Online Conditional Vine Copulas: Forecasting Electricity Demand

Abstract

We propose an online multivariate distributional forecasting model in which we condition the parameters of a vine copula on covariates. An online algorithm updates the parameters sequentially using only the most recent observations. Existing approaches require full batch re-estimation using all available data at each update step, which makes them computationally slow in large-scale forecasting applications and less suitable for settings in which dependence structures change over time. Simulation results show that the online procedure achieves forecasting performance comparable to that of the batch estimator, while reducing computation time substantially. We illustrate the method in a forecasting application to electricity net-demand across five regions of Great Britain and show that (i) the conditional vine copula model outperforms unconditional Gaussian and vine copula benchmarks and (ii) that online estimation improves forecast accuracy when batch re-estimation is not available. The algorithm is implemented in the Python package ondil.