基于多步特征选择的股市预测方法

Chinese Stock Prediction Method Based on Multi-Step Feature Selection

摘要: 智能金融预测模型在投资决策中发挥着重要作用。为解决高维数据引起的维数灾难和过拟合问题，有效提高多元非线性金融时序预测精度，设计了一种基于多步特征选择的GA-LSTM预测框架。首先，通过基于过滤和嵌入的特征选择方法从大量指标中筛选有效股票影响因子，然后将剩余变量输入遗传算法优化后的长短期记忆神经网络来预测股票收盘价。为了验证所提模型的有效性，将此模型与传统降维模型PCA、LASSO和预测模型ARIMA、MLP、SVR、RNN、GRU对比分析，在中国银行数据集上的实验结果表明：基于机器学习新的组合方法不仅可以大规模降维，预测误差MSE、MAPE也低于对比模型，显著提高了预测精度。最后，将此模型应用在中国不同行业代表性的股票中取得较好预测效果，再次证明此模型在金融时间序列智能特征提取和预测上的应用价值。

Abstract: Intelligent financial forecasting model plays an important role in investment decision-making. In order to solve the problem of dimensional disaster and overfitting caused by high-dimensional data and effectively improve the prediction accuracy of multivariate nonlinear financial time series, a GA-LSTM prediction framework based on multi-step feature selection is designed. First, the effective stock impact factors are screened from a large number of indicators by the feature selection method based on filtering and embedding, and then the remaining variables are input into the long-short-term memory neural network optimized by the genetic algorithm to predict the stock closing price. In order to verify the effectiveness of the proposed model, this model is compared with traditional dimensionality reduction models PCA, LASSO and prediction models ARIMA, MLP, SVR, RNN, and GRU. The combined method can not only reduce the dimensionality on a large scale, but also the prediction error MSE and MAPE are lower than those of the comparison model, which significantly improves the prediction accuracy. Finally, this model is applied to the representative stocks of different industries in China and achieves good forecasting effect, which proves the appli-cation value of this model in intelligent feature extraction and forecasting of financial time series again.