A new perspective on non-ferrous metal price forecasting: An interpretable two-stage ensemble learning-based interval-valued forecasting system

Abstract

An accurate non-ferrous metal price prediction model is critical for formulating national economic policies, planning company production, and mitigating risk. Existing research improves the performance of prediction models based on point data but neglects the value of interval data and model interpretability, resulting in suboptimal predictions. Hence, this study proposes an interval-valued forecasting system for non-ferrous metal prices via interpretable two-stage ensemble learning. An interval-valued data preprocessing module is designed to improve predictive ability and enhance modeling diversity in terms of data by introducing various interval-valued mapping strategies. To enhance the modeling diversity of the predictors, a meta-predictor module that incorporates four advanced deep-learning models that produce various sub-predictors is proposed. A two-stage ensemble learning module is developed to obtain final interval-valued non-ferrous metal prices based on all sub-predictors. In the first stage, based on temporal fusion transformers, different deep-learning models are combined to reduce the bias in individual predictors. In the second stage, based on an attention mechanism, different interval-valued mapping strategies are combined to improve forecasting performance. Multiple comparative experiments and analyses are conducted using real non-ferrous metal market data. In an empirical study, the proposed system achieved the best results. Taking a copper dataset as an example, the system results for the IMAPE, IRMSE, IARV, and UI were 0.57826 %, 62.51197, 0.02147, and 0.14651, respectively. The results show that the proposed system not only outperforms both individual and advanced ensemble models in terms of accuracy and robustness but also offers valuable interpretable insights for improving interval-valued forecasting power.