Non-ferrous metals play a significant role in social development. It is important for policy makers and entrepreneurs to forecast non-ferrous metals price accurately. However, existing methods are hard to obtain satisfactory results because the fluctuation rule of non-ferrous metal price is increasingly complex. Therefore, it is necessary to develop more accurate and stable forecasting method. In this paper, a multi-model fusion based non-ferrous metal price forecasting method is proposed. The dual-stage signal decomposition algorithm which combines complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and variational mode decomposition (VMD) is innovatively introduced into non-ferrous metal price forecasting. First, CEEMDAN is used to decompose the original price into several subsequences. Second, the most complex subsequence with maximum sample entropy (SE) is further decomposed by VMD. Dual-stage decomposition reveals essential features such as long-term trend and periodic fluctuations hidden in original sequence and thus lowers prediction difficulty. Besides, particle swarm optimization (PSO) is used to select optimal parameters for VMD. Finally, all subsequences are predicted by long short-term memory network (LSTM) and the results are integrated as the final prediction result. In the empirical study of London Metal Exchange (LME)'s copper, aluminum and zinc price, the proposed method is superior to all benchmark methods in terms of RMSE, MAE and MAPE. The results demonstrate that the proposed method is effective and robust. © 2022 Elsevier Ltd
