Retailers that offer perishable items are required to make ordering decisions for hundreds of products on a daily basis. This task is non-trivial because the risk of ordering too much or too little is associated with overstocking costs and unsatisfied customers. The well-known newsvendor model captures the essence of this trade-off. Traditionally, this newsvendor problem is solved based on a demand distribution assumption. However, in reality, the true demand distribution is hardly ever known to the decision maker. Instead, large datasets are available that enable the use of empirical distributions. In this paper, we investigate how to exploit this data for making better decisions. We identify three levels on which data can generate value, and we assess their potential. To this end, we present data-driven solution methods based on Machine Learning and Quantile Regression that do not require the assumption of a specific demand distribution. We provide an empirical evaluation of these methods with point-of-sales data for a large German bakery chain. We find that Machine Learning approaches substantially outperform traditional methods if the dataset is large enough. We also find that the benefit of improved forecasting dominates other potential benefits of data-driven solution methods.