Demand forecasting by incorporating sales patterns of product groups

• Main Author: Liew, You Sheng
• Other Authors: Jagath C Rajapakse
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Mathematics of computing::Probability and statistics
• Online Access: https://hdl.handle.net/10356/166041
• Institution: Nanyang Technological University

Accurate demand forecasting is crucial for industries who have both high lead time in productions and high inventory cost for holding stock. If demand is underestimated or overestimated, it can have negative consequences for the re- tailer’s cost and revenue. For example, stock-outs can disappoint consumers, while unsold items may need to be discarded. In recent years, machine learning has become a popular approach for making demand forecasting predictions, offering significant improvements over statistical methods. However, the main focus was on predicting products with a significant amount of data, and improving on those accuracies that were already strong, as well as reducing explainability of increasingly complex models used for fore- casting. Weaker predictions were usually attributed to a lack of data, such as products that are newer or have variance that is not explainable by themselves. This paper aims to improve weaker performing predictors, therefore improving overall forecast accuracy while not compromising on explainability. We introduce a novel approach by extracting features from existing data, and incorporating sales data across different products via clustering on the extracted features. We evaluate their performance in improving weak predictors by using machine learn- ing techniques to perform forecasting. Deeper insights on the data like feature ranking and analysis can also be retrieved using factor loading analysis. The best performing technique is an autoencoder-based deep clustering ap- proach combined with feature extracted data, which allows us to capture complex patterns and relationships between similar time series data. We use a combination of clustering algorithms to group similar products together, then apply machine learning techniques to make our demand forecasts, focusing on improving stability and strengthening the forecasts for weaker predictions. After segregating strong and weak predictors into different prediction pipelines, the ensemble technique shows significant improvement in terms of accuracy and stability across months.