Demand Forecasting and Inventory Management of Perishable Inventory - with a Focus on Blood Platelet Transfusions

Abstract

Inventory management of perishable products has seen extensive study over the years; the perishable nature capturing the real-world phenomena of expiration after a limited shelf life. Such problems are challenging as they involve balancing demand fulfillment with minimal wastage. An added dimension to such problems, given the rise of machine learning, is to estimate future demand. Demand forecasts can be helpful for decision making, in particular they can be used for finding the optimal ordering quantity for the products. The central thesis of this dissertation is that by forecasting the demand and utilizing it in the inventory management process, we can build a more robust inventory system that takes additional information into consideration when making decisions. Firstly, five different demand forecasting methods, ARIMA (Auto Regressive Integrated Moving Average), Prophet, lasso regression (least absolute shrinkage and selection operator), random forest, and LSTM (Long Short-Term Memory) networks are utilized and evaluated via a rolling window method. Subsequently, we study the structural properties of the optimal ordering policy for perishable products with fixed shelf lives in a periodic-review single-item inventory system over a finite horizon, where demand forecasts are available. The objective is to find the optimal ordering policy that minimizes the total expected cost, consisting of a linear ordering cost, inventory holding cost, wastage cost, and shortage cost, over a finite horizon. We show that the optimal policy is a state-dependent base-stock policy in which the base-stock values are a function of the system’s state, the inventory level, a vector of current and previous demand forecasts, and previous demand values. Moreover, we explore the monotonicity properties of the optimal policy. The monotonicity properties motivate us to propose a heuristic in which the order quantity is an affine function of the inventory level and forecast-dependent target inventory levels. We evaluate the performance of the proposed heuristic on platelet transfusion data for hospitals in Hamilton, Ontario. Experimental results show that the proposed heuristic is effective in minimizing the total cost while maintaining low on-hand inventory levels.