We study the problem of anomaly detection (AD) and explanation algorithms in multivariate time series. We introduce the main details of the state-of-the-art AD architecture, discussing several limitations of the existing techniques and the aspects we need to consider to address the open challenges. To date, we dispose of a plethora of AD techniques, which exhibit shaky performance depending on the data characteristics. In this context, we must unrealistically assume that the user would have a deep understanding of the data and she would trust the actions taken by the applied model. We thus envision algorithms that automatically learn complex data dependencies irrespective of time shift and scale of multivariate time series variables. Hence, this must enable users to interpret the evolution of the studied events to interact and refine the anomaly outcomes and perform an effective anomaly causal analysis.