Since the sixteenth century, there was already the idea that purchasing power should influence the value of each currency. In order to understand the relationship between exchange rate and inflation, linear autoregressive models always presented difficulties to beat the random walk. Possible frictions in foreign exchange operations may hinder arbitrage close to the equilibrium exchange rate considered by financial agents. As the exchange rate distances itself from the value considered fair, the convergence becomes more intense, because the costs would no longer be so relevant to the potential profit of the operation. In the proposed nonlinear model, there are two different regimes: one near equilibrium (random walk behavior) and one behavior away from it occurring simultaneously, but with variable weights. For different levels of the exchange rate relative to the equilibrium, one regime gains more weight and the other loses relevance. This thesis aims to evaluate the predictive nature of the exchange rate movement. The nonlinear model ESTAR is used to create baskets of currencies to be bought and sold and the aggregate return based on exchange rate movements is computed. Finally, we consider the interest rate effects on the model to set up currencies portfolios in order to simulate the return on an investment using this strategy. For the baskets of currencies, the model generated good returns and low risks, based on both standard deviation and drawdown. This characteristic was observed in the in-sample model and in the out-of-sample model, which indicates a strong predictive power. Considering the interest effect, portfolios with fewer currencies showed positive returns, but this advantage is lost by increasing the number of currencies.