A data-based approach to newsvendor problems subject to purchase price uncertainty

Abstract

Poor procurement decisions, especially involving perishable or short life-cycled products, which will have to be disposed of, can cost companies large portions of their profits. The newsvendor problem addresses inventory decisions to assist retailers in deciding just the right order quantity while still subject to uncertainty. Efficient time series forecasting techniques, including the use of machine learning models, have helped improve financial results by offering insight on future outcome-based decisions. In this dissertation, a comprehensive study was developed around the situation in which a retailer is faced with the problem of stochastic purchase prices and must decide when is the best day to place an order, as well as how much to buy to restock his perishable supply. To support the decision-making process, the problem was modeled as a variant of the newsvendor problem, subject to two decision variables: when to place the order and how much to buy. SARIMA, Prophet, MLP, RNN, and LSTM models were used for time series forecasting and were assessed in their ability to support the decision-making by forecasting future purchase prices. All forecasting- based decisions outperformed the zero-information scenario in terms of total costs. Two models (MLP and RNN) outperformed the others in terms of supporting the decision of when to buy.