Risk-averse Learning with Non-Stationary Distributions