高频交易(HFT)对市场价格波动的快速捕捉和高效套利能力受到现在金融市场的广泛关注。传统方法在处理高频数据时通常缺乏全面建模能力,因其数据复杂、噪声干扰以及趋势变化迅速等特性,对实时决策的精准和模型解释性提出了巨大挑战。针对上述问题,本文提出了一种基于结构信息嵌入与动量优化的在线学习模型(SOC, Structural Online Classification)。SOC模型通过多层次特征工程构建时间序列特征、局部极值特征和全局关系特征,以充分嵌入高频交易数据的结构信息;结合双层聚类方法(K-Means结合层次聚类)对高维特征进行降维与优化,显著增强分类器的透明性与可解释性。利用L2正则化与协方差正则化策略改良模型,结合Adam优化器实现高效的动量优化。本文在沪深300指数、UR股票等高频数据集上对SOC模型进行了性能验证。实验结果表明,SOC模型在分类准确性、均方误差和F1值等多个指标上均表现优异,其中沪深300指数的分类准确率达到98.73%,显著优于传统在线学习模型。通过对比传统神经网络模型与在线学习模型(SOC)在分类与回归任务中的表现,定量分析了在线学习模型的改进方向。实验结果表明,SOC模型在预测精度、泛化能力及内存效率(内存用量减少67.5%)等方面均显著优于传统模型,验证了在线学习机制在动态数据环境下的有效性。High-frequency trading (HFT) has drawn extensive attention in the current financial market due to its rapid capture of market price fluctuations and efficient arbitrage capabilities. Traditional methods often lack comprehensive modeling capabilities when dealing with high-frequency data, as the data is complex, subject to noise interference, and characterized by rapid trend changes, posing significant challenges to the accuracy of real-time decision-making and model interpretability. To address these issues, this paper proposes a structural online classification model (SOC) based on structural information embedding and momentum optimization. The SOC model constructs time series features, local extremum features, and global relationship features through multi-level feature engineering to fully embed the structural information of high-frequency trading data. It combines a two-layer clustering method (K-Means combined with hierarchical clustering) to reduce the dimensionality and optimize high-dimensional features, significantly enhancing the transparency and interpretability of the classifier. The model is improved using L2 regularization and covariance regularization strategies, and the Adam optimizer is employed to achieve efficient momentum optimization. The performance of the SOC model was verified on high-frequency datasets such as the CSI 300 Index and UR stocks. Experimental results show that the SOC
暂无评论