McMaster University
From Forecasting to Optimization in Data-Driven Multi-Item Inventory Systems
Abstract
dc:description.abstractInventory systems often need to manage the ordering and stocking of multiple items with uncertain and interrelated demands arising from diverse populations. Traditional inventory models usually assume that demands are independent or identically distributed for simplicity and tractability. This thesis develops data-driven solutions for forecasting and optimization in multi-item inventory systems with auto and cross-correlated demands under uncertainty and distribution heterogeneity. The studies are motivated by the inventory management of human Immunoglobulin (Ig) products. Chapter 2 proposes a cluster-based forecasting framework to address the inherent heterogeneity in demand for Ig products. The core contribution is to directly incorporate the clustering task's outcome into the forecasting performance, rather than treating patient grouping as a separate preprocessing step. Clustering choices, such as the number of clusters and feature weights, are selected based on forecast performance. The framework outperforms standard Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) benchmarks. Chapters 3 and 4 study multi-item inventory ordering policies with temporally and cross-item correlated demands. Chapter 3 studies optimal ordering policies in a multi-item, multi-period inventory system under a separable cost function across items. To capture the auto and cross-correlations in the demand process, this study extends Autoregressive Moving Average (ARMA) demand models to a Vector ARMA (VARMA) framework. This approach allows for a formal analysis of optimal policies and provides closed-form order-up-to levels. Experiments are then conducted on synthetic, real Ig demand and retail sales datasets. The results show significant cost improvements compared to independent ordering systems. Chapter 4 studies the stochastic joint replenishment problem (SJRP), where product demands are interdependent across items and time periods. The problem is solved using a deep reinforcement learning (DRL) method, Proximal Policy Optimization (PPO), with a gated and factorized action design that separates two key decisions, in line with SJRP's main decisions: whether to place a joint order and how much to order at each period for each item. It has been tested against both classical ordering policies and DRL methods using synthetic, Ig and retail datasets. The approach leads to cost savings, especially when the joint ordering costs are moderate to high, compared to baselines. Together, the chapters present frameworks that improve both forecasting accuracy and ordering efficiency under uncertainty in complex, real-world multi-item inventory environments.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rahimi, Zhaleh
- Advisor dc:contributor.advisor
-
- Down, Douglas
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.71548/25540