The warehouse that knew the weather

A supermarket chain found that a two-degree rise in forecast temperature changed sales of certain products by more than any promotion it had ever run.

The products were predictable in hindsight and not in advance: salad, ice cream, barbecue charcoal, and, unexpectedly, a particular brand of insect repellent whose sales tracked humidity rather than heat. The chain's contribution was not the discovery of these links but the decision to connect the forecast to the ordering system, so that a warm weekend produced a larger delivery without anyone making a judgement.

Automating a judgement of this kind raises a question that recurs in every forecasting application: what happens when the forecast is wrong. The chain's answer was asymmetric, and asymmetry is usually the right shape. Over-ordering salad costs the margin on the wasted stock. Under-ordering on the hottest Saturday of the year costs a customer who drives to a competitor and may not come back. Because the two errors are unequal, the optimal response is not the most likely outcome but a deliberate lean towards the expensive side.

This is the part most often got wrong by people new to forecasting. A forecast gives a distribution; a decision needs a number; and the number is produced by combining the distribution with the cost of being wrong in each direction. Choosing the central estimate is correct only when the two costs are equal, which they almost never are.

The chain's own summary was blunt. The weather data was free and had been available for decades. What they built was not a prediction but the willingness to act on one before the evidence arrived.