Data-driven inventory management under customer substitution

• Data-driven solution for multi-product inventory with customer substitution • Method validated on both synthetic and real-world retail data • New metric identifies when integrated models outperform traditional ones • Managerial insights on data use and substitution for inventory decisions Retailers selling perishable goods typically offer multiple products in a product category, e.g., fresh food or fashion. Managing the inventories of these products is especially challenging due to frequent stock-outs and resulting substitution effects within the category. Furthermore, the true product-specific demand distributions are usually unknown to the decision maker. Digital technologies have massively expanded the available data and computing power, which can help improve inventory decisions. In this paper, we present a novel solution approach for the multi-product newsvendor problem. Our method is based on modern machine learning techniques that leverage large available datasets, e.g., data on historical sales, weather, store locations, and special days, and are able to take complex substitution effects into account. We evaluate our approach on two real-world datasets of a large German bakery chain and find that our data-driven approach outperforms the benchmark on the first dataset and performs competitively on the second dataset. We then analyze our approach in a controlled environment with synthetic data to pinpoint the factors that determine its performance. While saving computation time, our approach outperforms relevant benchmarks for a wide range of cost parameters and substitution levels when the demand distribution depends on exogenous data; it is competitive otherwise.