Abstract
Exchange rate forecasting involves many challenges in research. Due to the difficulty of selecting superior variables to design a good forecasting model, few empirical studies have discussed the influence of explainable variables. In this paper, some new forecasting model is constructed; we adopt the Particle Swarm Optimization (PSO) to select the optimal input layer neurons to predict NTD/USD exchange rates by the Back Propagation Network (BPN), called PSOBPN model, which could help to obtain the better performance in back propagation. There are several steps in this experiment: first, we divided the whole data into six periods of sliding windows. Second, we selected superior variables by the PSO and GA method, and we selected 10 variables within the entire 27 variables. Finally, we forecasted the exchange rate by BPN with the selected variables. The results showed that the PSOBPN achieves the best forecasting performance and is closely matched with the actual exchange rate. It is better than GABPN with optimal selection. Also, it is better than randomly selected models.