Sales forecasting through aggregation: integrating neural-Bayesian and knowledge-based methods (ENN.5323)

In many industrial sectors, there is a vast increase in the range of products. For example, a large supermarket today handles roughly 60 times more products than 60 years ago. At the same time, the product life cycles become shorter and the demand patterns more irregular. Forecasting and even understanding these sales patterns thus becomes more and more difficult, but also more and more important. The goal of this project is to develop methods and procedures to forecast and explain the sales of large groups of products. The standard approach, separate modelling of the demand for individual products, is unfeasible or at best not optimal: too many different products with too few data for each product to build reliable models. One solution is to aggregate the products in product families. Using standard methods and/or neural networks, forecasts are generated at the level of these product families, which are then translated back to the level of the individual products. In this project, we will develop procedures to systematically and optimally divide the products into product families.
The second approach to be studied is the joint modelling of the demand for a large amount of products in a single neural-network model. In this way, the different tasks, forecasting the sales for individual products, can "learn from each other". The basic idea is that the network manages to learn a set of generic features, in this case typical sales patterns. The model for each individual product then follows from a product-specific combination of these features. This kind of multitask learning is an ideal testbed for Bayesian methodology. Recently developed tools can be applied to derive the posterior distribution of model parameters given the data and prior assumptions. The final result can be interpreted as a form of implicit aggregation.
The more knowledge-based explicit aggregation and more data-oriented implicit aggregation are, to some extent, complementary. Knowledge, obtained in the search for systematic explicit aggregation methods, will be implemented as prior assumptions in the neural-Bayesian approach. The other way around, the more data-oriented approach will direct the search for systematic procedures. The final goal is to combine the two approaches into a successful solution: systematic procedures for a first global subdivison of products in product families integrated with advanced forecasting tools based on the neural-Bayesian approach for further fine-tuning.

Er zijn nog geen resultaten bekend.

Five companies are involved in this project.