统计与数据科学学院知识库

        在全球资源日益减少和近几年国际安全局势愈发紧张的背景下，国际原油市 场也随之波动不安，使得国际原油价格预测成为当下备受关注的课题之一。随着 近几年国际原油价格波动频率显著增加，其价格变化具有高度不确定性。作为“工 业血液”，原油在工业生产、航空航天等领域中扮演着关键角色，是保障国家经 济和社会安全发展的重要资源。因此，国际原油价格的波动必然对各产业的发展 和升级产生深远影响，加强国际原油价格预测工作，提供原油价格变动的前瞻性 洞察，提高预测准确性和可靠性对国际原油市场稳定、促进各国工业、经济发展 具有重要意义。

       本文收集了1997年到2025年的全球APSP原油价格数据以及相关影响因素 数据，提出了一种基于STL分解与RF-XGBoost深度集成学习模型的新型国际原 油价格预测方法。引入Lasso回归嵌入选择重要特征后，在原始序列的基础上建 立了不同的单一机器学习模型和组合机器学习模型，单一模型RNN的预测性能 最优（MSE：0.0357，R2：0.8181），但模型拟合效果仍有上升空间。为尽可能 提高预测精度，本文利用STL分解将复杂的原油价格序列数据拆解为趋势、季 节性和残差三部分，有效降低了数据复杂性，使特征更易于提取。然后，针对趋 势子序列建立了随机森林RF模型，经Optuna超参数优化后的趋势模型更是表 现出卓越的预测性能。另外，在处理季节子序列和残差子序列时发现，国际原油 价格序列具有十分规律的季节周期波动；残差子序列表现为平稳的白噪声。最后， 将RF预测趋势和季节性、残差通过XGBoost集成融合，预测的综合性能有效提 高（MSE：0.0048，R2：0.9999），证实了STL-RF-XGBoost 模型在国际原油价 格预测任务中的可行性与优良性。

        Against the backdrop of increasingly scarce global resources and the growing tension in international security in recent years, the international crude oil market has experienced significant fluctuations, making the prediction of international crude oil prices a topic of intense interest. In recent years, the frequency of crude oil price volatility has markedly increased, exhibiting high uncertainty. As the "lifeblood of industry," crude oil plays a critical role in industrial production, aerospace, and other fields, serving as a vital resource for ensuring national economic and social security. Consequently, fluctuations in international crude oil prices inevitably exert profound impacts on the development and upgrading of various industries. Strengthening the prediction of crude oil prices, providing forward-looking insights into price changes, and improving the accuracy and reliability of forecasts are of great significance for stabilizing the international crude oil market and promoting industrial and economic development worldwide.

       This paper collects global APSP crude oil price data and related influencing factor data from 1997 to 2025 and proposes a novel international crude oil price prediction method based on an STL decomposition and RF-XGBoost deep ensemble learning model. After introducing Lasso regression for feature selection, different single machine learning models and combined machine learning models were constructed on the original sequence. Among them, the single RNN model demonstrated the best predictive performance (MSE: 0.0357, R²: 0.8181), but there remained room for improvement in model fitting. To maximize prediction accuracy, this study employs STL decomposition to break down the complex crude oil price sequence into three components: trend, seasonality, and residual, effectively reducing data complexity and making features easier to extract. Subsequently, a Random Forest (RF) model was established for the trend subsequence, and after Optuna hyperparameter optimization, the trend model exhibited outstanding predictive performance. Additionally, when analyzing the seasonal and residual subsequences, it was found that the international crude oil price sequence exhibits highly regular seasonal cyclical fluctuations, while the residual subsequence behaves as stationary white noise. Finally, the integration of RF-predicted trends with seasonality and residuals via XGBoost significantly improved the overall predictive performance (MSE: 0.0048, R²: 0.9999), confirming the feasibility and superiority of the STL-RF-XGBoost model in international crude oil price prediction tasks.