Gold futures, as an essential hedge asset, have received much attention. In this paper, the VMD-reconstruction-integration framework combining rolling windows is proposed to predict gold futures prices. The Fine to Coarse (FTC) method is used to reconstruct the decomposed IMFs, and the non-parametric regression (NR) model and extreme learning machine (ELM) model are used to predict the reconstructed long-term trend term and short-term disturbance term respectively, which effectively avoided the information leakage problem in the decomposition process and improved the prediction accuracy of gold futures price. The empirical results show that after avoiding the decomposition leakage problem, the model under the decomposition framework still has a certain improvement effect. In addition, the R-VMD-NR/ELM model has the best prediction effect, and compared with the R-VMD-NR/SVR model, the six evaluation criteria improved by 0.8883, 9.7188, 0.0021, 1.0492, 0.008, and 0.02, respectively.