Safety stock levels in dynamic markets. A neural approach

The target of the study is the design of an innovative decision support tool to solve the problem of a correct safety stock dimensioning in a production system, in case of independent demand of an highly dynamic and rapidly changing market. Given the complexity of the demand, the definition of the stock level using traditional techniques, based on long period elaboration of statistical parameters, can be inadequate due to the high uncertainty of the data related to programmed and received customer requirements. A classification problem was set, implemented on two different neural architectures, to define a short term service level to compare the safety stock, defined with the classical formulas, with a virtual one based on moving average and standard deviation values and give as output the necessity to increase-confirm-decrease the calculated level. The different architectures were trained, tested and evaluated in various configurations of knots and levels, on time series generated by a pseudo-random composition of probabilistic demand distributions, to evaluate the response in a way that could be at most independent from the data structure.