Forecasting exchange rates using ARMA and neural network models.

摘要

Exchange rate movement is an important subject of market study and is of great deal of interest to international investors. It plays a key role in open economies today and characterized by extremely high volatility. Although a lot of theories on exchange rate behavior have been developed, forecasting exchange rates still shows humbling results and remains as a considerable point of interest for academics and market practitioners.The purpose of this thesis is threefold: first to identify the type of time series and neural network models that perform best in short-run second to forecast exchange rates in and out of sample and third, to compare the forecasting abilities of time series and neural network models to random walk performance.We examined numerous ARMA and neural network models and attempted to forecast exchange rate of Brazilian Real for one month ahead. We observed that excluding some minor cases, ARMA models do not outperform random walk. Our experiments with ARMA showed one more time that a RW model is superior to ARMA models, both constrained and unconstrained.Neural network, however, outperforms the random walk, and shows better performance than ARMA, which is quite reasonable result given the characteristics and abilities of the neural network. Nevertheless, based on empirical studies and research the random walk performs better all other models in short and medium run and reasons for such triumph are subject to further investigation.