Forecasting Compositional Time Series with Exponential Smoothing Methods