D-Vine GAM Copula based Quantile Regression with Application to Ensemble Postprocessing

TL;DR

This paper introduces GAM-DVQR, an extension of D-vine copula quantile regression that incorporates covariate effects via GAMs, improving weather forecast postprocessing by modeling complex dependencies and temporal effects.

Contribution

The paper proposes GAM-DVQR, a novel method that parametrizes copulas with covariates using GAMs, enhancing the modeling of temporal and spatial effects in probabilistic weather forecasting.

Findings

GAM-DVQR outperforms EMOS and EMOS-GB in temperature forecast postprocessing.

The method effectively detects time-dependent correlations and relevant predictors.

Using a static training period improves model sustainability.

Abstract

Temporal, spatial or spatio-temporal probabilistic models are frequently used for weather forecasting. The D-vine (drawable vine) copula quantile regression (DVQR) is a powerful tool for this application field, as it can automatically select important predictor variables from a large set and is able to model complex nonlinear relationships among them. However, the current DVQR does not always explicitly and economically allow to account for additional covariate effects, e.g. temporal or spatio-temporal information. Consequently, we propose an extension of the current DVQR, where we parametrize the bivariate copulas in the D-vine copula through Kendall's Tau which can be linked to additional covariates. The parametrization of the correlation parameter allows generalized additive models (GAMs) and spline smoothing to detect potentially hidden covariate effects. The new method is called GAM-DVQR, and its performance is illustrated in a case study for the postprocessing of 2m surface temperature forecasts. We investigate a constant as well as a time-dependent Kendall's Tau. The GAM-DVQR models are compared to the benchmark methods Ensemble Model Output Statistics (EMOS), its gradient-boosted extension (EMOS-GB) and basic DVQR. The results indicate that the GAM-DVQR models are able to identify time-dependent correlations as well as relevant predictor variables and significantly outperform the state-of-the-art methods EMOS and EMOS-GB. Furthermore, the introduced parameterization allows using a static training period for GAM-DVQR, yielding a more sustainable model estimation in comparison to DVQR using a sliding training window. Finally, we give an outlook of further applications and extensions of the GAM-DVQR model. To complement this article, our method is accompanied by an R-package called gamvinereg.