Hedging the Drift: Learning to Optimize under Non-Stationarity