Demand Forecasting in Industrial Contexts: From Traditional Approaches to Foundation Models

Demand forecasting is a key aspect of supply chain management. However, producing accurate forecasts is highly challenging, particularly for intermittent and erratic patterns or when dealing with short time series. This study evaluates and compares the performance of statistical models, machine learning (ML) and deep learning (DL) models, as well as recent foundation models, in a real-world industrial context. The analysis is based on time series data from a large steel company. Results show that statistical models still achieve acceptable performance. ML models, especially LightGBM, can improve prediction accuracy by exploiting cross-learning and incorporating exogenous features. DL models underperform, as their potential cannot be fully exploited in data-scarce scenarios. Foundation models, leveraging the knowledge acquired during the large pretraining phase, prove to be highly effective in low-data contexts. Both in zero-shot and fine-tuned settings, they can outperform traditional approaches, though at the expense of a significantly higher computational cost. In particular, the fine-tuned Moirai model emerges as the best overall performer, achieving a 95% improvement in BIAS% and an 11% improvement in MAE% compared to the baseline Window Average, but requiring substantially longer training and inference times than simpler, traditional models.