基于BSFinformer模型的金融数据特征选择及预测

Feature selection and prediction of financial data based on BSFinformer model

金融领域的长时间序列预测正在面对复杂的市场和众多金融产品的挑战,传统的时序数据预测方法在处理线性分布数据时表现良好,但对于特征参数冗余和非线性长序列金融产品数据的预测效果有限.为了解决这一问题,提出一种长时间序列预测方法BSFinformer(Boruta-SHAP+Finformer),利用金融数据的时间相关性并综合运用BorutaSHAP,Finformer等技术来完成特征选择及预测功能.该方法首先引入Boruta-SHAP模块,利用XgBoost和SHAP分析方法进行特征选择,从给定的特征集中识别出与金融时间序列预测任务相关的重要特征,并解释这些特征对预测的影响.其次,利用Transformer结构和自注意力机制,改进为Finformer模块,将长序列金融数据分解为趋势、周期和残差成分,结合稀疏自注意力机制.在多个真实金融数据集上进行了实验评估.实验结果显示,BSFinformer对金融产品的价格预测表现出优异的性能,与其他预测方法相比,能准确捕捉长期趋势和周期性来实现高质量的预测.具体地,和传统的Transformer模型相比,在三个实验数据集上,BSFinformer的均方误差分别降低了52%,16%和19%,平均绝对误差分别降低了34%,25%和11%,为金融数据的长期时间序列预测提供了一种有效的解决方案.

The long⁃term series prediction in the financial domain faces challenges due to complex markets and numerous financial products.Traditional methods in time series forecasting perform well in handling linear distributed data,but their effectiveness is limited when dealing with redundant feature parameters and nonlinear data of long sequence financial products.To address this issue,this study proposes a method in long⁃term series prediction called BSFinformer(Boruta⁃SHAP+Finformer).This method leverages the time correlation of financial data and integrates techniques such as Boruta⁃SHAP and Finformer to accomplish feature selection and prediction tasks.Firstly,the Boruta⁃SHAP module is introduced,which utilizes such analytical methods as XgBoost and SHAP for feature selection.It identifies important features related to tasks of financial time series prediction from the given feature set and explains the impact of these features on the prediction.Secondly,the Finformer module is developed by improving the Transformer structure and incorporating self⁃attention mechanisms.It decomposes long sequence financial data into trend,cycle,and residual components,and combines sparse self⁃attention mechanisms.The BSFinformer model is evaluated on multiple real financial datasets through experiments.The experimental results demonstrate that the BSFinformer model exhibits excellent performance in price prediction of financial products.Compared to other forecasting methods,the BSFinformer model accurately captures long⁃term trends and periodicity to achieve high⁃quality predictions.Specifically,compared to the traditional Transformer model,the BSFinformer model reduces Mean⁃Square Error by 52%,16%and reduces 19%,and Mean Absolute Error by 34%,25%and 11%on the three datasets,respectively.It provides an effective solution for long⁃term series prediction of financial data.