Seasonal Decomposition and Trend using Hybrid STL- FNN with Application

Most researchers in the field of time series follow one approach, which is to rely on time series models in general. Several models have emerged in engineering, economics, and health applications. These models are known as neural networks, which imitate human nerve cells. This study suggests a hybrid algorithm called STL-FNN. It combines two methods - seasonal trend decomposition STL and feed-forward neural network (FNN). The STL method analyzes the original data into three subseries: seasonality, trend, and residual. To predict each of the three sub-series, the FNN neural network uses the seasonal component and separates them. Finally, all predicted outputs are combined as a total time series product. The results indicate that the STL-FNN can check the performance of the hybrid model. It uses the relative absolute error (MAE) criterion. We analyzed actual data to test the hybrid model’s ability to predict. The model used the average monthly spending of foreign visitors in the United Kingdom from January 1986 to February 2020