Exoskeletons offer many possibilities within the contexts of manual labour as well as rehabilitation. A lot of effort has gone into improving the usability and efficiency of exoskeletons in order to extent their use cases. In order to control an exoskeleton, it is necessary to predict the intended movement of the user. A basic operation within the domain of intention estimation is that of elbow joint angle estimation. We have chosen to work with this prediction problem with measurements from a Force Myography (FMG) sensor armband. This problem is traditionally solved by training a machine learning model on data collected from a single person, used for intention estimation for specifically that same person. We instead wish to optimize this approach such that the solution can quickly be adapted to different users, reducing the amount of required training data for each novel person. Along this line, this paper proposes a novel boosting approach, Exo-Ada, specifically designed for use with exoskeletons in order to efficiently transfer to the domain of a new user. Exo-Ada is built on top of 2-Stage TrAdaBoost and uses a learner component made up by a rectified Convolutional Neural Network (CNN), making use of reference points and dilated convolutions. Exo-Ada outperforms several baselines on a test on FMG sensor measurements from multiple people.