Optimizing Supply Chain Operations Using Advanced Time -Series Mixer Models for Demand Forecasting and Inventory Under Uncertain Demand

Abstract

Effective inventory management in the presence of uncertain demand necessitates forecasting models that strike a balance between accuracy and cost-efficiency. This study introduces an approach that incorporates Reversible Instance Normalization (RevIN) with Time-Series Mixer (TSMixer), referred to as RevIN-TSMixer, a transformer-based model built on the Multi-Layer Perceptron Mixer architecture, leveraging Multi-Layer Perceptron (MLP) across temporal and feature dimensions. The proposed approach is evaluated in two scenarios: error optimization, aimed at minimizing forecasting errors, and total cost optimization, emphasizing practical cost-effectiveness. Results demonstrate that RevIN-TSMixer consistently outperforms the original TSMixer model, achieving a 2% reduction in forecast errors and significantly lowering total costs. These advancements are credited to RevIN's ability to address data shifts and improve adaptability to fluctuating demand patterns. To further enhance supply chain management, the study integrates a continuous review (r, Q) inventory policy with the proposed forecasting model, replacing the traditional Economic Order Quantity (EOQ) model's constant demand assumption with average forecasted demand. This integration includes safety stock calculations, reorder point determination, and comprehensive cost analyses, reducing shortages and optimizing inventory strategies. Validated through experimental scenarios, the proposed framework highlights the synergy between accurate forecasting and adaptive inventory management. By demonstrating robust forecasting accuracy and significant cost reductions, RevIN-TSMixer establishes a state-of-the-art solution for supply chain optimization under uncertain demand conditions.

Keywords: Demand Forecasting, Inventory, supply chain, Time series model, TS Mixer