Forecasting the distribution of aggregated time series: a bootstrap approach