| 面向股票预测的混合特征筛选与图卷积增强的多层感知机模型构建 |
| 点此下载全文 |
| 引用本文：潘豪,高美娜,刘晓梅.面向股票预测的混合特征筛选与图卷积增强的多层感知机模型构建[J].上海第二工业大学(中文版),2026,43(2):218-226 |
| 摘要点击次数: 213 |
| 全文下载次数: 10 |
|
|
| 中文摘要:上市公司的股价波动不仅取决于自身基本面, 还会受到其他相关股票价格变动的交叉影响。为些, 本研究提出了一种基于混合特征筛选(hybrid fature selection, HFS) 与图卷积网络(graph convolutional network, GCN) 的多层感知机(multi-layer perceptron, MLP) 模型, 即HFS-GCN-MLP。首先, 采用混合特征筛选策略(ElasticNet 捕捉线性信号, XGBoost 捕捉非线性信号), 从40 个因子数据特征和12 个基本面数据特征中分别筛选并取其并集; 其次, 基于6 个市场数据特征构建关联图, 通过GCN 提取空间依赖嵌入; 最后, 将嵌入与筛选后的因子拼接后输入MLP, 对2025 年第一季度的日频数据进行涨跌二分类预测。在两个独立样本集(各包含13 只A 股股票) 上的实验结果显示,HFS-GCN-MLP 相比基准模型, 预测准确率分别显著提升了2.66 个百分点和2.94 个百分点。此外, 消融实验进一步验证了并集特征筛选机制与GCN 模块对整体性能的关键贡献。研究结果表明, 将线性与非线性特征筛选与图卷积增强深度学习相结合, 可有效提高股票涨跌预测的准确率。 |
| 中文关键词:股票走势预测 混合特征筛选与图卷积网络的多层感知机模型(HFS-GCN-MLP) 图卷积网络 特征选择 |
| |
| Hybrid Feature Selection and Graph Convolutional Enhanced Multi-Layer Perceptron Model for Stock Prediction |
|
|
| Abstract:The stock price fluctuation of a listed company depends not only on its own fundamentals but is also subject to the crossinfluence of price changes in other related stocks. This study proposes a multi-layer perceptron (MLP) model based on hybrid feature selection (HFS) and graph convolutional network (GCN), termed HFS-GCN-MLP. First, a hybrid feature selection strategy is implemented: Elastic Net captures linear relationships, while XGBoost identifies nonlinear patterns across 40 factor-based features and 12 fundamental features, with the selected variables combined through a union operation. Subsequently, a financial correlation graph is constructed using six market-related data features, then extract spatial dependency embeddings through GCN. The extracted embeddings are then concatenated with the selected features and fed into MLP to perform next-day binary (up/down) price movement classification using daily trading data from the first quarter of 2025. Tests on two independent sample sets (each containing 13 A-share) show the HFS-GCN-MLP model outperforms baselines by 2.66% and 2.94% accuracy gains. Ablation experiments further validate that the critical contributions of the union feature selection mechanism and the GCN module to overall performance. Results show that integrating linear and nonlinear feature selection with graph convolutional-enhanced deep learning can effectively improve the accuracy of stock price movement prediction. |
| keywords:stock movement prediction HFS-GCN-MLP graph convolutional network feature selection |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|