This repository is the official implementation of [Conditional Sig-Wasserstein GANs for Time Series Generation]

Authors:

Paper Link:

To setup the conda enviroment:

conda env create -f requirements.yml

This repository contains implementations of synthetic and empirical datasets.

• Synthetic:
  • Vector autoregressive (VAR) data
  • Autoregressive conditionally heteroscedastic (ARCH)
• Real-world data:
  • Stock data: https://realized.oxford-man.ox.ac.uk/data

We compare our SigCGAN with several baselines including: TimeGAN, RCGAN, GMMN(GAN with MMD). The baselines functions are in sig_lib/baselines.py

To reproduce the numerical results in the paper, save weights and produce a training summaries, run the following line:

python train.py -use_cuda -total_steps 1000

Optionally drop the flag -use_cuda to run the experiments on CPU.

To evaluate models on different metrics and GPU, run:

python evaluate.py -use_cuda

As above, optionally drop the flag -use_cuda to run the evaluation on CPU.

The numerical results will be saved in the 'numerical_results' folder during training process. Running evaluate.py will produce the 'summary.csv' files.