Abstract
We analyse the predictive and the forecasting ability of various Generalized Autoregressive Score (GAS) and GARCH frameworks for European Union Allowances (EUAs) daily returns (EUAs returns) for the period 22/04/2005-28/02/2019. We further examine the impact of different distributional assumptions on risk prediction. The Model Confidence Set (MCS) is employed to compare and select a superior predictive model of Value-at-Risk (VaR) thresholds. We find that GAS under skewed t-student error distribution and gjr-GARCH under general error distribution deliver excellent results for the Value-at-Risk (VaR) prediction for EUA at 1% and 5% levels, respectively. These results are robust with respect to three back-testing procedures (i.e., Unconditional Coverage, Conditional Coverage, and Dynamic Quantile tests). These results are of particular importance for the development of EUA pricing policies and risk management strategies.