Probabilistic reconciliation by conditioning


Giorgio Corani, IDSIA (Dalle Molle Institute for Artificial Intelligence), Switzerland


I will show how to perform probabilistic reconciliation via conditioning. First we create an incoherent joint distribution p(B,U) on bottom and upper time series, based on the base forecast. Then we condition it on the constraint AB=U, where A is the aggregation matrix specifying how bottom-level series are aggregated to form the upper ones. The conditioning can be solved analytically for Gaussian base forecasts; in this case the reconciled distribution has the same point forecast and variance of MinT. I will then show how to compute the reconciled distribution via sampling for count time series and I will discuss some properties of the reconciled distributions.