Abstract

This study developed neural network models using deep learning with a range of 4 to 10 neurons to forecast the stochastic demand of a group of 20 clients over three years. The models were compared with traditional approaches such as simple moving average, weighted moving average, simple exponential smoothing, and double exponential smoothing. The results revealed that the deep learning model with 10 neurons was the most efficient, significantly reducing forecast errors compared to traditional methods and machine learning. This study highlights the ability of neural networks to capture complex patterns and adapt to stochastic variability, providing greater accuracy and consistency in demand prediction. These findings are crucial for decision making in environments with uncertain and variable demand, improving planning and resource management. Additionally, the deep learning model with 10 neurons showed an average reduction in mean absolute deviation by 15%, compared to 25% in traditional methods. This improvement in accuracy underscores the potential of neural networks in demand forecasting applications.