MCI-GRU: Stock Prediction Model Based on Multi-Head Cross-Attention and Improved GRU (Neurocomputing).

📖 Overview

MCI-GRU is a stock forecasting architecture that replaces the GRU reset gate with an attention mechanism for dynamic temporal feature selection, incorporates a Graph Attention Network (GAT) to model inter-stock dependencies, and employs multi-head cross-attention to infer latent market states. By fusing temporal, cross-sectional, and latent representations, MCI-GRU achieves state-of-the-art performance on CSI 300/500, S&P 500, and NASDAQ 100 benchmarks, delivering higher ARR, Sharpe, and Calmar ratios, and has been deployed in live fund management systems.

🏆 Experiment Results

Performance Comparison

🚀 Quick Start

Prerequisites

• Python 3.8 or higher
• CUDA-compatible GPU (recommended for training)
• 8GB+ RAM

Installation

• Clone the repository

git clone https://github.com/WinstonLiyt/MCI-GRU.git
   cd MCI-GRU

• Create and activate virtual environment

# Using conda
   conda create -n mcigru python=3.8
   conda activate mcigru
   
   # Or using venv
   python -m venv mcigru_env
   source mcigru_env/bin/activate  # Linux/Mac
   # or
   mcigru_env\Scripts\activate     # Windows

• Install dependencies

pip install -r requirements.txt

Data Preparation

• Prepare your dataset in CSV format with the following columns:

• Update the data path in the script:

filename = '/path/to/your/dataset.csv'

Usage

Training and Prediction

Run the main training script for different market indices:

# CSI 300 Index
python code/csi300.py

Ablation Studies

Run ablation studies to analyze component contributions:

# Ablation studies for CSI 300 Index
python ablation/csi300.py

📁 Project Structure

MCI-GRU/
├── code/                   # Main implementation
│   ├── csi300.py           # CSI 300 index model
│   ├── csi500.py           # CSI 500 index model
│   ├── sp500.py            # S&P 500 index model
│   └── nasdaq100.py        # NASDAQ 100 index model
├── ablation/               # Ablation study implementations
│   ├── csi300.py           # CSI 300 ablation studies
│   ├── csi500.py           # CSI 500 ablation studies
│   ├── sp500.py            # S&P 500 ablation studies
│   └── nasdaq100.py        # NASDAQ 100 ablation studies
├── requirements.txt        # Python dependencies
├── README.md               # This file
└── LICENSE                 # MIT License

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 Citation

If you find this repository helpful, please cite our paper:

@article{zhu2025mci,
  title={MCI-GRU: Stock prediction model based on multi-head cross-attention and improved GRU},
  author={Zhu, Peng and Li, Yuante and Hu, Yifan and Xiang, Sheng and Liu, Qinyuan and Cheng, Dawei and Liang, Yuqi},
  journal={Neurocomputing},
  volume={638},
  pages={130168},
  year={2025},
  publisher={Elsevier}
}

Disclaimer: This code is for research purposes only. Past performance does not guarantee future results. Please conduct your own research before making investment decisions.