Computational Intelligence approaches in forecasting demand of essential medical products

• Main Author: Chung, Suhwan
• Other Authors: Jagath C Rajapakse
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Computer and Information Science; Time series forecasting; Demand forecasting; Meta learning; Global forecasting; Cross-series forecasting
• Online Access: https://hdl.handle.net/10356/182464
• Institution: Nanyang Technological University

It’s widely accepted in the forecasting community that complex methodologies may not necessarily result in better out-of-sample predictions compared to simpler methods. The idea is backed by the numerous findings from influential forecasting competitions and previous literature. For instance, in the influential M4 forecasting competition, eight purely neural network-driven models tend to struggle due to insufficient sample size for reliable parameter fitting and suffer from the risk of fitting to the random noise. In fact, having more data in a time series context doesn’t necessarily mean to having more reliable insights from each individual time series. Instead, it indicates the availability of numerous “related” series can be considered in the forecasting process that can provide reliable information. In addition, as more related series becomes accessible, the use of more complex models become a more viable approach when trained jointly across all series by “borrowing” information from other series. Forecasting models that can be trained across joint (group) series have a competitive edge over traditional uni-variate forecasting techniques such as Holtwinters, ARIMA, and others. Despite numerous efforts, constructing forecasting models that can train across various disparate series remains challenging since the weak or irrelevant series and features could affect to the prediction of varying forecasting model. To mitigate issues cased by weekly connected series, we recommend developing models that analyze structural features of time series instead of constructing models using past observed values. In our second chapter, we propose a two-stage meta-learning process that resembles the Model Agnostic Meta Learning (MAML) frameworks where the focus is to improve performance on new forecasting tasks with just a few gradient updates. Other challenges related to joint series training observed in our literature review involved with training across heterogeneous time series. The issue becomes more noticeable when a forecasting function is required to be estimated from all relevant time series simultaneously, which is also known as Global Methods. In our third chapter, we propose a unique method to address the issue of identifying similarities among series. This method uses time series characterization algorithms to cluster series and apply forecasting functions to subsets based on their similarity. In our fourth chapter, we bridge the gap between research and industry application of cross-series forecasting by examining the interactions between autoregressive orders and their impact on performance during joint-series training.