Time series analysis is primarily concerned with maintaining stability to make accurate forecasts about the future. Numerous statistical models have been developed for this purpose, focusing either on mean stability or variance stability. Neural network models, inspired by the structure of the human nervous system, have subsequently been adopted to enhance predictive capabilities.
In this work, we propose a hybrid model that integrates the structural properties of time series with modular neural networks (MNN). The model decomposes a given time series into four sub-series based on its key components—seasonality, trend, dispersion, and remainder (STDR)—and each component is processed using a corresponding neural sub-network. The predictions from these subnetworks are then combined to form the final network output.
The proposed STDR–MNN hybrid model efficiently captures complex temporal dependencies within time series data and is implemented using MATLAB 2022A. The model’s performance is evaluated using real datasets to demonstrate its effectiveness in forecasting segmented time series with improved accuracy and stability.