Can we use deep learning to analyse charts and transactions to predict cryptocurrency prices?

Can we use deep learning to analyse charts and transactions to predict cryptocurrency prices?

I'm pretty sure it works for stocks so it should work for high-volume cryptos too.
Totally worthless for 99% of cryptos though. Most of those are just pump & dumps.

I'm thinking about coding it and release the source code just to see the world burn with everyone trying to predict the future. Just Imagine.

Where to even start, my script that saves gdax historical data and current data is done, but now I'm at a loss.

I-i can recognize handwritten digits... how do price prediction?

Just feed it data and train it.

I guess the mental block I have is that the market is a continuous event, yet I have to decide to feed data in increments.
Since it's a neural net how much of the algorithm should I even try to create? Theoretically I really only need to pick my time increment.

Can I train it from now to backwards? (As opposed to starting with the earliest historical data I have)
How to make it both train as new market data comes in as well as learn based from it's own decisions?

Easy in theory until I imagine how to implement it...

Isn't that what Augur is supposed to do?

The basic ML you are talking about is easy. Knowing which questions to ask is the hard part.

no, the truth is so many people tries to predict the prices and has automated actions set up that happen so fast and in cascades that you can't simply guess what will happen. it has both automated and psychological components and the way news is presented can not be guessed nor could the billions of events shaping the markets.

it's in effect random.

>Can I train it from now to backwards?
neural nets don't give a fuck about the order you present your data to them statistically.

the net will of course be different every time you build it.

one thing tho if you train your network with deltas as inputs you need to switch them if you go backwards in time (always p1-p0) but this is just programming technicality.

Hmm I envision you have several nets doing different things, one could scan for news and guess if it's positive or negative.
Someone posted a thread about this recently I think, (((stock brokers))) were getting fleeced cause of some guys algorithm

Thx m8
>the net will of course be different every time you build it.
Ok yeah I forgot that it plateaus rather quickly. Is continuous training not a thing (ie. adding to the model as the market moves forward in time rather than manually rebuilding over and over)?

It does not work for stocks, unless you have insider info

You would need a supercomputer for this. Neuralnet learning time grows exponentialy. If you want to analyse transactions it would have to be real time analysis, not possible on any home-tier computer. Even if it was, you would lack ability/rights to place orders fast enouh. What you have in mind is something like high frequency tradin, but it is not available for average Joe's

The only (((analysis))) you can make with crypto, is trying to predict when the whales will manipulate the price.
AKA psychology

>You would need a supercomputer for this.

Bullshit. My desktop computer with GPU can train a substantially sized neural network in just a matter of minutes.

>Neuralnet learning time grows exponentialy.

And, again, bullshit. The most computationally complex aspect of training a neural network is the optimization process which effectively has a cubic time complexity.

>Bullshit. My desktop computer with GPU can train a substantially sized neural network in just a matter of minutes.
hahhahaahahaahhaaahhaaaa

>I don't know what I'm talking about but I'll still give my opinion

Does it work with currencies and index funds?

Maybe.
Depends on what the input factors are and how good the quality of data is.

If its just garbage in, its going to be garbage out.

not really m8. Some technical analysis might work, but only as "self-fullfilling" prophecy. since indexes are very well diversified you cannot pin point major factors that drive the index price, so the only option left for you is time-series analysis and it's pretty accepted that time-series analysis simply does not work for financial assets. The higher market capitalization the harder it is to create a prediction model. if you want to play around with machine learnign it would be better idea to collect variables about altcoins with low market cap, dependent variable could be price increase, where let's say higher than 20% increase is 1 and all else is 0. train your model with differently lagged variables and who knows, you might actually predict something with logit or some other method

wow Veeky Forums i'm genuinely impressed by the responses. you're not as retarded as a couple of years ago.

if it were to work properly you would need a supercomputer and enormous sets of data. like with the captchas on Veeky Forums, you're providing computer vision training data to google, because they need as much data as they can get.

i'd imagine that people are trying to this type of thing but they're playing with other people's money and they're making money for themselves through fees and salaries, not through their trading system.