Thanks for the great feedback. I have no expectations for this other than the learning, and it's already been successful on that front. Just seemed like a fun thing to poke at when most other hobbyists seem to be doing image analysis and language modeling. I've crawled a couple of forums and I get that there are a lot of people out there who think they can readily use these techniques to make money. I doubt very much that this will be the outcome in my case :).
Where I am now I am just trying to figure out how to treat the data, whether to normalize or stationarize and how to encode inputs, etc. The reason that I am working with daily prices is that the fantasy output of this would be a model that can inform a one day grid trading strategy. It may very well be that daily prices won't work for this.
Whether there's anything like an equilibrium in cryptoasset markets where there are no underlying fundamentals is debatable. While there's no book price, PoW coin prices might be rationally describable in terms of (average_estimated cost of energy + cost per GH/s + 'speculative value')
A proxy for energy costs, chip costs, and speculative information
Are there standard symbols for this?
Can cryptoasset market returns be predicted with quantum harmonic oscillators as well?
What NN topology can learn a quantum harmonic model?
https://news.ycombinator.com/item?id=19214650
"The Carbon Footprint of Bitcoin" (2019) defines a number of symbols that could be standard in [crypto]economics texts. Figure 2 shows the "profitable efficiency" (which says nothing of investor confidence and speculative information and how we maybe overvalue teh security (in 2007-2009)). Figure 5 lists upper and lower estimates for the BTC network's electricity use.
https://www.cell.com/joule/fulltext/S2542-4351(19)30255-7
Here's a cautionary dialogue about correlative and causal models that may also be relevant to a cryptoasset price NN learning experiment:
https://news.ycombinator.com/item?id=20163734
Cool stuff, and I didn't mean to discourage you at all. Some of the most interesting challenges in datascience arise in finance.
Forex perhaps is just a pathologically tricky beast to trade well, even though it is the easiest to access. I think perhaps cryptos would be an easier start in terms of there being more inefficiencies and autocorrelation in the market.
In terms of data treatment, I recommend starting with Marco de Prado's Advances in Financial ML. I don't agree with some of his methods, but it is a practical book that highlights a lot of the issues you'll face. You can then draw your own conclusion how to treat them.
Where I am now I am just trying to figure out how to treat the data, whether to normalize or stationarize and how to encode inputs, etc. The reason that I am working with daily prices is that the fantasy output of this would be a model that can inform a one day grid trading strategy. It may very well be that daily prices won't work for this.