The use of the electrical activity from the muscles may provide a natural way to control exoskeletons or other robotic devices seamlessly. The major challenges to achieve this goal are human motor redundancy and surface electromyography (sEMG) variability. The goal of this work is to find a feature extraction and classification procedures to estimate accurately elbow angular trajectory by means of a NARX Neural Network. The processing time-step should be small enough to make it feasible its further use for online control of an exoskeleton. In order to do so we analysed the Biceps and Triceps Brachii data from an elbow flexo-extension Coincident Timing task performed in the horizontal plane. The sEMG data was pre-processed and its energy was divided in five frequency intervals that were fed to a Nonlinear Auto Regressive with Exogenous inputs (NARX) Neural Network. The estimated angular trajectory was compared with the measured one showing a high correlation between them and a RMSE error maximum of 7 degrees. The procedure presented here shows a reasonably good estimation that, after training, allows real-time implementation. In addition, the results are encouraging to include more complex tasks including the shoulder joint.