Abstract: Based on the characteristics of the current RMB exchange rate， a new forecasting model should take into account the qualitative studies of exchange rate theories， the quantitative research of historical tendencies and the impact elicited by policy issuing. The nonlinear auto regressive model with exogenous inputs， NARX for short， was adopted to forecast RMB exchange rate， and the non-deliverable forward exchange rate， known as NDF， was selected as an external input of the NARX model to improve the forecasting performance in face of sudden policy and news. By empirical testing， it was found that the NARX model with NDF performs much better than the NARX model without NDF. Moreover， selecting the training data from a wide range and with the historical values impacting exchange rate tends to affect the models generalization performance to a large extent.

Key words: NARX model, RMB, exchange rate forecasting