We propose a novel dynamic approach to forecast the weights of the global minimum variance portfolio (GMVP). The GMVP weights are the population coefficients of a linear regression of a benchmark return on a vector of return differences. This representation enables us to derive a consistent loss function from which we can infer the optimal GMVP weights without imposing any distributional assumptions on the returns. In order to capture time variation in the returns’ conditional covariance structure, we model the portfolio weights through a recursive least squares (RLS) scheme as well as by generalized autoregressive score (GAS) type dynamics. Sparse parameterizations combined with targeting towards nonlinear shrinkage estimates of the long-run GMVP weights ensure scalability with respect to the number of assets. An empirical analysis of daily and monthly financial returns shows that the proposed models perform well in- and out-of-sample in comparison to existing approaches.