Optimizing Sustainability: Benchmarking Time-Series Forecasting Models on Purchasing Data to Reduce Overproduction

This study delves into computational sustainability, leveraging AI methods to address sustainable consumption and production challenges. Utilizing historical purchasing data from a prominent Icelandic supermarket chain spanning from 2018 to 2023, the study explores various dimensions of data aggregation, including macro-categories, micro-categories, and product labels. Six distinct forecasting methodologies are employed: regression-based, LLM, foundation models, and statistical approaches. Evaluating these models using Root-Mean-Squared-Error reveals different findings dependent on dataset aggregation. As a result, datasets with finer aggregation exhibit higher predictive accuracy, with the LLM model consistently outperforming others across macro and micro-category datasets. While one of the foundation models demonstrates comparable performances, LLM’s efficiency is notable despite GPU utilization and longer processing times. The regressor emerges as the most effective predictor for datasets categorized by product labels, with LLM and one foundation model also displaying commendable performance. These findings offer valuable insights for reducing overproduction and optimizing production planning in the production industry, underscoring the importance of leveraging advanced forecasting models to promote sustainability. Further details and implications are discussed in the paper.