LSTM_Predict_Vol

Replicate the study 'Can LSTM outperform volatility-econometric models?' with SSE 50 etf English | 中文

---

Realized Volatility Prediction using HAR and LSTM Models

This project aims to reproduce the core ideas from the paper "Can LSTM outperform volatility-econometric models?" by applying them to the minute-level high-frequency data of the SSE 50 ETF. It provides a comparative analysis between a classic econometric model (HAR-RV) and a deep learning model (LSTM) for predicting daily realized volatility.

✨ Features

• Data Processing: Implements a complete workflow for calculating daily Realized Volatility from high-frequency, minute-level data.
• Model Implementation: Contains full implementations of both the classic HAR-RV model and an LSTM model based on TensorFlow/Keras.
• Interactive Visualization: Generates interactive charts using Plotly, allowing for detailed exploration and analysis of the prediction results.
• Modular Code: The code is structured cleanly, making it easy to understand and extend.

📂 File Structure

.
├── 50ETF_1min.csv          # Raw minute-level data
├── volatility_prediction.py # Main script
├── requirements.txt        # Project dependencies
├── interactive_prediction_comparison.html  # Generated interactive chart (combined)
├── subplots_prediction_comparison.html     # Generated interactive chart (subplots)
├── prediction_comparison.png               # Generated static chart
└── README.md               # This file

🚀 How to Run

1.  Clone the Repository     bash     git clone [YOUR_REPOSITORY_URL]     cd [YOUR_PROJECT_FOLDER]    

2.  Create and Activate a Virtual Environment     bash     # Create the environment     python3 -m venv venv     # Activate on macOS/Linux     source venv/bin/activate     # Activate on Windows     .\venv\Scripts\activate    

3.  Install Dependencies     bash     pip install -r requirements.txt    

4.  Run the Script     bash     python volatility_prediction.py         After execution, the script will automatically generate *.png and *.html chart files in the folder.

📊 Experimental Results

This project was backtested on the SSE 50 ETF data from 2005 to 2022. The performance metrics of the two models on the test set are as follows:

• HAR-RV Model: MSE = 0.000020, MAE = 0.002487
• LSTM Model: MSE = 0.000021, MAE = 0.002869

The results indicate that the classic HAR-RV model slightly outperformed the baseline LSTM model in this experiment. This aligns with the original paper's conclusion that deep learning models require careful fine-tuning to surpass strong benchmarks.

Prediction Results Comparison Chart:

📚 References

• Rodikov, G., & Antulov-Fantulin, N. (2022). Can LSTM outperform volatility-econometric models?. arXiv preprint arXiv:2202.11581.

---

基于HAR与LSTM模型的已实现波动率预测

本项目旨在复现论文《Can LSTM outperform volatility-econometric models?》中的核心思想，通过使用上证50ETF的分钟级高频数据，对比分析了经典的计量经济学模型（HAR-RV）与深度学习模型（LSTM）在预测日度已实现波动率方面的表现。

✨ 项目特点

• 数据处理：实现了从分钟级高频数据计算日度已实现波动率（Realized Volatility）的完整流程。
• 模型实现：包含了经典的HAR-RV模型和基于TensorFlow/Keras的LSTM模型的完整实现。
• 结果可视化：使用Plotly生成了可交互的预测结果对比图表，方便对模型细节进行缩放和分析。
• 模块化代码：代码结构清晰，易于理解和扩展。

📂 文件结构

.
├── 50ETF_1min.csv          # 原始分钟级数据
├── volatility_prediction.py # 主代码文件
├── requirements.txt        # 项目依赖库
├── interactive_prediction_comparison.html  # 生成的合并交互图
├── subplots_prediction_comparison.html     # 生成的堆叠交互图
├── prediction_comparison.png               # 生成的静态结果图
└── README.md               # 本说明文件

🚀 如何运行

1.  克隆代码库     bash     git clone [您的代码库URL]     cd [您的项目文件夹]    

2.  创建并激活虚拟环境     bash     # 创建     python3 -m venv venv     # 激活 (macOS/Linux)     source venv/bin/activate     # 激活 (Windows)     .\venv\Scripts\activate    

3.  安装依赖     bash     pip install -r requirements.txt    

4.  运行脚本     bash     python volatility_prediction.py         脚本运行后，会自动在文件夹内生成*.png和*.html的结果图表文件。

📊 实验结果

本项目在2005年至2022年的上证50ETF数据上进行了回测。在测试集上，两个模型的性能指标如下：

• HAR-RV 模型: MSE = 0.000020, MAE = 0.002487
• LSTM 模型: MSE = 0.000021, MAE = 0.002869

预测结果对比图:

从结果来看，经典的HAR-RV模型在本次实验中的表现略优于基准的LSTM模型，这与原论文中“深度学习模型需要精细调参才能超越强基准”的结论相符。

📚 参考文献

• Rodikov, G., & Antulov-Fantulin, N. (2022). Can LSTM outperform volatility-econometric models?. arXiv preprint arXiv:2202.11581.