Abstract
Deep Learning (DL) offers powerful tools for demand forecasting by capturing complex nonlinear patterns and adapting to dynamic market conditions. Accurate forecasts are vital for optimizing production planning, reducing costs, aligning with customer demand, and efficient resource allocation. Forecast accuracy depends heavily on both dataset characteristics and DL hyperparameters, which influence model complexity and learning behavior. Although research efforts are focused on using data properties in demand classification and hyperparameter tuning for better DL accuracies, the efforts exerted in analyzing their impacts are few. This paper investigates how demand characteristics, such as variability, zero demand frequency, and spikiness, and DL hyperparameters of Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) affect multi-period forecast accuracy. Three types of demand patterns are analyzed: smooth demand, erratic demand without spikes, and erratic demand with spikes. Demand Complexity Index (DCI) is proposed as an integrated metric of demand characteristics, including demand variability, the amount of zero demands, and the degree of spikiness of the demand. To handle zero-demand periods and normalize accuracy across datasets, Weighted Mean Absolute Percentage Error (WMAPE%) is used to assess forecasting accuracy. Results show that the Coefficient of Variation (CV) is the most influential data feature, while Learning Rate is the most impactful hyperparameter affecting forecast accuracy. Demand complexity significantly influences forecasting accuracy, with WMAPE increasing by up to 14.6% per unit rise in DCI for GRU and 11.3% for LSTM, highlighting the need for complexity-driven model optimization. The main contribution of this work is introducing an integrated framework to tailor hyperparameter selection to input demand characteristics, enabling improved accuracy and faster processing.”