Rooney used as an example a project that his company did in the agricultural field using probabilistic methods. The use case was a farmer with 100 hectares of land. The cattle needed to be fed hay during the winter so the farmer had two choices. To buy hay from the market or set aside land on which to grow their own hay. Both would cost the farmer money because hay from the market is expensive, and if too much hay is produced, the excess has be sold at below market rate. “How can our farmer optimise how much land they use to grow the hay?” asked Rooney.

"We can get out something useful using just a few hundred data points."

DataJavelin had already developed a model alongside the Met Office to look at crop yields depending on temperature and precipitation. From that model they got yield measurements over 60 years. Rooney said: “These are quite small datasets so we can get out something useful using just a few hundred data points on our previous model.”

He explained that in the past if there was no access to probabilistic methods, they would have had to rely on average temperature and precipitation figures for each month. But with the probabilistic method, the noise is taken into account and it is easier to marginalise over this noisy data. “If we believe that we know exactly what the precipitation is going to be, we are going to end up believing we know exactly how much feed we will be able to grow and we can optimise to grow exactly the right amount to feed our cattle,” said Rooney. “Somewhere around 14 hectares.”

From the model, Rooney and his team plotted a graph with hectares versus the cost to the farmer along the axes. It showed that the probability of losing money was greater from underproducing hay by setting aside too little land than from overproducing. “By growing that little bit extra, we’ve de-risked the project by not having to buy as much hay from the market,” he said.