The complexity of supply chain problems, more specifically the case of on-stock chains, is due to performance indicators variety, antagonism, and the difficulty of understanding the effects and interactions of different performance drivers with regard to these indicators. As mathematical formalization is essential to optimize the performance of these chains, this paper generally aims to study the contribution of Machine Learning to mathematically link the evaluation parameters of an on-stock supply chain to its action parameters. This work is based on an academic case study that seeks to mathematically formalize the problem of delivery delay in an on-stock supply chain. To this end, several Machine Learning algorithms have been tested and compared. This experience highlighted the impossibility of obtaining a labeled dataset through data collection from the real system. It thus demonstrates the necessity to use a simulation system, in particular, discrete event simulation, to generate this dataset.