nowadays in trade and economic issues, prediction is proposed as the most important branch of science. existence of effective variables, caused various sectors of the economic and business executives to prefer having mechanisms which can be used in their decisions. in recent years, several advances have led to various challenges in the science of forecasting. economical managers in various fields are trying to achieve effective forecasting methods with lower error possibilities. one of these fields is predicting changes in exchange rate that all economic experts can use in different fields like macro-economic decisions and economic planning. others believe that it is a random phenomenon which can’t be predicted. in normal economic calculations models are assumed linear for simplicity. doing these calculations with the non-linear models will cause bias which in predicting. through non-linear models, various phenomena such as turbulence can be also investigated. this thesis uses different tests like bds, hurst and predict correlation dimension. consequently, by these tests, predictability of exchange rate and the state of being linear or non-linear will be determined. what’s more, we showed the exchange rate is predictable and bearable and not random. using hurst exponent and runs tests we proved that this variable has a long memory. also bds test ascertained that behavior of the variable is non- linear. for proving stability we had to test augmented dicky- fuller and philips- pron. the results show this variable is not stable. and for proving predictive power of non- linear models, neural network is more power full than garch, which is more power full than arima.