad and the within-group variance makes any average misleading; make it too narrow and there are too few projects to support a distribution, a trade-off documented across the studies the review covers [6]. One cited study found the method outperforms earned-value and Monte Carlo approaches only when similarity between the new project and the reference class is sufficiently high, and for genuinely novel work such as first-of-a-kind engineering there may be no comparable class at all [6]. Historical outcomes also tell you what happened without explaining why, so a database of actuals does not replace causal understanding of the work.
None of that sends you back to a single guess. It means the reference class is a judgment you document — inclusion criteria, similarity checks, and how you adjust when no clean class exists — rather than a number a tool prints.
Next steps: stop requesting a lone date. Build a small database of completed work with actual effort and overruns, define the criteria a project must meet to enter the reference class, and present forecasts as ranges with probabilities. Keep the commitment as a separate decision, and feed every completed project's actuals back into the database so calibration improves over time.
References
- InfoQ, "What We Know and Don't Know About Software Development Effort Estimation" (summary of Jørgensen, IEEE Software; average overruns around 30%) — https://www.infoq.cn/article/software-development-effort-estimation
- "Anchoring and Adjustment in Software Estimation," controlled experiment (n = 23; means 6.8, 8.3, 17.4 months; high vs low p < 0.001; low vs control not significant) — https://static.aminer.org/pdf/20170130/pdfs/sigsoft/7qkzmx60mzlkvyj1ee8sa3riirtofuf9.pdf
- "Numerical Anchors and Their Strong Effects on Software Development Effort Estimates," three experiments with 381 software professionals — https://daneshyari.com/article/preview/461496.pdf
- Jørgensen and Grimstad, "The Impact of Irrelevant and Misleading Information on Software Development Effort Estimates: A Randomized Controlled Field Experiment," IEEE Transactions on Software Engineering (cited) — https://people.brunel.ac.uk/~csstmms/SAC2018\_final.pdf
- Flyvbjerg, "From Nobel Prize to Project Management: Getting Risks Right," arXiv:1302.3642 (preprint; reference class forecasting, APA endorsement, rail uplifts) — https://arxiv.org/pdf/1302.03642
- "Reference Class Forecasting: Promises, Problems, and a Research Agenda Moving Forward," Project Management Journal review, 2025 — https://www.tandfonline.com/doi/pdf/10.1080/09537287.2025.2578708
Series: developer-tools (8 articles)
- …
- 5Stop Asking for a Point Estimate: Forecast Software Work as a Calibrated Range This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article