Regular prices and temporary discounts are important elements for retailers’ and brands’ pricing decisions. These two variables need to be considered separately because consumer sensitivities to their changes typically differ. Although the scope and richness of retail datasets have grown rapidly in recent years, most of them only record actual prices paid by customers and lack direct information about regular prices and discounts. A systematic review involving close to five hundred publications that investigated pricing variables using retail scanner data confirms this, as 63% of them only observed actual prices. To solve this missing data problem, these studies often adopted heuristics to decompose actual prices into regular prices and discount depths. However, there are many such heuristics and their accuracies have not been assessed. This research introduces DEPART, a new machine learning approach based on regression trees with a publicly available R package and benchmarks it against previously used heuristics in two different datasets. The results show that on average the proposed method outperforms previously used heuristics by 26.5% or more. Additional analyses illustrate the potential economic benefits of adopting DEPART to improve pricing decisions.