基于LSTM和GRU神经网络的上海原油期货价格预测研究

Forecasting Shanghai Crude Oil Futures Prices with LSTM and GRU Neural Networks

摘要: 本研究针对上海原油期货价格预测问题，系统对比了ARIMA、VARIMA、LSTM与GRU四类模型的预测性能。实证结果表明：传统线性模型ARIMA和VARIMA因受限于线性假设，在测试集上预测误差显著(RMSE > 9.5)，尤其对突发性波动事件响应滞后；而深度学习模型展现出显著优势，其中GRU以双门控耦合机制实现效率与精度的最优平衡(MAE = 4.6928，参数量较LSTM减少25.6%)，LSTM则凭借三门分立结构在长期趋势捕捉中表现稳健(R² = 0.9647)。消融实验进一步验证了门控设计的必要性——移除LSTM遗忘门导致误差激增50%，印证了其在噪声过滤中的核心作用。研究成果为油价预测提供了兼具精度与效率的深度学习解决方案，并为构建智能化风控体系提供了理论依据。

Abstract: This study systematically compares the predictive performance of four models—ARIMA, VARIMA, LSTM, and GRU—for forecasting Shanghai crude oil futures prices. Empirical results demonstrate that traditional linear models (ARIMA/VARIMA), constrained by their linear assumptions, exhibit significant prediction errors on the test set (RMSE > 9.5), particularly showing delayed responses to abrupt volatility events. In contrast, deep learning models demonstrate remarkable advantages: GRU achieves an optimal balance between efficiency and accuracy through its dual-gate coupling mechanism (MAE = 4.6928, with 25.6% fewer parameters than LSTM), while LSTM exhibits robust performance in capturing long-term trends owing to its triple-gate structure (R² = 0.9647). Ablation experiments further validate the necessity of gating design—removing the forget gate from LSTM increases errors by 50%, confirming its critical role in noise filtering. The research outcomes provide a deep learning solution that combines accuracy and efficiency for oil price forecasting, offering a theoretical foundation for constructing intelligent risk management systems.