| Vose Software

Industry: Agriculture and Food Supply Chain
Product: ModelRisk
Application: Optimizing Food Supply Chain Operations Under Uncertainty


The $61M Reserve a $49M Budget Thinks It Can Cover

A global agribusiness runs eleven critical supply-chain nodes — grain origination, ports, shipping lanes, processing clusters and distribution hubs — spread across four regions. Each carries an annual disruption probability and a severity if it fires. Add them up as independent risks and the 95th-percentile annual disruption cost is about $49M; that is the contingency reserve the planning spreadsheet implicitly sizes. Model the same eleven nodes with the correlation that actually governs them — nodes in the same region share a climate, political and logistics shock, so they tend to fail together — and the P95 jumps to $61.5M, with the P99 rising from $87.6M to $106.9M. The independence assumption does not make the company a little optimistic; it under-reserves the 1-in-20 year by $12.6M and the 1-in-100 year by $19.4M.

The reason is structural, not a tuning artefact. What exhausts a disruption budget is never the average year — it is the year a whole region goes down at once: a drought that hits every origination point in South America, or a chokepoint event that closes both the sourcing and the transit node in the same corridor. A model that treats those nodes as independent literally cannot generate that year often enough, so it prices a tail that does not exist.

The annual disruption cost is heavy-tailed by construction

The headline is the distribution of total annual disruption cost across all eleven nodes, with regional shocks correlated:

Distribution of annual supply-chain disruption cost

The mean annual cost is $17.6M, but the distribution is severely right-skewed: the P95 (Value-at-Risk) is $61.5M, the P99 is $106.9M, and the expected shortfall beyond the P99 — the average of the worst 1% of years — is $147.4M. There is a 12.2% probability the annual disruption bill exceeds $40M. A contingency set to the mean is overwhelmed roughly one year in eight; the number that matters for solvency is the tail, and the tail is an order of magnitude above the average.

Why a point estimate fails here

A single expected-cost figure throws away the two things that decide whether a contingency budget survives a bad year. First, the aggregation: disruption cost is a sum of frequency times severity across eleven nodes, and the sum has a much wider distribution than any single node. Second — and decisively — the correlation: regional nodes share a driver, so their disruptions are not independent draws. An expected value is identical whether the nodes are independent or perfectly coupled; only the distribution's tail tells them apart, and the tail is exactly what a reserve must cover.

Correlation is imposed, and it is what fattens the tail

The model places each node's disruption on a one-factor structure: a shared regional factor plus node-specific noise, so within-region nodes fail together while regions remain largely independent. The same regional factor also inflates severity, because a region-wide event drives shared scarcity and surge freight. Comparing the resulting cost distribution against the independence assumption makes the effect unmistakable:

Cumulative distribution of annual cost, correlated versus independent

The two curves share almost the same body — the means are close, $17.6M correlated against $15.5M independent — but they diverge in the tail, where the P99 is $106.9M correlated against $87.6M independent, a $19.4M (×1.22) understatement. The correlation is genuinely present in the simulation, not just asserted: the empirical pairwise correlation of disruption events is 0.24 within a region and 0.00 across regions, exactly the structure the one-factor copula imposes. Assume independence and you price the wrong tail.

Sizing the contingency reserve

The reserve question is "how much do we need to set aside to cover the year with a given confidence?" — and the answer depends on whether you respect the correlation:

Contingency-reserve sizing, correlated versus independent

To cover the year with 95% confidence requires $61.5M under the correlated model but only $48.9M if you assume independence — a $12.6M gap. At 99% confidence the requirement is $106.9M correlated against $87.6M independent, a $19.4M gap. The correlated curve sits consistently to the right: every confidence level demands a larger reserve once you admit that regions fail together. The independence assumption is not conservative — it is the cheaper, wrong answer.

Which region owns the tail

Tornado of each region's contribution to the P95 annual cost

Removing each region's disruptions in turn and measuring the drop in the P95 cost ($61.5M) ranks the regions by tail contribution. East Africa / Red Sea is the largest single driver (a $14.0M reduction in the P95 if its disruptions were eliminated), followed by SE Asia (-$11.8M) and South America (-$10.4M), with North America / Europe the smallest (-$7.0M). This is where correlation pays off as insight: the regions that own the tail are the ones whose nodes move together, so dual-sourcing and chokepoint diversification there buy more tail reduction per dollar than the same spend spread evenly.

What the model changed

  • The contingency reserve was raised from ~$49M to $61.5M at 95% confidence — closing a $12.6M under-reserve that the independence assumption had hidden.
  • Correlation was made an explicit modelling input, not an afterthought — the one-factor regional structure, verified at 0.24 within-region empirical correlation, is what generates the year a whole region fails.
  • Diversification spend was concentrated on East Africa / Red Sea and SE Asia — the two regions the tornado identified as owning the P95 tail — rather than spread thinly across all eleven nodes.
  • The board was given a tail, not an average — a P99 of $106.9M and an expected shortfall of $147.4M, framed as the solvency exposure a single expected-cost figure had concealed.

ModelRisk Functionality Used

  • Bernoulli disruption events with LogNormal severity at each of eleven nodes — the frequency-severity backbone of supply-chain risk aggregation.
  • A one-factor regional copula correlating within-region disruptions (and inflating severity in bad-region years), verified by the 0.24 within-region versus 0.00 cross-region empirical correlation.
  • Aggregation of correlated risks into a total annual-cost distribution whose tail — not whose mean — separates the correlated and independent worlds.
  • A contingency-sizing sweep reading the reserve required for each confidence level directly off the cost distribution.
  • Sensitivity tornado ranking the four regions by their contribution to the P95 cost, redirecting diversification spend to the regions that move together.

A supply-chain risk budget built on summed independent risks is a bet that no region ever fails as a whole. Monte Carlo simulation in ModelRisk replaces that bet with a correlated distribution — turning "we expect about $17.6M of disruption a year" into "the 1-in-20 year costs $61.5M and the 1-in-100 year costs $107M, and the independence assumption was hiding $19M of that tail."