As the crude oil price movement is influenced by increasingly diverse range of risk factors in the crude oil markets, the crude oil price exhibits more complex nonlinear behavior and poses higher level of risk for investors than ever before. To model the crude oil risk changes at higher level of accuracy, we proposed a new multiscale approach to estimate Value at Risk. It takes advantage of Variational Mode Decomposition (VMD) model to extract and model the main risk factors in the multiscale domain, where the individual characteristics of these risk factors are modeled using ARMA-GARCH models. The Convolutional Neural Network (CNN) based nonlinear ensemble model is employed to aggregate these risk forecasts as the ensemble members and produce the optimal ensemble forecasts. Empirical evaluation of the performance of the proposed model has been conducted using the extensive dataset, constructed with daily price observations in the major crude oil markets, Experiment results confirm that the proposed risk forecasting models produce an improved forecasting accuracy for the typical risk measures such as Value at Risk.