I'm working on an neural network for trading...

I'm working on an neural network for trading. The main reason I'm doing this is because I want to get better at AIs in general, so I have no idea about trading and finances.
I made an BTC/USD prediction bot and the results are mixed. Sometimes is very accurate, sometimes very wrong.
Right now I'm only analyzing the BTC/USD price from a single market. Would it be a good idea to analyze USD/EUR, BTC/ETH, etc. and find patterns in the correlation of multiple markets? Or would I be better off improving the current long sequence prediction in a single market basend on common patterns?

Attached: asdf.jpg (640x480, 47K)

Other urls found in this thread:

journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0212320&type=printable
twitter.com/NSFWRedditImage

Seems like an obvious yes.

yes.

Alright, seems like the market correlation idea isn't that bad.
I'm not sure that I can do both, my GPU RAM is already maxed out with the current model. But I'll try.
What markets would be the best to try at first? It's pretty important for the training process to have input data that forms a pattern, otherwise I'll only train it to predict shit.

Please respond. Should I take regular currencies like USD/EUR, another cryptomarket like BTC/ETH or some company stock? I guess there might be some correlation with all of them but I would like to have the best option to evaluate the initial model.

''neural network trading''
You are finding a Pearson Index of correlation
What do you want? The correlation between BTC and usd?

what's in for us user?

The most dominant currencies my man, highest consistant volume on the most reliable exchanges.

> Sometimes is very accurate, sometimes very wrong.

could u explain? in your pic i can see only prediction since 8:37

No technology on this board. Only marketing. Gtfo

>Pearson Index
I didn't know about that, thanks for the information! This will certainly prove useful.

What I want is the "best" correlation. The correlation I can get the best data from or the market with the most stable correlation to BTC/USD. My uniformed guess is that USD/EUR correlates more with BTC/USD than the stock of Volkswagen.
>most dominant currencies my man, highest consistant volume on the most reliable exchanges.
What would that be?

Should it prove to be somewhat reliable, I will invite a lot of beta testers.

This is one of the situations where the prediction is accurate. The blue line is 45 minutes in the past and predicted a rise. And a rise did happen 45 minutes later. But as I said, there are a lot of other situations where the prediction didn't come true.

>i copied Siraj's youtube tutorial on machine learning, please spoonfeed me
fuck off

I watched his videos but my model is very different from his. I started off with copying the model in this paper and extended it with sequence-prediction and several visual chart auto-encoders, over multiple timeframes and averaged out.
journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0212320&type=printable

Let's be real - you want an ai to trade for you so you have minimal agency in your trading to avoid roping. Pure cope.

Well, yeah. Why do you make it sound like a bad thing?

Lol finding the 'best' correlation is the hard part. Foreign exchange rates as you said might be good. S&P500? Gold price?

What type of neural network are you using? How did you decide your number of layers/neurons? I feel like there are better ways to apply neural nets for bitcoin price prediction. Something I've been meaning to try but haven't fully fleshed out/too lazy to do:

Use data of the order book and trade history for btc prediction. This could actually be applied to most high volume cryptos/stocks. So data would be:
Bid and ask orders (number and price point). The could limit this to +/- 10% for most relevant orders.
Trade history
I think most importantly: order deletions. I don't think any exchange APIs offer this data though. So a workaround would have to be crafted. Maybe something like 'order deletions = current order size - trade history at price point'.
Maybe also 'orders exceeding 5btc in size'. Value here can be tweaked ofc

There was a very good paper I came across which showed some promising results using a similar approach. The order book was treated as a heat map. The heat map was fed into a convolutional layer (typically used in image classification). Forgot what the rest was. Will post link once I get on my PC in a few hours

It is a good approach.
>LSTM + CNN
the best tools for your problem
Your model will suffer a lot from barts: massive unexpected dumps and pumps.
Your model is basically detecting other bots trading against other bots.

6 years ago, some Asian PhDs did the same, using state-of-the art ML tools back then.
They mentioned they had problems with barts as well.

How much data have you fed your model?
Is your data bullish or bearish?

>Bid and ask orders (number and price point). The could limit this to +/- 10% for most relevant orders.

Your model is basically detecting other bots trading against other bots.

this

Thanks, this is excellent advice! I haven't thought about including trade history in the model.
Maybe I can make some sort of web parser for trade and deletion history.
I'm very interested in the paper.

I'm already using a combination of several CNNs and LSTMs. I found good hyper parameters after playing with the setup for a few weeks now. Sadly I'm limited by my hardware.

Could this be compensated by including my own prediction in future predictions over several iterations? I figure other bots use the same strategy or find the same patterns, and I don't know how much they affect the market overall.
For training I used the last year of BTC/USD history. Making separate models for bullish/bearish market is on my list.

>chart auto-encoders what do you need that for?

>>LSTM
why also these?

Decreasing the input size of the LSTM while also classifying common patterns.