Cold-Start Demand Prediction for New Products: A Meta-Learning Approach on the M5 Competition Dataset

Zichao Li, Zong Ke

Keywords: cold-start forecasting, meta-learning, supply chain, demand prediction, few-shot learning, retail analytics, time-series

TL;DR: This paper proposes a meta-learning framework for cold-start demand forecasting that reduces errors by 18% via hierarchical adaptation and probabilistic few-shot learning, validated on retail benchmarks (M5, Favorita, Amazon).

Abstract: We present a hierarchical meta-learning framework for cold-start demand forecasting in retail supply chains. Our method combines Transformer-TCN architectures with model-agnostic meta-learning (MAML) to enable accurate predictions for new products with minimal historical data. Evaluated on the M5 dataset and a real-world case study, the framework reduces forecasting errors by 32\% compared to state-of-the-art approaches while requiring only seven days of observations. Key innovations include category-aware task sampling and probabilistic few-shot adaptation, addressing critical limitations of existing methods in data-sparse scenarios. The system's practical utility is demonstrated through deployment with a multinational retailer, achieving \$2.3M annual cost savings.

Submission Number: 2