Multi-Horizon Demand Forecasting Using Hierarchical Attention Mechanism with Learned Temporal Embeddings

Accurate demand forecasting is critical for supply chain management and enterprise competitiveness, yet traditional methods struggle with complex temporal patterns, promotional effects, and demand volatility. This paper proposes a Transformer-based intelligent supply chain management system for multi-horizon demand prediction and inventory optimization. The architecture employs learned positional embeddings capturing domain-specific periodicities, hierarchical prediction heads for different forecasting time scales, and quantile regression for probabilistic forecasting with uncertainty quantification. Comprehensive experiments on the M5 Forecasting benchmark dataset demonstrate superior performance with mean absolute error (MAE) of 2.24 and mean absolute percentage error (MAPE) of 8.2%, representing 22.5% and 24.1% improvements over vanilla Transformer baselines, and 54 % improvement over traditional ARIMA methods. Comparative analysis against six baseline methods including Random Forest, LSTM, GRU, and TCN validates the effectiveness of attention mechanisms for capturing long-range dependencies in retail demand data. Business impact assessment confirms 15-25% inventory reduction while maintaining 95%+ service levels, directly enhancing enterprise competitiveness through improved cash flow and customer satisfaction. The system enables risk-aware decision-making through prediction intervals, supporting robust inventory optimization under demand uncertainty.

• Publication year

  2025

• Venue

  International Symposium on Fault-Tolerant Computing

• Publication date

  2025-11-21

• Fields of study

  Not labeled

• Identifiers
  DOI 10.1109/FTCS68006.2025.11405761
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

  2018cited by this paper
• The bullwhip effect in supply chains

  2015cited by this paper
• Neural network ensemble operators for time series forecasting

  2014cited by this paper
• Logistics and supply chain management : creating value-adding networks

  2005cited by this paper
• Long Short-Term Memory

  1997cited by this paper

• No citing papers are available for this paper.