Options
2021
Conference Paper
Title
Demand Forecasting Using Ensemble Learning for Effective Scheduling of Logistic Orders
Abstract
We present forecasting models based on extreme gradient boosting to predict demand using real-world data of a German intermediary company in the media sector. The data set comprised the daily demand of 196,767 products from three years (mid-2017 to mid-2020) and meta information for each product including product type affiliation. Models were trained separately for each product type either to predict the demand on group or product level. For the latter, training was in rolling format based on the last 12 weeks to then predict the product's short-term demand one-week ahead. Performance, evaluated via the coefficient of determination, is especially precise for specific product types. Engineered features consisting of seasonal information, statistical indices, and general performing indices obtained via fuzzy c-means clustering over time improved the prediction. Especially, predictions for the upcoming week on product level are challenging but of high value for future business decisions regarding inventory planning and purchase orders.
Author(s)