search

birdx0810 / timegan-pytorch

This repository is a non-official implementation of TimeGAN (Yoon et al., NIPS2019) using PyTorch.
MIT License
80 stars 35 forks source link

Synthetic data generation #4

Closed kamasyali closed 2 years ago

kamasyali commented 2 years ago

When I implement this with the stock data, the generated synthetic data is not so good. For example, opening price of a day should be very close to the closing price of the previous day. However, for some days I do not observe this pattern. Did you actually generate synthetic data for stock dataset? If yes, can you also share your generated data?

ghost commented 2 years ago

...opening price of a day should be very close to the closing price of the previous day. However, for some days I do not observe this pattern.

I think it would be normal that "some days" would have the poor results that you mentioned for a few reasons.

  1. The hyperparameters were not optimized for this dataset.
  2. I don't think the evaluation metric (TSTR) would be sensitive to this issue. Therefore, there is a possibility that the model generated these "noisy" examples.

Did you actually generate synthetic data for stock dataset? If yes, can you also share your generated data?

Unfortunately, I did not save my results on the stock data, as I am using it for other purposes.

You could try to generate samples using the original implementation here. Hope this helps.