Decision based model selection for the newsvendor problem

This thesis studies the performance of a decision based model selection (DBMS) method for newsvendor problems. In the newsvendor problem an optimal stock quantity has to be determined based on the distribution of the demand. The DBMS method distinguishes itself from conventional model selection methods by selecting models based on the performance of its decision, rather than on the goodness-of-fit to the data or its predictive power. We begin with giving expressions of the expected regrets caused by using DBMS when two models are available. We subsequently discuss the robustness of a decision, knowing partial information of the underlying model. We derive the maximum regret which is achieved by a two-point distribution for the demand. Lastly, we numerically illustrate the performance of DBMS when the underlying model assumes this two-point distribution and show that the DBMS performs better than a purely data-driven approach, especially when the sample size is small.