XGBoost与ARIMA在股票画像中的应用研究

Research on the Application of XGBoost and ARIMA in Stock Portrait

摘要: 本研究基于中国石油2023~2024年的实时数据，构建了ARIMA和XGBoost模型，以评估和比较它们在股票趋势预测中的性能。结果显示，XGBoost在训练集上具有强大的非线性拟合能力，但在测试集上错误波动性更高。ARIMA模型虽解释能力较弱，但在短期预测中更稳定，适合低延迟场景。建议在高频交易中使用ARIMA模型(响应时间小于50毫秒)，而将XGBoost优先用于事件驱动策略。本研究的局限性在于仅使用单一资产和静态时间窗口数据。未来研究将引入多因素(如VIX指数、情绪数据)及动态滚动预测框架以提升混合模型的泛化能力。本研究为优化混合模型在特征工程、权重分配及评估系统方面的理论支持，并为金融机器学习从“模型堆砌”向“机制驱动型混合模型”转型提供参考。

Abstract: This study, based on PetroChina’s 2023~2024 real-time data, builds ARIMA and XGBoost models to evaluate and compare their performance in stock trend prediction. Results show XGBoost has strong non-linear fitting ability on the training set but higher error volatility on the test set. ARIMA, though less explanatory, is more stable in short-term forecasting, suiting low-latency scenarios. It’s recommended to use ARIMA (with a response time of less than 50 ms) for high-frequency trading and prioritize XGBoost for event-driven strategies. This study’s limitations lie in its single-asset and static time-window data. Future work will introduce multi-factors (e.g., VIX index, sentiment data) and dynamic rolling prediction frameworks to enhance the hybrid model’s generalization. The study offers theoretical support for optimizing hybrid models in feature engineering, weight allocation, and evaluation systems, and references for shifting financial machine learning from “model piling” to “mechanism-driven hybrids”.