Industry: Agriculture and Food Supply Chain Product: ModelRisk Application: Crop Yield Prediction
A grain trader has pre-sold 270,000 t of corn forward off a 30,000-hectare production region — a commitment of exactly 9.0 t/ha. The agronomy team's deterministic planning sheet returns a season yield of 9.6 t/ha, comfortably above the contract. On that single number the trader signed the offtake, locked the logistics, and assumed the only open question was the upside. The deterministic sheet was not wrong about the average. It was silent about the only thing that mattered: how often the season lands short.
Run the season 60,000 times in ModelRisk and the comfort evaporates. The mean yield is 9.25 t/ha and the median 9.40 t/ha — already below the 9.6 best-guess — but the headline is the spread. There is a 43% probability that regional yield falls below the 9.0 t/ha contract volume, and a 29% probability it falls below the 8.2 t/ha cost break-even. The trader was one ordinary weather year away from buying corn on the spot market to honour a contract it had priced as a sure thing.
The temptation in a 30,000-hectare model is to treat each field as an independent draw and let the law of large numbers do the work — average 30,000 fields and the spread "averages away" to almost nothing. That is exactly the mistake. A drought does not visit one field at a time. A single regional climate season — the rainfall total, the heat window — lands on every hectare simultaneously. The right unit of uncertainty is the season, not the field, and at the season level the spread is large and the lower tail is fat.
The model builds yield as agronomic potential scaled by a shared regional rainfall multiplier (Gamma-shaped, with a left drought tail), a shared heat-stress regime (a Bernoulli ignition times a Beta severity), and a small idiosyncratic soil-and-management residual that genuinely does average out across the footprint. The shared factors are what keep the P5 at 7.07 t/ha instead of collapsing onto the mean.
The real decision is not "what is the yield" but "how much can be safely committed forward". The threshold curve answers it directly: the probability the season meets a given commitment, swept across commitment levels.
At the 9.0 t/ha contract the uninsured rainfed region clears it only 57% of the time. To pre-sell at 90% confidence the trader can commit just 6.68 t/ha — about 200,000 t, not 270,000 t. Layering a regional shortfall-insurance product that backstops delivered yield to 9.2 t/ha lifts the probability of meeting the existing 9.0 t/ha contract from 57% to 100% — turning an over-committed position into a covered one. The curve, not the mean, sizes both the forward book and the insurance.
Sweeping each driver across its P10–P90 range, growing-season rainfall is the dominant lever at ±1.78 t/ha around the median, followed by the heat-stress regime at ±1.30 t/ha. Agronomic potential and variety choice — the one input the trader actually controls — moves yield ±0.88 t/ha, and the soil-and-management residual only ±0.56 t/ha. The ranking tells the agronomy team where to spend: weather risk must be transferred (insurance, forward-cover discipline), not engineered away.
The two biggest drivers are not independent. A dry season bakes bare soil hotter, so low rainfall and severe heat co-occur — and the loss tail is the corner where both go wrong at once.
The probability of landing in the joint drought-and-severe-heat corner is 8.1%, against the 6.2% a naive independence assumption would predict — a tail roughly 30% heavier than the spreadsheet that multiplies marginal probabilities together. That extra weight is precisely where the contract shortfalls cluster.
A season's yield is not a number, it is a distribution with a correlated weather tail. ModelRisk is what lets a trader pre-sell against the distribution instead of against a best guess.