Stock Movement Prediction Used Bi-typed Hybrid-relational Market Knowledge Graph via Dual Attention Networks

📌 Introduction

This project focuses on predicting stock movements by leveraging a bi-typed hybrid-relational market knowledge graph and Dual Attention Networks (DAN). The approach integrates graph-based modeling and deep learning techniques to improve accuracy in financial market predictions.

🎯 Objectives

• Construct a market knowledge graph incorporating stock and executive relationships.
• Used Momentum spillover effect for better model.
• Utilize Dual Attention Networks (DAN) for feature extraction and relation learning.
• Enhance prediction accuracy using graph-based deep learning methods.
• Implement and evaluate the model in PyTorch Geometric.

🛠️ Technologies Used

• Programming Language: Python
• Deep Learning Framework: PyTorch Geometric
• Graph Processing: NetworkX
• Machine Learning Libraries: NumPy, Scikit-learn, Matplotlib
• Data Handling: Pickle, Pandas

🚀 Installation & Setup

1️⃣ Install Dependencies

pip install torch torchvision torchaudio torch-geometric numpy scikit-learn matplotlib networkx dill

OR

pip install -r requirements.txt

2️⃣ Download the Dataset

Ensure the data folder contains processed stock and executive relations.

The two datasets for SMP with their folder names are given below.

CSI100E CSI300Epip install -r requirements.txt

3️⃣ Run the Training Script

python main1.py

🔍 Execution Workflow

📌 Data Processing

• Load stock-executive interaction data (interactive.pkl).
• Construct market knowledge graphs (explicit & implicit relations).
• Generate graph-based features for model training.

📌 Model Training

• Utilize Dual Attention Networks (DAN) for feature extraction.
• Train the model on hybrid-relational graph data.
• Optimize using Adam optimizer and cross-entropy loss.

📌 Model Evaluation

• Evaluate AUC (Area Under Curve) and Accuracy on test data.
• Example output from training:

  epoch32, train_loss=0.6495, eval_auc=0.5456, eval_acc=0.5189, test_auc=0.5516, test_acc=0.5274
  epoch50, train_loss=0.6014, eval_auc=0.5468, eval_acc=0.5356, test_auc=0.5325, test_acc=0.5233
  

📊 Results Visualization

• The model constructs market knowledge graphs showing stocks & executive relations.
• Performance metrics (AUC, accuracy) are plotted to assess training progress.
• Example knowledge graph visualization:

  import networkx as nx
  import matplotlib.pyplot as plt
  G = nx.Graph()
  # Add stock & executive nodes, then visualize the graph
  nx.draw(G, with_labels=True, node_size=600, font_size=8)
  plt.show()

🔥 Challenges & Solutions

Challenge	Solution
Large-scale graph data	Efficient data preprocessing & sparse tensor storage
Model overfitting	Regularization (Dropout, BatchNorm)
API Latency in real-time	Optimized graph sampling techniques

📈 Future Improvements

• Implement Transformer-based graph learning for better feature extraction.
• Deploy as a real-time prediction API.
• Enhance knowledge graph with macro-economic indicators.

📖 References

1. PyTorch Geometric Docs - https://pytorch-geometric.readthedocs.io/
2. NetworkX Graph Visualization - https://networkx.org/
3. Dual Attention Networks Research - https://arxiv.org/

📌 GitHub Repository : https://github.com/Ayussh-Raj/Stock-Movement-Prediction/tree/main