Estimate accuracy and contingency: how sure can you be?
An estimate is a forecast made with incomplete information. The useful questions are how wide the range is at this stage, what sits inside it, and how much to hold for what you cannot yet see. This article covers accuracy ranges, the reasons projects overrun, and a better way to set contingency than a flat percentage.
By the Caprix AI team · · 9 min read
Key takeaways
- Accuracy depends on how much of the design is defined. Early estimates carry wide, lopsided ranges.
- Overruns are more common than underruns because unknowns add cost far more often than they remove it.
- Contingency is for risk within the agreed scope. It is not a fund for changes of mind.
- A risk-based contingency, built from probability and impact, is easier to defend and to manage than a flat percentage.
- Report a range and its basis, not a single number.
Accuracy follows definition
The accuracy of an estimate is set mainly by how much of the project has been defined when it is prepared. Professional bodies publish classifications to make this explicit. One widely used system, from AACE International for the process industries, grades estimates from Class 5 to Class 1.
| Class | Project definition | Typical purpose | Typical accuracy range |
|---|---|---|---|
| 5 | 0% to 2% | Concept screening | Low: −20% to −50%. High: +30% to +100% |
| 4 | 1% to 15% | Feasibility | Low: −15% to −30%. High: +20% to +50% |
| 3 | 10% to 40% | Budget authorisation | Low: −10% to −20%. High: +10% to +30% |
| 2 | 30% to 75% | Control or tender | Low: −5% to −15%. High: +5% to +20% |
| 1 | 65% to 100% | Check estimate or bid | Low: −3% to −10%. High: +3% to +15% |
Building work is not a process plant and the bands differ by sector, but two lessons carry over. The range narrows only as definition increases, not because more effort went into the arithmetic. And the range is lopsided: the upside is larger than the downside at every stage.
Why the range is lopsided
Cost cannot fall below zero but it can double. More practically, an estimate is a list of the things someone thought of. Anything forgotten can only add cost.
Research on large projects, notably by Bent Flyvbjerg and colleagues, found that cost overruns are the norm and gave two explanations. Optimism bias is the human tendency to plan around the best case. Strategic misrepresentation is the pressure to present a low figure so that a project is approved.
The remedy they propose is reference class forecasting: instead of building the estimate only from the inside, look at what similar completed projects actually cost against their early estimates, and adjust for that record. A database of completed projects is how a cost consultant applies the same idea in practice.
What contingency is for
Contingency is often treated as one pot. It is clearer as separate allowances, each with an owner.
| Allowance | Covers | Released when |
|---|---|---|
| Design development | Detail not yet drawn within the agreed scope | The design is completed |
| Construction risk | Site and delivery risks, such as ground conditions | The risk has passed |
| Escalation | Price movement to the dates of spend | Work is let on firm prices |
| Client change | Changes to the brief | Held by the client, outside the estimate |
Mixing these is how contingency disappears early. A design development allowance spent on a client change leaves nothing for the detail it was meant to cover.
From a flat percentage to a risk-based figure
A flat 5% or 10% is quick and familiar, and it tells nobody what is being covered. A risk-based allowance starts from a register. For each risk, estimate the probability that it occurs and the cost if it does. The expected value is the two multiplied together.
An illustrative register for a building project:
| Risk | Probability | Cost if it occurs | Expected value |
|---|---|---|---|
| Unforeseen ground conditions | 30% | RM400,000 | RM120,000 |
| Design development of services | 50% | RM300,000 | RM150,000 |
| Authority requirements | 20% | RM250,000 | RM50,000 |
| Material price movement | 40% | RM350,000 | RM140,000 |
| Total | RM1,300,000 | RM460,000 |
The total expected value, RM460,000, is the average outcome over many similar projects. It is not a safe allowance for this one: if the first two risks both occur the cost is RM700,000. So the allowance is set by how much confidence the client wants. Holding only the expected value leaves a real chance of exceeding it. Holding more buys a higher chance of staying within the allowance.
Larger projects take this further with Monte Carlo simulation, which samples each risk thousands of times and reports the result as a distribution. The figure at 50% confidence is called P50 and the figure at 80% is P80. The method is more elaborate but the idea is the same: contingency is a statement about confidence.
Reporting an estimate honestly
- State the base date of the prices and what the estimate includes and excludes.
- Give a range appropriate to the design stage, alongside the central figure.
- Say what the estimate was built from: which benchmarks, from where, adjusted how.
- Show contingency separately with what each allowance covers.
- Update it as the design develops, and explain each movement.
A client who is given a single number will remember it as a promise. A client who is given a range and its basis has been told the truth about an early estimate.
Where Caprix AI fits
Estimate gives an order-of-cost figure with an elemental breakdown from completed-project benchmarks, and states when those benchmarks have been adjusted from other states, so you know how firm the location basis is. Trends and the escalation question after an estimate support the escalation allowance. When the design firms up, the BOQ Builder is where the estimate becomes a priced bill.