GRACE

This repository contains the pre-release code for the (F)TGCN-based quantile and mean models as presented in our paper, "Big Portfolio Selection by Graph-Based Conditional Moments Method." You can access the paper on Arxiv.

Environment

Main Settings: Python 3.9 & Pytorch 1.11.0 & CUDA 10.2
Minor Settings: To be completed.

Data

Stock Data: The price and volume data for each stock, sector-industry relation data, and wiki relation data can be downloaded from the official repository of Feng (2019); see the stock data repository.
Factor Data: Daily Fama-French five factors can be downloaded from the homepage of Kenneth R. French; see the factor data download link.

Data Pipeline

Script	Description
`compute_factor_loading.py`	Calculates factor loadings from raw End-of-Day data and factor data.
`construct_feature.py`	Generates the network input (including lagged values) for each trading day.
`construct_label.py`	Generates the label for each trading day.

Network

Script	Description
`model.py`	Specifies the model architecture of the network.
`my_dataset.py`	Defines the dataset structure based on PyTorch.
`load_data.py`	Loads the relation data.
`(F)TGCN.py`	Implements the agent used for training the (F)TGCN.
`train_(F)TGCN.py`	Trains the (F)TGCN-based quantile (mean) model.
`hypothesis_test.py`	Performs the Kupiec and Christofer tests.
`QCM.py`	Implements QCM learning from conditional quantiles.
`inference_(F)TGCN.py`	Obtains four moments from the trained models.

Reproduce the Results for NASDAQ-Wikidata with FTGCN Models

# Ensure you have updated the data path and log directory in each file.

# Step 1: Construct features and labels
python data_pipe/compute_factor_loading.py
python data_pipe/construct_feature.py
python data_pipe/construct_label.py

# Step 2: Train models
# Mean model 
python network/train_FTGCN.py --tau 0.0 --mse-loss --lam 0.1 
# Quantile models
python network/train_FTGCN.py --tau 0.005 --lam 0.1 
python network/train_FTGCN.py --tau 0.01 --lam 0.1 
...
python network/train_FTGCN.py --tau 0.99 --lam 0.1 
python network/train_FTGCN.py --tau 0.995 --lam 0.1

# Step 3: Inference and QCM learning
python network/inference_FTGCN.py

Cite

If you find this code helpful, please consider citing our paper:

@article{zhu2023big,
  title={Big portfolio selection by graph-based conditional moments method},
  author={Zhu, Zhoufan and Zhang, Ningning and Zhu, Ke},
  journal={arXiv preprint arXiv:2301.11697},
  year={2023}
}

Contact

Please feel free to raise an issue in this GitHub repository or email me if you have any questions or encounter any issues.

ZhuZhouFan / GRACE