Asymmetric loss function in product-level sales forecasting: An empirical comparison

Abstract

In the paper we study the behavior of models estimated using asymmetric loss function for the prediction of product‐level sales. The paper is focused on the deriving of a loss function from the newsvendor model where the cost of sales over‐ and underprediction are not equal. We describe the properties of the asymmetric loss function and validate its performance on transactional sales data. The results show that when costs of sales over‐ and underprediction are non‐equal, the prediction function obtained using asymmetric loss leads to lower economic costs compared with symmetric one. Our findings suggest implementing this type of forecasting method to predict product‐level sales in the retail and restaurant industries to better accommodate business goals when solving inventory planning tasks.