How are forecast generated?

We use Facebook Prophet (http://facebook.github.io/prophet/) for forecasting, which is a time-series based model. The business context training uses 180 days of customer history. In addition to cost, anomalies are used as an additional signal while forecasting, and then we predict up to 1 year in the future at monthly, quarterly and yearly aggregation.

The forecast at higher level business contexts is not equal to the sum of its child contexts in the hierarchy, because the forecast is not aggregated. And each underlying curve may have a different pattern, but when they are superimposed on each other as in a higher-level context, it may be on a different curve with different signals, personalities, and patterns. So until lower-level contexts are forecasted and higher-level contexts are aggregated from them, it is highly unlikely mathematically to give forecasts that sum up to the forecasts of high-level contexts.