Metal price movement is a complicated process with significant fluctuations, exhibiting nonlinear characteristics. With the proposed Heterogeneous Market Hypothesis (HMH), we argue that the macroscale nonlinear price movement consists of a diverse range of data components such as chaotic and multi-scale data characteristics at the microscopic level. In order to model the hidden data features, we propose a new Curvelet based forecasting algorithm. To model the dynamic chaotic data characteristics in the original time series, we project the original time series into the reconstructed phase space using the time delay method. The Curvelet denoising method is used for separating and reducing the disruption from noise in the transformed state space. Results from empirical studies conducted in the major metal markets suggest that the proposed model achieves statistically more robust and superior performance than traditional benchmark model.