| Vose Software

Industry: Construction and Infrastructure
Product: ModelRisk
Application: Managing and Mitigating Cost Overruns in Large-Scale Infrastructure Projects


The 8% Contingency That Was Really a Coin Flip

The deterministic estimate said $500M. The owner added 8% contingency for an approved budget of $540M. The probabilistic rebuild in ModelRisk showed the mean simulated cost at roughly $544M, the P90 at $604M, and a 51% probability that the project would breach the $540M budget. Eight percent is the construction-industry default. On this project, with this risk profile, the right number was closer to 16%.

Cost-overrun analysis in infrastructure is not really about cost. It is about correlation. Steel, fuel and equipment costs do not drift independently — they move together when oil prices spike. Material costs and labor costs are linked through schedule extension. Treat the cost stack as nine independent triangular distributions and Monte Carlo will give the right mean and badly under-state the tail.

Bottom-up stack, ten components

The bottom-up estimate decomposes into ten cost lines:

Line Most-likely ($M) Distribution CoV
Steel & rebar 90 LogNormal + 6% shock 18%
Concrete & cement 75 LogNormal 10%
Earthworks & aggregate 45 LogNormal 16%
Mechanical/Electrical 70 LogNormal 14%
Direct labor 110 LogNormal 12%
Equipment & fuel 55 LogNormal + 6% shock 20%
Indirect (PM, overhead) 40 LogNormal 8%
Permits & insurance 15 Triangular 5%
Schedule extension (0) LogNormal months × $3.2M/mo
Change orders (0) Beta of contract value

The cost-component distributions are LogNormal, not Normal. The reason is simple: a cement cost cannot go below zero; the upside (a 20% price shock) is observed more often and is larger than the downside (a 10% drop). LogNormal captures both. Normal allows negative costs and symmetric tails that the industry data refuses to give back.

The two single biggest moving parts — steel and equipment & fuel — also carry a 6% probability of a 25–40% price shock event (a supplier failure, a geopolitical fuel-price jump, a strike). A shock-augmented LogNormal captures what a plain LogNormal would smooth out.

The copula that owners forget

The classic in-house cost model treats every line as independent. The real world has clear correlation structure:

  • Steel and concrete and aggregate all move with raw-materials inflation — pairwise correlations of about 0.50.
  • Fuel correlates with the heavy-equipment line (0.70) and with materials transport (0.35).
  • Direct labor and indirect cost share the same project-overhead exposure (0.45).

ModelRisk's Gaussian-copula construction lets the team specify these as rank correlations that hold across the entire joint distribution — most importantly in the tail. Independence would have given an annual mean of $544M with a much tighter spread. The copula widens the P90 by roughly $32M, because in years where steel is up, fuel is usually up too. That is exactly the kind of compounding risk single-point cost estimators do not see.

Cumulative cost trajectory — $500M transit package, 36-month build

The deterministic S-curve plots cleanly to $500M at month 36. The simulated P10–P90 envelope tells the real story: by month 20 it already straddles the approved $540M budget cap, and by month 36 the P90 trajectory is around $604M — $64M above the budget line. The fan widens through construction because every monthly draw of steel, fuel and labor compounds the gap. An owner setting contingency by industry rule-of-thumb (8% on top of the point estimate) is implicitly betting that the P50 trajectory holds — and the simulation says the P50 trajectory sits essentially on the budget cap, with a 51% chance of finishing above it.

How much contingency for what confidence?

The contingency-sweep curve plots P(total cost ≤ budget) directly against the contingency percentage, for both the unhedged baseline and the hedged variant — letting the board read off the confidence level for any candidate number:

P(total cost within budget) vs contingency %

Reading off the unhedged curve: an 8% contingency ($540M) buys ~49% confidence, a 20% contingency ($600M) buys ~89% confidence, and a 21% contingency ($604M) clears the 90% line. The 80% probability-of-meeting-budget target — the standard owner board-approval threshold — needs roughly 16%, not 8%. The hedged curve sits visibly above the unhedged one at every contingency level, with the gap widening above 15%: the hedging program buys more probability than it costs everywhere, but the marginal value of hedging is highest exactly where the unhedged curve is still climbing through the high-stakes 70–85% confidence band.

What drives the tail

Tornado: drivers of total cost variability

The top three contributors to total cost variability are change orders, direct labor and steel & rebar. Change orders — modeled as a Beta scaled to 20% of contract value, mean 8% — is the single line the owner most directly controls (through design freezes and scope discipline). The team's mitigation priorities follow the tornado, not the deterministic spreadsheet's largest absolute line items.

Cost and schedule fail together

Cost overrun and schedule extension are not independent draws — they share the copula factors (steel / fuel / labor) and are linked structurally through the per-month extension charge. The joint-density chart makes the dependence visible:

Joint cost-and-schedule outcome — where the tail actually lives

The joint quadrant — total cost above $540M and schedule more than 6 months over plan — is roughly 2–3× more likely than independence would predict. That is the trial where everything goes wrong together: commodity shock, labor squeeze, schedule extension, idle equipment, LD exposure. An owner who books contingency for cost overrun and a separate reserve for schedule extension is double-paying for the marginal trials and severely under-providing for the joint trials. The copula-aware view says: book a single reserve sized to the joint quadrant, not two independent ones.

What the model changed

  • Contingency reset to $100M (20%) from $40M (8%) for board approval — the P80 reading.
  • Steel and fuel hedging programs initiated — reducing P90 cost by roughly $19M.
  • Change-order budget split off into its own owner-controlled $40M pool, tracked separately from contractor contingency.
  • Two states' bids reviewed. The model flagged that the deterministic spreadsheet had systematically under-priced packages where the copula tail was deepest; two pending bids were uplifted by 4–6%.

ModelRisk Functionality Used

  • LogNormal + shock-mixture distributions for steel and fuel, capturing both day-to-day volatility and the rare but real 25–40% price jumps.
  • Gaussian copula correlation structure across ten cost lines — the $32M extra P90 spread that independence assumption hid.
  • Beta-of-contract-value for the change-order pool, calibrated to the owner's historical change-order rate of 8% mean.
  • Tornado on absolute cost-variability contribution — the ranking that re-ordered mitigation priorities to change-order discipline first, hedging second, schedule-compression third.
  • Contingency curve as a deliverable for the capital committee, replacing the single "10% contingency" line with a P50/P80/P90 menu.

Cost overrun is not a number. It is a distribution that depends as much on what moves together as on what moves. Monte Carlo simulation in ModelRisk is what makes both visible — and once they are visible, contingency stops being a percentage rule-of-thumb and starts being a defensible confidence level.