Identifying potential future elite athletes is important in many sporting events. The successful identification of potential future elite athletes at an early age would help to provide high-quality coaching and training environments in which to optimize their development. However, a large variety of different skills and qualities are needed to succeed in elite sports, making talent identification generally a complex and multifaceted problem. Due to the rarity of elite athletes, datasets are inherently imbalanced, making classical statistical inference difficult. Therefore, we approach talent identification as an anomaly detection problem. We trained a nonlinear one-class support vector machine (one-class SVM) on a dataset (N=951) collected from 14-year-old junior soccer players to detect potential future elite players. The mean area under the receiver operating characteristic curve (AUC-ROC) over the tested hyperparameter combinations was 0.763 (std 0.007). The most accurate model was obtained when physical tests, measuring, for example, technical skills, speed, and agility, were used. According to our results, the proposed approach could be useful to support decision-makers in the process of talent identification.