yklin / Metropolis-Hasting-WGAN-on-stock-prediction

Use metropolis hasting to enhance gan on stock prediction

This repository try to use Metropolis-Hastings Algorithm and WGAN to do the stock prediction.

💨 Quick start

1. Clone the repositroy and enter the folder

$ git clone [e-mail]:ChickenBenny/Metropolis-Hasting-WGAN-on-stock-prediction.git
$ cd Metropolis-Hasting-WGAN-on-stock-prediction

2. Set up the virtual environment

$ python -m venv venv

• Windows

$ venv\Scripts\activate

• Mac / Linux

$ source venv/bin/activate

3. Install package

$ pip install -r requirements.txt

4. Run the demo notebook. However, GAN is an unstable model, so you may need to run it for a longer period of time or adjust the hyperparameters to obtain better results.

🔽 Reference

1. Metropolis-Hastings GAN : https://arxiv.org/abs/1811.11357
2. Wasserstein GAN : https://arxiv.org/abs/1701.07875
3. Repo from borisbanushev : https://github.com/borisbanushev/stockpredictionai
4. Stock price prediction using Generative Adversarial Networks : https://thescipub.com/abstract/jcssp.2021.188.196

🌟 Idea

Incorporate the ideas from Boris Banushev's repository, as it focuses on implementing GAN-based models for stock prediction. These models leverage the power of GANs' ability to generate realistic and diverse samples, making them suitable for handling stock market dynamics, especially during high-volatility situations.

1. Try using GAN to predict stock prices and simulate the stock distribution.
2. Try enhancing the sampling process using the Metropolis-Hastings algorithm to achieve better convergence and explore a broader range of parameter values.
3. Attempt to improve the prediction accuracy by employing a VAE to extract latent variables and enhance the prediction effect.

🖥️ Result

• RMSE and MAE from testing dataset

  RMSE	MAE
  2.077	1.673

👀 Pros and Cons

• Pros
  1. Improve the stability of WGAN model.
  2. Improve the adaptability of the model.
  3. Might capture comple or highly volatile distributions.
• Cons
  1. Computational complexity and sesitivity to hyperparameters.
  2. Data limitations and highly dependency on data quailty.

🙈 Notice

The model exhibits instability and high volatility. Therefore, if you are interested in making stock predictions using GAN-based models, it is recommended to fine-tune the hyperparameters, explore various data preprocessing techniques, or consider alternative models.

License

MIT © ChickenBenny