---
title: "Planning Fallacy: Fix It With a Reference Class"
description: "Why time and cost estimates for projects run optimistic, and how reference-class forecasting corrects them, with a worksheet of 12 made-up projects to try."
url: "https://staging.khoi.co.uk/en/blog/du-an-tre-nhom-tham-chieu"
lang: en-GB
author: "Lê Quốc Khôi"
date: 2026-11-03
updated: 2026-11-03
topics: ["Data & analytics"]
tags: ["Planning fallacy", "Reference-class forecasting", "Project management", "Forecasting bias", "Forecasting"]
---

# Why project estimates run optimistic, and the reference-class fix

## Key takeaways

- The planning fallacy is the tendency to underestimate the time and cost of your own project. Buehler and colleagues found participants too optimistic across many tasks, academic and non-academic (Buehler et al., 1994).
- The root is the inside view: building a scenario from the details of this project instead of looking at how similar projects ended (Flyvbjerg, 2006). When participants were told to connect their predictions to relevant past experience, the optimistic bias disappeared (Buehler et al., 1994).
- Construction costs came in 44.7% above forecast on average for rail, 33.8% for bridges and tunnels and 20.4% for roads; Flyvbjerg sees no improvement in accuracy over 70 years of data (Flyvbjerg, 2006).
- Reference-class forecasting takes three steps. For UK roads, to leave only a 20% chance of overrun you add 32%; for metro and rail it is 57% (Flyvbjerg, 2006).
- The 12 jobs at the end of the post are hypothetical and every figure is mine. A 10-week plan was met on time in 1 of 12 jobs; promising 13 weeks leaves 2 jobs late.

The site crew of an interior-fit-out contractor in Can Tho adds up the tasks and gets 10 weeks for a new job. Over three years the crew has done 12 similar jobs, and only one finished on plan. That is a hypothetical example, and I made up the data to show the sum. Why is 10 weeks almost certainly too low, and what fixes it?

## What is the planning fallacy, and why does every team fall for it?

The [planning fallacy](https://staging.khoi.co.uk/en/glossary#planning-fallacy) is underestimating the time and cost of what you plan to do. Flyvbjerg, citing Lovallo and Kahneman, explains it as planners taking an inside view: focusing on the components of the project at hand instead of the results of similar projects already completed (Flyvbjerg, 2006).

Buehler, Griffin and Ross tested three hypotheses and report that the results supported each (Buehler et al., 1994). One: people underestimate their own completion times but not other people's. Two: when predicting, they focus on plan-based scenarios rather than past experience. Three: they explain earlier failed predictions with external, transient and specific causes, so past experience seems irrelevant this time.

Two details you can use at once. In their fourth study the optimistic bias vanished for people instructed to connect their predictions with relevant past experience. In the fifth, observers outside the task overestimated other people's times and made more use of past experience (Buehler et al., 1994). So before you fix a number, make the team write down the actual duration of a few recent comparable jobs, and ask someone outside the team to estimate independently. That is my inference, not a result of the paper; outsiders also overestimated, so their view is a reference point, not the answer. This post uses only the abstract of the paper, so there are no sample sizes.

## How much does optimism cost?

A great deal, and the strongest evidence is in infrastructure. Flyvbjerg pulls together data on transport projects: actual construction costs ran above forecast by 44.7% on average for rail, 33.8% for bridges and tunnels and 20.4% for roads; for rail passenger forecasts, actual ridership was on average 51.4% below forecast, and 84% of projects were out by more than 20% (Flyvbjerg, 2006). These are figures for cost and demand, not for construction time.

He argues this is not random noise. If poor data or models were the cause, errors would be roughly normally distributed around zero and shrink over time. In fact the distributions are significantly non-normal, the means differ from zero, and there is no improvement over the 70 years for which cost data exist. He concludes the problem is bias, caused by optimism bias and by strategic misrepresentation (Flyvbjerg, 2006). He says comparative research finds similar problems with new plants, IT systems and new products; the post has not checked that research.

## How does reference-class forecasting work?

In three steps, and none of them asks you to guess the specific mishaps of your project (Flyvbjerg, 2006):

1. Pick a class of similar past projects, broad enough to be statistically meaningful and narrow enough to be comparable.
2. Build the distribution of outcomes for that class from real data.
3. Place the new project in that distribution to get the most likely outcome.

Flyvbjerg retells a story Kahneman uses. A team designing a new school curriculum in Israel each estimated 18 to 30 months. An expert on the team was asked how long similar projects had taken; he said about 40% eventually gave up, and he could not think of any that finished in under seven years or took more than ten. The team ignored this and finished eight years later, and hardly anyone used the result (Flyvbjerg, 2006). The story comes from Flyvbjerg's telling and has not been checked against the original.

In the UK, the Department for Transport and the Treasury have used the method since 2004. The uplift added to a budget depends on the type of project and the risk of overrun the approver will accept (Flyvbjerg, 2006):

| Type of project | Uplift at the 50th percentile | Uplift at the 80th percentile |
|:--|--:|--:|
| Roads (172 projects) | 15% | 32% |
| Rail, including metro, light rail and high-speed (46 projects) | 40% | 57% |
| Bridges and tunnels (34 projects) | 23% | 55% |

In words: for roads there is a 50% chance that cost stays within the estimate plus 15%, and an 80% chance that it stays within the estimate plus 32%. For the Edinburgh tram line, the business case showed a base cost of £255 million with a 25% allowance, about £320 million in all. The reviewer, Ove Arup, applied the 57% uplift at the 80th percentile to arrive at £400 million; at the 50th percentile a 40% uplift gives £357 million. They judged the promoter's estimate to look optimistic (Flyvbjerg, 2006).

## How do you apply it to a small job?

Do the same three steps with your own data, and remember that a small sample gives only a rough estimate.

**A hypothetical example.** The Can Tho contractor has records of 12 finished jobs. I made up these figures. Overrun against plan, sorted low to high (%): 0, 5, 8, 10, 12, 15, 18, 20, 25, 30, 40, 55.

- Pick the class: 12 interior jobs of the same type and similar size, done by this same crew.
- Build the distribution: half the jobs overran by 15% or less; 10 of 12 overran by 30% or less.
- Place the new job in it: a 10-week plan becomes 11.5 weeks for a 50% chance of being on time, and 13 weeks for roughly 80%.

Only one job in 12 finished on plan, so promising 10 weeks means being on time about once in 12. Suppose the contract carries a penalty of VND 5m for each week of delay (an assumption). The average overrun is 238 divided by 12, about 19.8%, nearly 2 weeks: promise 10 weeks and you lose about VND 9.9m in penalties per job on average. Promise 13 weeks and only two jobs run late, one by 1 week and one by 2.5 weeks: an average of about VND 1.5m.

The cost of promising 13 weeks is that you may lose the bid to a rival who promises 10. When winning depends on the low number, the problem is no longer the team's psychology. This resembles what Flyvbjerg calls strategic misrepresentation; he writes about public projects, and I am applying it to bidding. He says the remedy is accountability: forecasters carry the risk of their forecasts and an independent body reviews them (Flyvbjerg, 2006). He also warns that a large reserve tends to be spent, so tight contracts must go with it.

Two other posts on this site look at planning and forecasting from different angles. [Feasibility checks before a business plan](https://staging.khoi.co.uk/blog/kiem-tra-kha-thi-truoc-ke-hoach) asks whether an idea is worth pursuing. [Starbucks forecasting through COVID](https://staging.khoi.co.uk/blog/starbucks-du-bao-giua-covid) covers the case where history can no longer be extrapolated. The full list of biases is in the framework [Checklist of forecasting biases](https://staging.khoi.co.uk/en/framework/forecast-bias-checklist).

## Going deeper: what do these two sources establish?

Flyvbjerg (2006) appeared in *Project Management Journal*, volume 37, issue 3, pages 5 to 15. The post uses the author's own copy posted online, numbered as 32 pages and laid out differently from the print version, so it gives no page numbers. The paper describes the method for transport projects, identifies the 2004 Department for Transport guidance, which he wrote himself with COWI, as its first practical use, and reports that none of the several hundred forecasts he examined was a reference-class forecast. What he attributes to Kahneman, Tversky and Lovallo is taken second-hand through him, not read directly.

First limit: the data are costs and demand for public infrastructure projects, not construction time for fit-out work. An uplift of 15% or 32% cannot be used for fit-out work in Can Tho; applying the same method to duration is my own step. What carries over is the method, and the data have to be yours. Second: Flyvbjerg says the method bypasses both optimism bias and strategic misrepresentation because it goes straight to outcomes, but he adds that where the motive is misrepresentation you also need accountability. Third: for Buehler and colleagues the post uses only the abstract, because the paper is not open access. The abstract has no sample sizes, scales or numerical effect sizes, so only the qualitative conclusions can be used.

## Frequently asked questions

### How does the planning fallacy differ from optimism bias and from deliberate misreporting?

Flyvbjerg separates two causes. Optimism bias is self-deception, with no intent. Strategic misrepresentation is deliberately understating costs and overstating benefits so that your project is approved ahead of others (Flyvbjerg, 2006). The planning fallacy is optimism bias applied to time and cost. The same low number can come from either cause, and each needs a different remedy.

### How many past projects do I need for a reference class?

Flyvbjerg gives no minimum. He asks only that the class be broad enough to be statistically meaningful and narrow enough to be truly comparable with the new project (Flyvbjerg, 2006). His data hold 172 road projects, 46 rail projects and 34 bridges and tunnels. With 12 jobs from one small contractor, treat the percentiles as rough and update them whenever a project finishes.

### If I know about the planning fallacy, can I avoid it?

Not reliably. Flyvbjerg cites Kahneman and Tversky: knowing about a cognitive illusion does not by itself make perception more accurate, but it can help you spot situations where you should set your impressions aside and weigh the evidence (Flyvbjerg, 2006). In the fourth study of Buehler and colleagues, the bias disappeared when participants were told to link their predictions to relevant past experience (Buehler et al., 1994). You need a compulsory procedure, not just understanding.

### Is a flat 20% contingency on every project enough?

There is no basis for one flat number. The uplifts in Flyvbjerg's study vary by type of project and by the risk you accept: for roads 15% at the 50th percentile and 32% at the 80th; for metro and rail 40% and 57% (Flyvbjerg, 2006). He also warns that a large reserve tends to be spent simply because it is there, so tight contracts must go with it.

## References

- 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*, 67(3), pp. 366–381. https://doi.org/10.1037/0022-3514.67.3.366
- Flyvbjerg, B. (2006) ‘From Nobel Prize to project management: getting risks right’, *Project Management Journal*, 37(3), pp. 5–15. https://doi.org/10.1177/875697280603700302

---

Canonical HTML version: https://staging.khoi.co.uk/en/blog/du-an-tre-nhom-tham-chieu