<p>We present Embedding Outlier Detection (EmbOD), a novel, flexible, unsupervised outlier detection (OD) method. Our OD technique transforms the original data into a new feature space, an embedding, based on ensembles of unsupervised methods. Moreover, we create the embedding with ensembles of hierarchical clustering, where we measure the distance between inliers and outliers. Our method produces a predictive model, i.e., once we have trained the model, we can query it about new data points. As part of our experiments, we analyze several configurations for EmbOD and provide statistical analysis. We also compared EmbOD with several classic and new OD methods. For these tests, we used numerous real datasets. Our results include metrics to compare performance and statistical analysis to compare the methods. Finally, our experimental results show that EmbOD ranks among the top methods when we consider both the precision score and the area under the receiver operating characteristic curve, combined with the Wilcoxon-Holm test. Furthermore, we use the Wilcoxon-Holm test to analyze EmbOD hyperparameters with the goal of finding an emerging setting.</p>