This paper compares forecasts of Brazilian monthly inflation rate generated from different linear and nonlinear time series and Phillips' curve models. In general, the nonlinear models had a better performance. The VAR model produced the smallest mean square forecast error (MSE) among linear models, while overall best forecasts were generated by the extended Phillips curve with a threshold effect, which presented a 20% smaller MSE than the VAR model. The Diebold e Mariano (1995) test indicated a significant difference between forecasts generated from the VAR and the expanded Phillips curve with a threshold.