不同时间窗下铁矿石期货价格预测精度研究——基于CNN-LSTM注意力模型的分析

Research on the accuracy of iron ore futures price forecasting under different time horizon——Analysis based on CNN-LSTM-Attention model

利用时间窗提升铁矿石期货价格预测精度对铁矿石期货市场平稳发展具有重要意义。本文选取2013年10月至2021年12月铁矿石期货价格及同期相关数据,采用STL分解方法对铁矿石期货价格进行特征分析,构造基于自注意力机制的CNN-LSTM模型,预测铁矿石期货价格并进行对比分析。结果表明:将铁矿石期货季节性规律应用于时间窗可以提升铁矿石期货价格预测结果准确性。在4、7、30、365天时间窗下,最佳预测结果是4天时间窗。模型预测结果的平均绝对误差MAE值为11.5,相较于LSTM、LSTM-ATT、CNN-LSTM基准模型分别降低了32.70%、19.12%、22.28%。构建模型具有较好的泛化性,MAE在7天、30天、365天时间窗下均为最低。基于此,应关注价格的时间窗特征,完善铁矿石期货市场环境,推动铁矿石现货市场保供稳价。

The use of time horizon to improve the accuracy of iron ore futures price forecasting is important for the smooth development of iron ore futures market.In this paper,we select iron ore futures prices and related data of from October 2013 to December 2021,use the STL decomposition method to characterize iron ore futures prices,construct a CNN-LSTM model based on the self-attention,predict iron ore futures prices,and conduct a comparative analysis.The results show that applying the seasonal pattern of iron ore futures to the time horizon can improve the accuracy of iron ore futures price prediction results.On 4,7,30 and 365-day time horizons,the best prediction result is4-day time horizon.The MAE value of the model is 11.5,which is 32.70%,19.12%,and 22.28%lower than that of LSTM,LSTM-ATT,and CNN-LSTM benchmark models.The model has good generalization,and the MAE is the lowest than 7,30 and 365-day time horizons.Based on this,attention should be paid to the price time horizon characteristics,improve the iron ore futures market environment,and promote the supply and price stability of iron ore spot market.