Portfolio selection under non-gaussianity and systemic risk: A machine learning based forecasting approach