Multivariate time series modeling and prediction problems are abundant in\nmany machine learning application domains. Accurate interpretation of such\nprediction outcomes from a machine learning model that explicitly captures\ntemporal correlations can significantly benefit the domain experts. In this\ncontext, temporal attention has been successfully applied to isolate the\nimportant time steps for the input time series. However, in multivariate time\nseries problems, spatial interpretation is also critical to understand the\ncontributions of different variables on the model outputs. We propose a novel\ndeep learning architecture, called spatiotemporal attention mechanism (STAM)\nfor simultaneous learning of the most important time steps and variables. STAM\nis a causal (i.e., only depends on past inputs and does not use future inputs)\nand scalable (i.e., scales well with an increase in the number of variables)\napproach that is comparable to the state-of-the-art models in terms of\ncomputational tractability. We demonstrate our models' performance on two\npopular public datasets and a domain-specific dataset. When compared with the\nbaseline models, the results show that STAM maintains state-of-the-art\nprediction accuracy while offering the benefit of accurate spatiotemporal\ninterpretability. The learned attention weights are validated from a domain\nknowledge perspective for these real-world datasets.\n