Contingency Sizing: Why 5% of EPC Gives the Riskiest Project the Least Protection
Take two projects with identical $230m budgets. Project A has $190m of firm-priced EPC and $20m of interconnection under an executed agreement. Project B has $150m of EPC and $60m of interconnection still resting on a cluster study.
Apply the conventional contingency — five percent of EPC — and A gets $9.5m while B gets $7.5m.
Run the actual cost distributions and A's contingency sits at the P84 confidence level. B's sits at P55.
The convention gives the riskier project less money and a coin-flip level of protection, for exactly the reason that makes it riskier: its uncertainty is concentrated in a line the percentage is not applied to.
This post closes Series E with the method that fixes it — decomposed contingency, sized against each line's own estimate class — and with the portfolio effect that stops it becoming absurdly conservative.
ℹ️ Note: All distributions and figures are labelled assumptions. The method is the point; the ranges should come from your own estimate classification and risk register.
What Is Contingency For?
Cost that is expected to be incurred but cannot yet be attributed to a specific item — the aggregate of known unknowns. It is not a buffer against things nobody thought of, and it is not a negotiating margin. It is the part of the estimate that exists because estimates have distributions.
That framing matters because it determines the method. If contingency is a statistical property of the estimate, the way to size it is to characterise the estimate's distribution — not to apply a percentage inherited from a previous deal.
The AACE Framework
The standard classification ranks estimates by how much of the project is actually defined, and attaches an expected accuracy range to each level.
| Class | Project definition | Typical purpose | Accuracy range |
|---|---|---|---|
| 5 | 0–2% | Concept screening, feasibility | −50% to +100% |
| 4 | 1–15% | Feasibility studies | −30% to +50% |
| 3 | 10–40% | Budget authorisation, funding approval | −20% to +30% |
| 2 | 30–75% | Control baseline, bid evaluation | −15% to +20% |
| 1 | 65–100% | Check estimates, bid validation | −10% to +15% |
Two features of that table do all the work in this post.
The ranges are enormous at the top. A Class 5 estimate can be out by a factor of two on the upside. A project at financial close does not have Class 5 estimates for everything — but it rarely has Class 1 estimates for everything either.
Every range is asymmetric. As the classification states directly, "the potential for cost overrun (positive side) is always greater than the potential for cost underrun (negative side)," reflecting that capital projects tend to exceed estimates more often than they stay within them, especially where definition is low.
That asymmetry is why a symmetric contingency is systematically short, and it is the same observation the network upgrades post made from the interconnection side.
The Problem With a Single Percentage
A project at financial close does not have one estimate. It has several, at different classes.
| Line | Amount | Effective class | Upside range |
|---|---|---|---|
| EPC, firm price signed | $190m | Class 1 equivalent | +6% (change orders) |
| Owner-supplied equipment | $12m | Class 2 | +20% |
| Interconnection and upgrades | $20m | Class 3 | +50% |
| Land, permitting, development | $8m | Class 2 | +20% |
A signed fixed-price EPC contract is not subject to estimation error in the ordinary sense; its exposure is change orders and claims, which is a much narrower distribution. Interconnection under a cluster study is the opposite — and as the network upgrades post established, the withdrawal exemption thresholds of 25% and 100% exist precisely because regulators expect estimates to move that far.
Applying one percentage to a budget composed of a firm contract and an unfixed estimate is applying the wrong distribution to both.
The Naive Decomposition, and Why It Overshoots
The obvious fix is to size contingency line by line and add it up:
EPC $190m × 6% = $11,400,000
Owner-supplied equipment $12m × 20% = $2,400,000
Interconnection $20m × 50% = $10,000,000
Development $8m × 20% = $1,600,000
-----------
Sum of individual maxima = $25,400,000
That is $25.4m of contingency on a $230m budget — eleven percent, and far more than any lender or sponsor would accept.
It is also wrong, because it assumes every line reaches its worst case simultaneously. They will not. Some will come in under, and the aggregate distribution is narrower than the sum of the individual ranges — the same portfolio effect that makes a diversified position less volatile than its components.
The Correct Method: Combine the Distributions
Run the lines as distributions and take a percentile of the total.
On Project A, with triangular distributions over the ranges above and the mode at the base estimate:
| Confidence | Total cost | Contingency | As % of base |
|---|---|---|---|
| P50 | $234,636,374 | $4,636,374 | 2.0% |
| P80 | $238,749,315 | $8,749,315 | 3.8% |
| P90 | $240,929,525 | $10,929,525 | 4.8% |
| P95 | $242,697,412 | $12,697,412 | 5.5% |
The P80 contingency is $8.75m — about a third of the naive sum of maxima. The portfolio effect is doing most of the work, and it is the reason a decomposed method does not produce absurd numbers.
Note also what the conventional figure turns out to be. Five percent of EPC on Project A is $9.5m, which on this distribution is the P84 level. The convention is not wrong here. It is a P84 number that nobody knew was a P84 number, which is a different criticism.
That distinction is worth dwelling on, because it explains why conventions survive. A rule of thumb that has been applied to hundreds of projects has been implicitly calibrated by the ones that failed: if five percent had been routinely inadequate, the convention would have moved. So on the typical project — one whose budget looks like Project A, dominated by a firm-priced EPC — the convention encodes real accumulated experience and lands in a sensible place.
The failure is not that the number is wrong on average. It is that the number cannot tell you when you are not average, and the whole point of a convention is that it is applied without asking. A project whose budget composition has shifted — because interconnection costs have risen faster than equipment costs, which is precisely what Series C documented — is exactly the case a historically calibrated rule handles worst, because the calibration came from a period when the mix was different.
Now Change the Mix
Project B has the same $230m budget with $150m of EPC and $60m of interconnection.
| Confidence | Contingency | As % of base |
|---|---|---|
| P50 | $6,078,583 | 2.6% |
| P80 | $16,050,090 | 7.0% |
| P90 | $21,091,758 | 9.2% |
| P95 | $24,742,817 | 10.8% |
Project B needs $16.05m to reach P80 — nearly twice Project A's requirement, on an identical total budget, because its uncertainty is concentrated in the line with the widest distribution.
And the conventional method gives it $7.5m, because five percent of a smaller EPC is a smaller number.
Project A conventional $9,500,000 → P84
Project B conventional $7,500,000 → P55
Same convention, same budget, and the riskier project ends up at a coin flip. The percentage is applied to the line that is least uncertain, so the more of a budget that sits outside the EPC contract, the less protection the convention provides — which is exactly backwards.
Contingency Is Not Management Reserve
A distinction that is often collapsed and should not be, because the two are controlled differently.
Contingency covers known unknowns: identified risks whose occurrence or magnitude is uncertain. A change order that has not been raised yet, a quantity that will differ from the take-off, an escalation that has not settled. It belongs inside the project budget and is drawn by the project.
Management reserve covers unknown unknowns: the residual that no risk register captures. It sits outside the project budget, is controlled at a level above the project, and is released only by a deliberate decision.
The distinction matters in project finance because only one of them is normally financed. Contingency is a line in the sources and uses statement, funded by the facility on the same terms as any other cost. Management reserve, where it exists at all, is usually a sponsor-level provision outside the financing — which means it is equity, at equity's cost, and it does not appear in the debt sizing.
Two practical consequences.
A financed contingency is cheap; a sponsor reserve is not. Contingency funded 75% by debt costs the sponsor a quarter of each drawn dollar. A cost overrun funded by the sponsor under its cost overrun undertaking costs a hundred cents. That is an argument for sizing the financed contingency properly rather than relying on the undertaking, and it runs directly against the instinct to keep the financed budget lean.
The distribution analysis sizes contingency, not reserve. A Monte Carlo over identified budget lines characterises known unknowns by construction. It says nothing about a tariff arriving mid-construction, a contractor insolvency, or a permit being challenged — which are the events that actually destroy projects. Those belong in the change-in-law and force majeure analysis from Series D, not in the contingency calculation, and a sponsor that believes its P80 contingency covers them has misread what was simulated.
Who Controls It, and When Is It Spent?
Contingency in a financed project is not the sponsor's to spend freely, and the control mechanism shapes behaviour in a way worth anticipating.
Drawings against contingency above a threshold typically require the consent of the lender or a certificate from the technical adviser. That is reasonable — contingency is lenders' money as much as the sponsor's — and it produces a specific dynamic during construction.
Early contingency drawdown is a leading indicator and is treated as one. A project that has consumed forty percent of its contingency at twenty percent completion has told its technical adviser something important, and the cost-to-complete test will start to bind long before any milestone is missed. Tracking contingency drawn against percentage complete is the single most useful construction monitoring metric, and it is reported far less often than schedule variance.
There is a perverse incentive at the end. Unspent contingency at completion reduces the final drawdown, which reduces the debt, which — since the equity contribution is fixed by the gearing ratio applied to actual cost — returns money to nobody in particular and slightly reduces leverage. Some sponsors respond by finding uses for residual contingency in the closing months. Lenders respond by requiring that unspent contingency be cancelled rather than drawn, which is the right answer and should be in the documents.
The related drafting point is what happens to contingency that is reallocated. A budget line that comes in under and releases funds should be able to support a line that comes in over, and a rigid line-by-line budget that forbids reallocation forces a contingency drawing where none was economically necessary. Most facilities permit reallocation within limits; the limits are worth reading.
What About Schedule Contingency?
The other kind, and it is not interchangeable with the money kind.
Schedule contingency is float — time built into the programme beyond the critical path requirement. It protects the dates, and Series D established how many dates a project actually has: guaranteed substantial completion under the EPC contract, the PPA's guaranteed commercial operation date, the longstop date, and the tax credit placed-in-service deadline.
Two things make schedule contingency harder to size than cost contingency.
Float is consumed by events that do not cost money. A permit appeal, a grid outage postponing energisation, a weather window missed. None of those draws on the cost contingency and all of them consume the buffer protecting four different deadlines with four different consequences.
Converting time to money is not linear. The delay LD analysis showed that the cost of delay steps at each date rather than accruing smoothly — liquidated damages begin, then a cap exhausts, then a termination right arises, and the placed-in-service deadline is a cliff with no damages attached at all because there is no counterparty. A model that values float at a daily rate has assumed away the structure that makes delay dangerous.
The practical instruction is to size cost and schedule contingency separately, and to convert the schedule buffer into money using the date ladder rather than a flat daily figure. A project with sixty days of float against a longstop date and thirty days against a placed-in-service deadline is exposed on the second, whatever the first says — and it is the second that has no remedy.
What This Means in Practice
Three consequences worth carrying.
Projects in congested interconnection zones are systematically under-provisioned. Series C established that network upgrade costs are the largest and least predictable line in many renewable budgets and that they move with other parties' decisions. Those are precisely the projects whose contingency a percentage-of-EPC convention understates.
A wrapped EPC deserves a lower contingency and usually does not get one. The sources and uses post noted that a wrap premium buys interface risk transfer. If the premium is paid and the contingency is not reduced, the risk has been paid for twice.
The right output is a confidence level, not an amount. "We have carried 5% contingency" is not a statement about protection. "Our contingency is a P80 against the decomposed distribution" is, and it is directly comparable across projects in a way the percentage is not.
How Do You Build This in Excel?
As a line-by-line distribution with a simulated total, and with the confidence level of whatever contingency you actually carry reported as an output.
The decomposition
For each budget line:
Base_Estimate
Estimate_Class (AACE 1–5, or firm-priced)
Low_%, High_% (from the class, or the risk register)
Distribution triangular is adequate; mode at base
The simulation
For each of N trials:
Total(n) = Σ over lines of Base × (1 + TRIANGULAR(Low, High, 0))
Sort, then:
PF_ContingencyP50 = PERCENTILE(Total, 50%) − Base
PF_ContingencyP80 = PERCENTILE(Total, 80%) − Base
PF_ContingencyP90 = PERCENTILE(Total, 90%) − Base
Excel can do this with a data table over RAND()-driven rows, or more cleanly with a small simulation in the workbook's own scripting. Ten thousand trials is ample; the answer is stable well before that.
The output that changes decisions
PF_ContingencyCarried = the amount in the budget
PF_ImpliedConfidenceLevel = percentile of (Base + Carried)
Project A: $9,500,000 → P84
Project B: $7,500,000 → P55
Publishing the implied confidence level is the single most useful thing this analysis produces. It converts a number everyone argues about into a statement about protection that a credit committee can actually price.
The cross-check against the cost-to-complete test
Contingency carried $7,500,000
Cost increase that fails cost-to-complete = contingency remaining
→ At what percentage complete does the buffer run out?
The construction post identified the cost-to-complete test as the real completion covenant. Contingency is what stands between a cost increase and a failed test, so the two analyses belong together.
ℹ️ Note: Take the ranges from the estimate class and the project's own risk register, not from this post. A firm-priced EPC with a strong contractor and a tight scope may have a +3% change order distribution; one with a weak scope definition and an aggressive price may be +15%. The method survives a change in inputs; a convention does not.
To build the decomposed contingency, the simulation and the implied confidence level, prompt Dezzmond with your budget lines and estimate classes.
What Do Sponsors and Lenders Actually Check?
- What confidence level does the carried contingency represent?
- Is contingency decomposed by line, or applied as a single percentage?
- What share of the budget sits outside the firm-priced EPC contract? That is where the uncertainty is.
- Is the interconnection estimate an executed agreement or a cluster study?
- Are the distributions asymmetric? They should be; every AACE class range is.
- Has the portfolio effect been applied, or is the total a sum of individual maxima?
- At what percentage complete does the contingency run out under an adverse case?
Frequently Asked Questions
What is contingency for?
Costs that are expected to be incurred but cannot yet be assigned to a specific item — the aggregate of known unknowns. It is a statistical property of the estimate, which is why it should be sized from the estimate's distribution rather than from a convention.
Why are cost estimate ranges asymmetric?
Because projects overrun more often than they underrun. Every AACE class range reflects this, from −10%/+15% at Class 1 to −50%/+100% at Class 5, and a symmetric contingency is therefore systematically short.
Why not just add each line's worst case?
Because that assumes every line reaches its maximum at once. On the worked example the sum of maxima is $25.4m while the P80 of the combined distribution is $8.75m — the portfolio effect accounts for most of the difference.
Why does 5% of EPC under-protect a risky project?
Because the percentage is applied to the line with the narrowest distribution. The more of a budget that sits outside a firm-priced EPC contract — interconnection, owner-supplied equipment, development — the less protection the convention provides.
Is contingency the same as management reserve?
No. Contingency covers known unknowns and sits inside the financed budget. Management reserve covers unknown unknowns, sits outside it, and is therefore sponsor equity at a hundred cents rather than debt at seventy-five. A distribution analysis sizes the first and says nothing about the second.
Does contingency cover schedule risk?
No. Schedule contingency is float, consumed by events that cost no money, and its value has to be computed against the date ladder — guaranteed completion, longstop, placed-in-service — because the cost of delay steps at each date rather than accruing at a daily rate.
What should be reported instead of a percentage?
The implied confidence level. "Our contingency is a P80 against the decomposed distribution" is comparable across projects and priceable by a credit committee; "we carried 5%" is not.
Closing: Series E, and the Number That Absorbs Every Other Error
Thirteen posts on project finance mechanics, and they end on the line that catches whatever the other twelve got wrong.
Contingency is the last defence in the capital budget. The sources and uses statement balances by adjusting equity; the cost-to-complete test forces the sponsor to fund a shortfall; and between those two sits contingency, which is supposed to absorb the ordinary variation so that neither mechanism has to be invoked.
Sizing it by convention means sizing the last defence without knowing how strong it is. And the convention has a specific and perverse failure: it scales with the firm-priced portion of the budget, so a project whose cost risk is concentrated in an unfixed interconnection estimate — which describes a great many projects in congested zones right now — receives the least protection precisely because it has the most exposure.
The fix is not more contingency. On Project A the convention was already generous at P84. The fix is to know which case you are in, which requires decomposing the budget by estimate class, combining the distributions rather than adding the maxima, and reporting a confidence level instead of a percentage.
That is a morning's work in a spreadsheet, and it converts the most-argued-about line in a capital budget into one of the few numbers in this series that can be compared directly across projects.
Series F begins next, on returns and valuation: what the different IRRs actually measure, and why the one most often quoted is the one that says least.
Sources: AACE International — 18R-97 Cost Estimate Classification System · Cenex — AACE Cost Estimate Classification System (Class 1–5) Explained · Australian Government Department of Infrastructure — Guidance Note 3B: Deterministic Contingency · Production Planning & Control — From Reference Class Forecasting to Decomposed Contingency in Infrastructure Projects