Multivariate product demand forecasting with neural networks

In retail, there are plenty of use cases that would benefit from predicting the future amount of product sales. Those use cases include cash flow management, campaign execution and inventory planning, all of which are the crucial components for the success of any retail business. Quantitative time series analysis is widely applied for predicting the product demand. It includes a set of well-established methods from statistics and econometrics. However, their capabilities are constrained by certain assumptions and they require careful statistical treatment on the data before the application. Artificial Neural Networks are powerful class of machine learning models which have shown outstanding success on the unstructured data. We proposed five different mathematical formulations to prepare and select hierarchical multivariate time series data to feed into a Long-Short Term Memory network. We referred to the formulations (or "schemes") as (1) Uni, (2) Uni-Store, (3) Uni-Product-Pcc, (4) Uni-Product-Mi, (5) Multi-Store. Each of them groups the product sales signals in multiple stores according to various association criteria to forecast the sales amounts. In the experiments, the mutual information (3) and correlation (4) based schemes demonstrated poor performance presumably due to the small number of selected products. Uni scheme produced the models that resulted in the minimum loss. The Multi-Store scheme produced the models that can be trained approximately three times faster than than that of the other schemes. Its average forecast error was not significantly higher than that of the Uni scheme.