This paper studies the growth rate of Colombian GDP between 1982 and 2008 with a SETAR model (Self-Exciting Threshold Autoregressive), based on the methodology proposed by Tsay (1989) and Tong (1990) for the detection of nonlinearities related to changeable regimens. The main results show empirical evidence of nonlinearity of threshold in the series associated with high or low rates of growth taken in an annual lag, remaining more time in the regime of higher growth rates than in less intensive dynamic regimes. Furthermore, the study compares the performance of the SETAR results with the forecasts generated by a linear autorregresive model in different horizons of prediction, based on a symmetrical loss function. Even though the performance of the forecasts of the SETAR model does not seem to improve with regard to the benchmark model, the results depend on the origin of the forecast.