An integrated data-driven method using deep learning for a newsvendor problem with unobservable features

We consider a single-period inventory problem with random demand with both directly observable and unobservable features that impact the demand distribution. With the recent advances in data collection and analysis technologies, data-driven approaches to classical inventory management problems have gained traction. Specially, machine learning methods are increasingly being integrated into optimization problems. Although data-driven approaches have been developed for the newsvendor problem, they often consider learning from the available data and optimizing the system separate tasks to be performed in sequence. One of the setbacks of this approach is that in the learning phase, costly and cheap mistakes receive equal attention and, in the optimization phase, the optimizer is blind to the confidence of the learner in its estimates for different regions of the problem. To remedy this, we consider an integrated learning and optimization problem for optimizing a newsvendor's strategy facing a complex correlated demand with additional information about the unobservable state of the system. We give an algorithm based on integrating optimization, neural networks and hidden Markov models and use numerical experiments to show the efficiency of our method. In an empirical experiment, the method outperforms the best competitor benchmark by more than 27%, on average, in terms of the system cost. We give further analyses of the performance of the method using a set of numerical experiments.