We consider sequential data that is sampled from an unknown process, so that the data are not necessarily iid. We develop a measure of generalization for such data and we consider a recently proposed approach to optimizing hyper-parameters, based on the computation of the gradient of a model selection criterion with respect to hyper-parameters. Hyper-parameters are used to give varying weights in the historical data sequence. The approach is successfully applied to modeling the volatility of Canadian stock returns one month ahead.