<p>Modeling and forecasting of the geomagnetic variation are important research topics concerning geomagnetic navigation and space environment monitoring. We propose a combined forecasting model using a dynamic recursive neural network called echo state network (ESN), the method of complementary ensemble empirical mode decomposition (EEMD) and the complexity theory of sample entropy (SampEn). Firstly, we use EEMD-SampEn to decompose the geomagnetic variation time series into many series of geomagnetic variation subsequences whose complexity degrees are transparently different. Then, we use ESN to build a forecasting model for each subsequence, selecting the optimal model parameters. Finally, we use the real data collected from the geomagnetic observatory to conduct simulations. The results show that the forecasting value of the combined model can closely conform to the tendency of geomagnetic variation field, and is superior to the least square support vector machine (LSSVM) model. The mean absolute error of the model for three-hour forecasting is less than 1.40nT when <i>Kp</i> index is less than 3.</p>