Forecasting Necessary Levels Of Bread Stock For "Direct-To-Shelf" Delivery Using CNN And LSTM Artificial Neural Networks

In the "direct-to-shelf"delivery model, a supplier (i) sells produce to clients to stock assigned store shelves, and (ii) buys back expired or damaged stock from the same clients. Predictions of necessary levels of stock are thus needed to meet market demands with efficiencies at scale. This paper examines the use of a CNN-LSTM combined method for the multi-variate prediction of necessary levels of stock for US based 'Martin's^® Famous Potato Rolls and Bread'. First the historic sales data was merged with historic weather data, then the merged dataset was divided twice into multiple 'micro-datasets': First based on ZIP Code and Product ID, then on Store ID and Product ID. Using this approach high performing 'micro-models' were developed capable of forecasting highly localised geographic demand for specific products. The results of the experiments demonstrate highest accuracy from a CNN-LSTM configuration using ZIP Code and Product ID to predict the 'Net Sales' target variable, achieving a RMSE of 2.006, MAE of 1.3892, and loss of 0.73.