Industry: Agriculture and Food Supply Chain Product: ModelRisk Application: Optimizing Food Storage Strategies Under Uncertainty
A distributor of chilled perishables budgeted spoilage at a flat 9% of consignment value, the figure a planner gets by running average temperatures and average dwell times through a spoilage curve. The real cold chain does not run on averages. Each consignment crosses four handling legs — farm pre-cool, line haul, distribution centre, retail back-room — and a weak refrigeration unit or a hot lane pushes every leg toward a temperature breach at once. When the distributor rebuilt the chain as a 200,000-trial Monte Carlo model in ModelRisk, the picture changed: the mean loss is 6.8%, the median 3.9%, but the P90 is 18% and the P95 is 24% — and 25% of consignments lose more than a tenth of their value.
The plug-in average was wrong in two directions at once. Spoilage rises with accumulated thermal dose along a saturating curve, so feeding it average inputs overstates the typical loss (Jensen's inequality pushes the deterministic answer to 9.3%, well above the simulated 6.8% mean) while completely missing the right tail of consignments that cook. A point estimate is not just imprecise here; it is biased high on the body and blind on the tail.
About 23% of consignments arrive essentially clean, the body of the distribution sits in single-digit percentages, and a long right tail runs past 30%. The black line marks the 9.3% deterministic plug-in — visibly to the right of the 6.8% true mean yet far short of the P90 of 18%. One number cannot be both the typical loss and the planning loss; the distribution shows why the distributor needs both.
The temptation is to treat four handling legs as four independent risks that wash out — by the law of large numbers, a sum of independent shocks tightens around its mean. Cold-chain failure does not behave that way, because the legs share a common cause.
Each consignment carries a single chain-reliability factor: an aging reefer, a poorly maintained unit, or a hot summer lane raises the breach probability and excursion severity on all four legs simultaneously. The model represents this with one Beta-distributed reliability draw per consignment that modulates every leg, so breaches are positively correlated rather than independent. A consignment that breaches at line haul is more likely to have breached at the DC too. That shared factor is what keeps the loss distribution skewed and fat-tailed instead of collapsing to a thin Normal — and it is why fixing one leg in isolation moves the average far less than the planner expects.
The mechanics underneath are physical, not statistical guesswork. Each breached leg contributes excursion temperature above the 4 °C setpoint × dwell hours to a cumulative thermal dose in degree-hours; spoilage fraction is a bounded, saturating function of that dose, capped at 85% (some product is always salvageable). Dwell times are triangular per leg, excursion magnitudes exponential — the loss is genuinely confined to [0, 0.85] with a right tail, never negative and never absurd.
Before committing capital to monitoring hardware, the distributor ranked what drives mean spoilage by perturbing each input across a plausible range.
The two largest levers are chain reliability (a 3.6 pp half-spread on mean spoilage) and the line-haul breach rate (3.2 pp) — both pointing at the refrigeration unit and the long-haul leg rather than the warehouse. Excursion severity (2.9 pp) and dwell time (1.5 pp) follow. The ranking told the distributor that money spent making the reefer and the line haul more reliable buys down spoilage faster than shaving warehouse handling time.
The intervention combined real-time temperature monitoring — alerting drivers to catch and correct excursions, cutting breach probability — with faster turns that shorten dwell at each leg. The model re-ran the full chain under the improved parameters.
Mean spoilage falls from 6.8% to 2.4%, and the tail compresses hard: the share of consignments losing more than 10% drops from 25% to about 6%. Per consignment worth $280,000 that is a mean loss falling from $19,000 to $6,700; across 1,500 consignments a year the expected spoilage bill falls from $28.5M to $10.1M — an $18.4M annual reduction. The decisive change is in the tail the average could never show, not the average itself.
Because dwell time is a lever management controls directly, the distributor swept it and tracked the probability that a consignment loses more than 10% of its value, under the current chain and the monitored chain.
On the current chain, halving relative dwell (to 0.6×) pulls the exceedance probability from about 27% down to 13%; on the monitored chain the same dwell cut reaches under 5%. The two curves together show the levers are complementary — monitoring shifts the whole curve down, faster turns slide along it — and that the 5% tolerance line is reachable only when both are pulled.
In a cold chain, the legs fail together, not independently — so the spoilage you must plan for lives in a fat-tailed distribution, and Monte Carlo is what makes that tail visible before it shows up on the dock.