Approximately 175 million loaves of bread are sold a year in the Norwegian groceries trade. Mesterbakeren accounts for one fourth of the sales. Having enough bread on the shelves is consequently essential.
The challenge of supplying freshly-made bread and baked goods is closely linked with the objective of offering a fresh product to customers in more than 600 stores across Norway – all while ensuring as little food waste as possible. A technology solution that ensures both is good for the company and for the environment.
More efficient, more profitable
Many factors come into play when determining the quantity of loaves to be delivered. Traditionally, this has been based on the sales figures in each individual store for each individual type of bread, with orders then being placed for the volumes sold in similar periods. KPMG saw that this process could be done more efficiently. However, investing in new technologies often involves crossing a threshold.
– When we received the proposal to develop a prediction model, we were a little skeptical. We're talking about large volumes of data and a lot of stores. And with fluctuations in sales, making calculations based on the sales figures is demanding. For someone who has worked with this for several years, there is also some reluctance involved in letting a model take over these calculations. Especially since it's a model so advanced that few of us understand exactly how it works, says Terje Jensløkken, Head of Sales Planning & Analysis for Mesterbakeren.
To ensure a good transition, Mesterbakeren chose to develop the model in short processes in collaboration with KPMG.
Eskild Næss, partner in KPMG, has had ongoing contact with Mesterbakeren over several years.
– We had to convince them that this is a technology that works, and, through a relatively short process, we managed to show results early on, which caught Mesterbakeren’s interest, he says.
Advanced technology, but easy to use
Ricardo V. Soares is a Data Scientist in KPMG, and he is one of the minds behind the prediction model that Mesterbakeren is now using daily. He says that it is complicated to understand how the algorithm works, but that it is easy for the end user to start using the model.
– The prediction model is based on artificial intelligence. It retrieves data from previous sales and looks at customer behaviour over a long-term period. Based on this data, the model manages to find a pattern and thus provide a prediction of the customer’s future purchases, and how many items should be ordered.
There are nevertheless some cases where changes may occur that make it necessary for the model to deviate from the prediction. Examples of this may be holidays with reduced opening hours, or if there are changes in traffic, parking facilities, or the competitive situation around the store. Jensløkken finds that the model works well, especially in interaction with competent employees.
– We see that the model hits the target pretty well in its calculations, but there are a number of factors that it doesn’t know – and this is where it becomes valuable to have competent employees who can make the appropriate corrections. We find that man and machine create better results together than they would have achieved individually.
A promising start
Mesterbakeren has been using the model for nearly three years, and the figures are promising.
– We can now compare the figures with historical sales, and we see that the model is virtually spot on with close to 0 per cent deviations in some periods. This means that the store sales predicted by the model match the actual sales. The total sales figure is frighteningly accurate, says Eskild with a smile.
- I hope we can teach others that new technology and machine learning is a winning combination.