The aim of this paper is to analyze the forecasting performance of alternative model for the US inflation rate over the period 1950.1-2002.7. NAIRU Phillips curve models forecasts are contrasted with those obtained by a special class of nonlinear time series models, the threshold autoregressive models. The forecast evaluation is conducted on point and density forecasts. The results show that overall the non linear specification are better able to capture the distributional features of the series, although in terms of MSFE the Phillips curve specification can yield noticeable forecasting gains for medium and long term horizons. Previous finding on the forecasting superiority of the simple naïve model are confuted.