Optimal reconciliation with immutable forecasts

The practical importance of coherent forecasts in hierarchical forecasting has inspired many studies on forecast reconciliation. Under this approach, base are produced for every series the hierarchy and subsequently adjusted to be a second reconciliation step. Reconciliation methods have been shown improve accuracy but will generally adjust series. However, an operational context, it is sometimes necessary or beneficial keep some variables unchanged after In paper, we formulate methodology that keeps pre-specified subset “immutable”. contrast existing approaches, these immutable need not all come from same level hierarchy, our method can also applied grouped hierarchies. We prove approach preserves unbiasedness forecasts. Our account correlations between errors ensure non-negativity perform empirical experiments, including application large-scale online retailer’s sales, assess proposed methodology’s impacts.