Alternatively, if using a neural network model for forecasting, you can train a single network to output multiple predictions (direct forecasts) for each future time step. The network could then learn about relationships between the forecasts.