A0673
Title: Gradient-boosted conditional vine copula models for multivariate temperature forecasting
Authors: David Jobst - University of Hildesheim (Germany) [presenting]
Annette Moeller - Bielefeld University (Germany)
Juergen Gross - University of Hildesheim (Germany)
Abstract: Weather forecasting currently relies on ensemble forecasts to address uncertainties in the future atmospheric states. However, these forecasts may exhibit biases and underdispersion. Therefore, the ensemble model output statistics (EMOS) method is frequently employed to separately postprocess the ensemble forecasts for each lead time. Unfortunately, the independence assumption among the postprocessed predictive distributions of different lead times is not always fulfilled. To restore these temporal dependencies, the postprocessed univariate distributions are combined with copula-based approaches, such as the ensemble copula coupling (ECC) or the Gaussian copula approach (GCA). As an alternative approach, a conditional vine copula (CVC) model is proposed, where the coefficients of the conditional bivariate copulas are estimated via gradient-boosting. In a case study, the suggested method is compared with ECC and GCA for the multivariate postprocessing of temperature forecasts for five lead times. The results show, that CVC provides an improvement in terms of calibration and identifies temporal dependencies better than the benchmark methods. Furthermore, the approach offers valuable insights into the covariate selection for the dependence model.
