In this work, we develop novel methods for Hierarchical Time Series forecasting. A key goal is model robustness so that node failures in the hierarchy do not particularly affect the results. Our is general enough to be extended to most optimization based approaches so that it is robust to node failures. We develop an optimization based method where we derive a formulation which produces robust solutions that are reconciled in sample. We test our method against real-world datasets ranging from transportation to demand prediction. We demonstrate improvement over existing, non-robust approaches.