Here is the pre-released code for the FTGCN-based quantile and mean models in our paper "Big portfolio selection by graph-based conditional moments method", [paper].
Main settings: Python 3.9 & Pytorch 1.11.0
Minor settings: To complete.
The price and volume Data of each stock, sector-industry relation data, and wiki relation data, could be downloaded from the official repositiy of Feng (2019); see [stock data].
In the meanwhile, the daily Fama French five factors could be downloaded from the homepage of Kenneth R. French; see [factor data].
| Script | Usage |
|---|---|
| compute_factor_loading.py | To calculate factor loadings from raw End-of-day data and factor data |
| construct_feature.py | Generate the network input (including lagged values) for each day |
| construct_label.py | Generate the label for each day |
| Script | Usage |
|---|---|
| model.py | The model specification of network |
| my_dataset.py | The dataset specification based on Pytorch |
| load_data.py | Load the relation data |
| (F)TGCN.py | The agent used for training (F)TGCN |
| train_(F)TGCN.py | Train a model of (F)TGCN-based quantile (mean) model |
| hypothesis_test.py | The Kupiec and Christofer tests |
| QCM.py | The QCM learning from conditional quantiles |
| inference_(F)TGCN.py | Obtain four moments from the trained models |
# Please make sure you have changed the log directory in each file.
# Construct features and labels
python compute_factor_loading.py
python construct_feature.py
python construct_label.py
# Train models
# mean model
python train_FTGCN.py --tau 0.0 --mse-loss --lam 0.1 --save_folder ...
# quantile models
python train_FTGCN.py --tau 0.005 --lam 0.1 --save_folder ...
python train_FTGCN.py --tau 0.01 --lam 0.1 --save_folder ...
...
python train_FTGCN.py --tau 0.99 --lam 0.1 --save_folder ...
python train_FTGCN.py --tau 0.995 --lam 0.1 --save_folder ...
# Inference and QCM learning
python inference_FTGCN.py
If you feel this code helps, please kindly cite the following 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}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent