Strawberries are a highly valuable soft fruit product. The UK was estimated to produce 76,000 tonnes of strawberries worth £325m in 2015. Whilst strawberries have grown rapidly in commercial significance, matching supply to demand is a complex and highly dynamic task. Estimating demand is inherently complex as it can be affected by weather, seasonality, marketing campaigns, product launches, promotions and special occasions like Christmas or Easter. Given the above factors and a desire to avoid the risk of stock outs, the supply chain tends to accumulate buffer stocks which in turn leads to wastage. This project aimed to identify ways to use machine learning models to improve supply and demand forecasting with the overall outcome to increase efficiency and reduce wastage within the strawberry supply chain.
The project started with an initial client workshop to map the different components of the supply chain, identify the variables to be predicted, estimate required levels of timeliness and accuracy required for the forecast information and identify client data sources. Following the acquisition of suitable supply and demand data the data science team built a bespoke machine learning data model and trained this on historic sales for each day of the year. This was combined with weather forecasts of temperature and rainfall at each location to accurately assess the impact of weather on sales whilst accounting for special one-off events.
The resultant store specific model for strawberry sales volumes demonstrated a noticeable improvement over the existing in-house prediction system with the performance varying by location. The results showed that machine learning can be used to reduce the mismatch between stocking levels and demand. However, the new model increased the number of under stocked events (increased risk of empty shelves). To place a value on the utility of the model required additional work to quantify the relative costs of under and over stocking.
The final report identified the need to quantify the impact of promotions on demand and to demonstrate the economic value of the overall model through incorporating further data on the costs of under- and over-ordering.