Training of reconciliation weights for probabilistic forecasts to optimise total energy (or variogram) score using Stochastic Gradient Descent with automatically differentiated gradients. See Panagiotelis, Gamakumara, Athanasopoulos and Hyndman, (2020) https://www.monash.edu/business/ebs/research/publications/ebs/wp26-2020.pdf for a description of the methods.
Maintainer: Anastasios Panagiotelis anastasios.panagiotelis@sydney.edu.au (ORCID)
This package carries out probabilistic forecast reconciliation via score
optimisation using the method described by wp;textualProbReco. Given incoherent (base) probabilistic forecasts formed over a
training data set, the function scoreopt finds linear
reconciliation weights that optimise total
score scoresProbReco over the training
data. Currently the energy score and variogram score are implemented. The
optimisation is carried out using the Adaptive Moments (Adam) variant
of Stochastic Gradient Descent developed
by adam;textualProbReco. Tuning parameters for this
optimisation can be
set using scoreopt.control. The gradients are found
using the automatic differentiation libraries of the Stan
project stanProbReco.
A version of the function that allows for simpler inputs is provided
by inscoreopt. Rather than using arguments that are
lists of realisations and lists of
functions, a matrix of data and a matrix of (point) predictions are the main
arguments. This function is less general
than scoreopt in two ways. First, there are only a
limited range of options for producing base forecasts (either from Gaussian
distributions or bootstrapping). Second, the scores are evaluated
using in-sample predictions rather than genuine forecasts.
Useful links: