Forecasting is not prophecy: dealing with high-dimensional probabilistic forecasts in practice
Invited talk for ISI-WSC 2019 in Kuala Lumpur
Many large organizations need to forecast huge numbers of related time series every week. Manufacturing companies forecast product demand to plan their supply chains; call centres forecast call volume to inform staff scheduling; technology companies forecast web traffic to maintain service levels; energy companies forecast electricity demand to prevent blackouts. In each case, what is required is a high-dimensional probabilistic forecast describing multivariate quantiles of the uncertain future, not a vector of point forecasts.
This raises several difficulties:
- it is analytically and computationally challenging to produce probabilistic forecasts for very high-dimensional time series;
- users find multivariate probability distributions difficult to use and interpret;
- the predictive accuracy of a high-dimensional probability distribution is not easy to measure.
I will discuss these problems and how they can be tackled.
R packages available at tidyverts.org.