Adaptation of the newsvendor model with asymmetrical loss functions for inventory management on Internet marketplaces

the article addresses the problem of distributing product inventories on online marketplaces in the context of rapidly growing e-commerce and unequal costs associated with surplus and shortage of stock. The aim of the study is to adapt the Newsvendor model with asymmetric loss functions to build a demand forecast that minimizes the seller’s total operating costs while improving the service level within regional warehouse clusters of the marketplace. The research question is how the choice of forecast quantile, accounting for the platform’s economic parameters, affects optimal ordering and warehouse distribution. The scientific novelty lies in refining the economic interpretation of the quantile via cost ratios and incorporating platform-specific factors into the objective function of the forecast. Materials and methods include an analytical review of classical demand models and machine learning methods, as well as formalization of the Newsvendor framework in terms of marketplace clusters and warehouses. An approach is proposed for selecting the quantile level based on parameters of price, procurement, storage, and penalties, complemented by normalized factors of sales stability and product visibility in search results. To train the forecast, a piecewise linear asymmetric loss function is applied, penalizing underestimation more or less than overestimation depending on the target service level. Methodologically, the solution jointly determines the forecast and the ordering policy at the cluster level. The results show that replacing symmetric error functions MSE and MAE with an asymmetric loss function aligns the forecasting task with the inventory optimization problem. Economic calibration of the quantile shifts the optimal order toward shortage for products with high reputational risks, and toward surplus for large, slow-moving goods. The introduced factors increase the model’s sensitivity to product assortment differences and enhance the interpretability of service levels. A mapping of under- and overestimation links forecast errors to specific types of platform costs. The discussion highlights the practical applicability of the approach in seller and marketplace decision support systems for cluster-based supply planning, reducing delivery times, and lowering storage fees. Limitations are associated with calibration of factor weights, robustness of recommendation system effects, and data availability on costs. Future prospects include quantitative validation with sales logs, Bayesian estimation of uncertainty, expansion to multi-echelon management, and joint optimization of redistribution between warehouses over time.

marketplace, Newsvendor model, asymmetric loss function, inventory management, demand forecasting