Newsvendor conditional value-at-risk minimisation: A feature-based approach under adaptive data selection

The classical risk-neutral newsvendor problem is to decide the order quantity that maximises the expected profit. Some recent works have proposed an alternative model, in which the goal is to minimise the conditional value-at-risk (CVaR), a different but very much important risk measure in financial risk management. In this paper, we propose a feature-based non-parametric approach to Newsvendor CVaR minimisation under adaptive data selection (NPC). The NPC method is simple and general. It can handle minimisation with both linear and nonlinear profits, and requires no prior knowledge of the demand distribution. Our main contribution is two-fold. Firstly, NPC uses a feature-based approach. The estimated parameters of NPC can be easily applied to prescriptive analytic to provide additional operational insights. Secondly, unlike common non-parametric methods, our NPC method uses an adaptive data selection criterion and requires only a small proportion of data (only data from two tails), significantly reducing the computational effort. Results from both numerical and real-life experiments confirm that NPC is robust with regard to difficult and large data structures. Using fewer data points, the computed order quantities from NPC lead to equal or less downside loss in extreme cases than competing methods.