SOT-FER: A multi-tier entropy-based time series forecasting framework with an application to manufacturing

Accurate forecasting is crucial for manufacturing systems’ production planning, inventory management, and resource allocation. However, managing multiple time-series datasets of varying complexity poses significant challenges, as traditional statistical models often fail to capture intricate patterns, and more advanced deep learning approaches—though powerful—can be computationally expensive. Nonetheless, these cutting-edge approaches are not always necessary since more straightforward machine learning (ML) techniques can achieve comparable performance in many cases. Therefore, balancing accuracy with efficiency thus requires identifying when to escalate to sophisticated models. This study introduces the System for Operational Time-series Forecasting and Entropy-based Review (SOT-FER), which uses a suite of entropy measures (Shannon, spectral, dispersion, permutation, and multiscale) alongside assessments of trend, seasonality, residual variability, and stationarity to quantify time-series complexity. SOT-FER employs hierarchical clustering to group series by level of pattern intricacy and guides the selective application of forecasting methods designated in tiers. Tier A features established statistical and ML models (State Space Exponential Smoothing, AutoRegressive Integrated Moving Average, Theta, Prophet, K-Nearest Neighbors, and Random Forest), while Tier B considers Long Short-Term Memory (LSTM) networks exclusively for the most challenging series. Applied to a real-world dataset of 128 monthly demand patterns from a U.S.-based manufacturing corporation, SOT-FER accurately pinpoints where advanced methods deliver significant gains, showcasing that selective Tier A + Tier B deployment improved forecast accuracy by over 50 % compared to a naïve baseline. This data-driven framework offers a scalable roadmap for improving forecasting strategies across diverse contexts by categorizing series according to inherent complexity.