| Vose Software

Industry: Agriculture and Food Supply Chain
Product: ModelRisk
Application: Crop Yield Prediction


A 9.6 t/ha "Best Guess" Hid a 43% Chance of Missing the Forward Contract

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.

Season corn yield distribution

Why a point estimate fails here

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.

How much can actually be pre-sold?

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.

Probability the season meets the commitment

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.

What drives the yield spread

Tornado of yield drivers

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.

Drought and heat strike together

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.

Joint density of rainfall and heat stress

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.

What the model changed

  • Forward book resized from the 270,000 t the deterministic sheet implied down to a 90%-confidence 200,000 t, with the balance left to be sold against confirmed yield in-season.
  • Shortfall insurance purchased to a 9.2 t/ha backstop, moving the probability of honouring the existing contract from 57% to 100% rather than carrying open spot-market exposure.
  • Weather treated as a single regional season, not 30,000 independent fields — restoring the P5 of 7.07 t/ha that a per-field average had erased.
  • Drought–heat coupling made explicit, recognising the joint loss corner is 8.1% vs 6.2% under independence.

ModelRisk Functionality Used

  • Shared-factor Monte Carlo simulation of 60,000 season paths, with one regional weather draw applied across the whole footprint to preserve correlated-tail spread.
  • Gamma rainfall and Beta-severity heat distributions giving a bounded, right-skewed loss tail consistent with agronomy.
  • Threshold-sweep analysis turning the yield distribution into a forward-commitment decision curve at a stated confidence level.
  • Tornado sensitivity analysis ranking rainfall, heat, variety and management by their effect on season yield.
  • Joint-density / dependency modelling quantifying the drought-and-heat co-occurrence the deterministic model treated as independent.

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.