Machine learning algorithms in intermittent demand forecasting: a review

Forecasting time series with intermittent characteristics poses significant technical challenges. Machine Learning (ML) techniques have the potential to revolutionise the existing state of practice by overcoming the challenges faced by standardised, conventional forecasting methods. In this paper, we shed light on technical details of major interest, such as hyperparameter tuning, data partitioning, training strategies, feature engineering, and mechanisms fostering the incorporation of exogenous variables, critical aspects not adequately covered by previous review approaches. Unlike earlier studies that have narrowly focussed on specific product categories, such as stock keeping units (SKUs) or spare parts, our research adopts a broader, product-agnostic perspective that synthesises findings across diverse industries and sectors. Moreover, our review underscores key elements to enhance the industrialisation of ML techniques by discussing the potential of transfer learning and other prominent methodologies that aim to overcome the practical implementation challenges of ML-oriented forecasting solutions. Finally, we identify research gaps, including the need for consistent benchmarking protocols and standardised evaluation frameworks, and we provide evidence supporting further exploration of emerging techniques (reinforcement learning, more sophisticated deep learning architectures, etc.). Our analysis could prove highly valuable for both researchers and practitioners operating at the intersection of intermittent demand forecasting and ML-based techniques.