DYNAMIC DEMAND FORECASTING IN SUPPLY CHAINSUSING HYBRID ARIMA-LSTM ARCHITECTURES

Abstract

Demand forecasting in supply chains is essential for efficient inventory management, production planning, and cost optimization. Traditional time series models, such as Autoregressive Integrated Moving Average (ARIMA), are effective in modeling linear relationships but often fall short in handling the nonlinear complexities characteristic of volatile demand patterns. Conversely, deep learning models like Long Short-Term Memory (LSTM) networks have proven adept at capturing intricate temporal dependencies, though they are computationally intensive and require substantial data. This study introduces a hybrid ARIMA-LSTM framework that leverages the linearity strengths of ARIMA in conjunction with LSTMs capacity for learning non-stationary, nonlinear patterns. The hybrid model decomposes the demand series, modeling linear trends with ARIMA and residual nonlinear components with LSTM. Empirical evaluations of real-world supply chain data reveal that this integrated architecture outperforms standalone ARIMA and LSTM models, achieving superior predictive accuracy and robustness in handling demand variability. Our findings underscore the hybrid models potential as an advanced predictive solution for dynamic, data-driven supply chain management.