In this paper, we propose a class of robust M-estimators for the orthogonal generalized autoregressive conditional heteroscedastic (GARCH) model. The method involves the estimation of only univariate GARCH models and hence easy to estimate and does not put additional constraints on the model. The forecasting performance of the class of robust estimators in predicting correlation and value-at-risk using various evaluation measures are investigated. We found empirical evidences of the better predictive potential of estimators such as least absolute deviation and B-estimator over the widely used quasi-maximum likelihood estimator when the error distribution is heavy-tailed and asymmetric. Applications to real data sets are also presented.