use GAN based augmentation

[danielhomola]

If the somewhat cleverer data augmentation #58 doesn't work that well, use Time-series gan , with this repo on github, which should give us infinite synthetic data), or this on bitbucket

An review of GANs (algos, theory & applications) from Jan 2020 can be found in this article.

[marcell-mate]

Evaluation

Evaluation of how good a generator really is is an area of current research, and a tough question given many popular metrics suffer from systemic weaknesses (eg Inception score or FID even for image generation)

• Here's a compelling article on the concept Improved precision-recall in GANs with the code on github here.

[danielhomola]

I guess for us with time series data as long as some key stats of the fourier series remain the same we're ok? I mean, we are not trying to generate class-based data (a time series for label 0 and another for label 1). We just want to generate any realistic looking OHLC data and then calculate our usual stats features on it and label it ourselves. So I guess for that we can measure the real data's key stats (maybe even the book mentions a few of these) and check if these are preserved in the generated data? Maybe I'm missing something.. let's discuss tomorrow!

[marcell-mate]

Not sure about the first one , I would need to do some more homework on that , probably you're right :D
Precision-recall here has nothing to do with classes, just an analogy where, briefly, precision is supposed to measure how much of what the generator does truly resemble the "real" feature space (quality), whereas recall shows how much of the "real" feature space can the model actually model (diversity). Seems like an approach which can be applied more generally than others, that's one of the reasons it caught my attention.
we can discuss tomorrow ! 👍

[danielhomola]

Ah got, thanks. So then it seems quite useful and general indeed!