A cost percentile does not describe the size of the overrun beyond it

A percentile identifies a threshold in a modeled cost distribution. It does not tell management how expensive the outcomes above that threshold could be. Review both the probability of exceeding a proposed budget and the size of the additional funding needed when that happens.

The distinction matters when two projects appear to have the same confidence level. One may occasionally exceed the budget by a manageable amount. Another may rarely require several times the approved funding. Treating their percentile labels as equivalent can hide a material difference in the decision.

The Basel Committee's market-risk terminology distinguishes a value-at-risk threshold from expected shortfall, which examines losses in the tail. The management-cost examples here use that general distinction between a cutoff and tail severity. They do not apply bank capital rules to an ordinary business project.

Consider two cost models, each with 100 equally weighted outcomes. In both, 94 outcomes cost $100,000 and one costs $150,000. In Model A, the remaining five outcomes cost $160,000 each. In Model B, they cost $500,000 each.

Define the 95th percentile here as the smallest cost at which cumulative probability reaches 95%. Under that convention, both models have a P95 of $150,000. Both also assign a 5% probability to exceeding a $150,000 budget. Those statements alone make the models look alike.

The conditional cost when the budget is exceeded is $160,000 in A and $500,000 in B. The required additional funding is therefore $10,000 in A and $350,000 in B. The same modeled exceedance frequency accompanies very different consequences.

The expected amount above the $150,000 budget across all outcomes is $500 in Model A: 5% multiplied by $10,000. It is $17,500 in Model B: 5% multiplied by $350,000. These unconditional averages include a zero excess in the other 95% of outcomes.

The average excess conditional on an overrun remains $10,000 and $350,000. A finance team preparing a response to an actual overrun needs that conditional view as well. Holding only the unconditional average does not establish that enough money will be available when the adverse event occurs.

Expected total cost is another measure again: $103,500 for A and $120,500 for B. Present each measure with its definition. A report titled expected shortfall should clarify whether it means tail total cost, excess above a threshold, or another explicitly defined loss quantity.

A tail average estimated from only a few adverse observations can itself be unstable. Show how many outcomes support it and whether they come from historical cases or a probability model. One exceptional observation may dominate the average; removing it requires an evidence-based reason, while retaining it requires an explanation of why that event belongs in the decision's scope.

Spreadsheet and statistical packages can interpolate percentiles differently in finite samples. The smallest-cumulative-probability convention used above is explicit, so the result can be reproduced. A different interpolation rule may place the reported percentile between $150,000 and the next outcome.

When many outcomes equal the percentile, calculating the average of values strictly above it may include less than the intended tail probability. A formal tail average can require partial weighting at the boundary. Ask the analyst to document how ties, weights and interpolation are handled before comparing models or periods.

Retain the underlying outcomes or a sufficiently detailed weighted distribution. A single P95 number cannot later be used to reconstruct the tail. If the source model provides only a few percentiles, say which severity conclusions cannot be supported from that limited output.

In a real project, identify the events behind the high-cost paths. They may involve a replacement supplier, a repeated commissioning stage, a regulatory redesign or a delay that extends several commitments. The relevant question is whether the model includes the complete cost consequence and whether the event is credibly represented.

Check for double counting. If a delay scenario already includes extra supervision and equipment hire, a separate generic contingency for those same costs may inflate the tail. Conversely, a cost model that excludes delayed revenue can understate the related funding pressure even when its project-cost distribution is complete.

The largest simulated outcome is not automatically an absolute maximum. It may be the largest observation drawn from a distribution with a longer tail, or it may reflect an artificial upper bound. Review that distinction before management treats the displayed worst case as a firm spending ceiling.

For each material adverse case, place the additional payments on a calendar and identify the earliest warning. A $350,000 requirement due immediately creates a different financing problem from the same amount spread over two years. Compare the dates with available facilities, approval lead times and funds that are genuinely usable.

Consider whether a practical option reduces severity: staged commitments, a second qualified supplier, a narrower initial scope or a defined stopping point. Include the cost and limitations of that option. A plan to stop after a failure does not save amounts already contractually committed before the failure becomes visible.

Keep management's risk preference separate from the model output. A team may accept a larger expected cost to reduce a severe funding exposure. Explain that tradeoff with the modeled amounts and operating constraints, rather than implying that one percentile automatically selects the right budget.

When scope, suppliers or timing change, compare the new threshold and tail severity with the earlier version. A lower P95 can coexist with a worse extreme tail if probabilities or event consequences have shifted. The approval record should show that movement instead of reporting only the favorable percentile change.

A concise decision table can include the proposed budget, exceedance probability, tail average, conditional additional funding and the dates that funding could be needed. Those measures give management enough context to discuss the remaining exposure and choose a response without treating a probability label as the complete financial case.