hierarchical-sigcwgan's Introduction

Hierarchical GAN for large dimensional financial market data Implementation

This repository is an implementation of the [Hierarchical (Sig-Wasserstein) GAN] algorithm for large dimensional Time Series Generation.

Link to the paper: https://doi.org/10.3905/jfds.2022.1.109

This codebase is based on work from COST Action 19130, supported by COST (European Cooperation in Science and Technology; www.cost.eu). We also thank an anonymous reviewer for useful comments.

Installation

There are two possible installation setups:

YML Installation (Recommended). Run the following command to automatically setup the hiersigcwganenv environment.
```
conda env create -f environment.yml
```
Manual Installation. Setup a new conda environment with python==3.8.3 and then run
```
pip install -r requirements.txt
```

Training on new datasets

In order to train on a new dataset the following changes to the code are needed:

Create a folder src/data, add your data file with the name of your dataset and add a data pipeline to data.py with the name get_{}_dataset with {} being the name of your dataset.
In hyperparameters_hierarchicalgan.py add an entry with the name of your dataset and the desired parameters to the following dictionaries:
1. Clustering_Hierarchical_GAN_CONFIGS
2. Base_SIGCWGAN_CONFIGS
3. CrossDim_SIGCWGAN_CONFIGS

All training results will be saved to generated_data/{}/seed=i with {} being the name of your dataset and as many seeds as specified.

Generating Scenarios

Once a model has been trained the parser in evaluate_hierarchical_gan.py can be modified to load the trained model and generate scenarios via the function generate_series_hierarchical_gan.

Replicating results from the paper

The data used in the first experiment are the closing prices from $03/05/2000$ to $07/05/2021$ (dd/mm/yyyy) of the continously rolled futures from Tables 2 and 4 in the paper. Due to copyright sharing issues we are not able to share the datasets but we do include the trained models in src/generated_data.