AI user again. Here a prediction from 96 hours ago and the real price

AI user again. Here a prediction from 96 hours ago and the real price.

Blue - previous price
Green dashed - prediciton
Black - real price

AI was slightly too pessimistic but in general it is fine.

+Grid
+Labes

for the fags who are lost otherwise.

Other urls found in this thread:

fooledbyrandomness.com/knowledge.pdf
twitter.com/NSFWRedditImage

can you give us the next future prediction from AI? thanks

So what is the current 96 hour prediction?

Current prediction.

Disclaimer:
DO NOT use this for trading! The AI is far from perfect and makes wrong predictions in more than 10% of all cases.

Again DO NOT use this for trading! Be warned.

why even bother to post this if you aren't going to give us a new one

also, you cant really prove it has worked if you didnt post that chart before, you could just have made it all up and claimed your dumb AI did it, its easy to draw a few lines a couple of days back and claim your TA knowledge predicted this

but he did post the projection before it happened, that's why I'm asking for the next one

nevermind you actually posted a new one

i just see it too often that ta faggots come in and claim their lines predicted this and then disappear

how much data do you feed the AI for the predictions?
do you only feed it short-term or long-term?

>continue the trend until zero or infinity
>AI

bwahaha

it's just a fucking linear extrapolation
>hurr AI

I only train it for mid-term predicitons: >1h and less than 1 week

Long-term is impossible imo.
Short-term cannot be used effectively by me due to technical limitations (i once tried it and the predictions are way better for this)

For each future datapoint prediction I feed it 960 values of different indicators.

For learning I fed it more than 10 million examples.

For the AI experts:
I split the data set in two parts:
learning (90%)
validation (10%)

Then I let it learn for many hours and always save the models with the best validation loss.

Cunts are fucked

Ad linear extrapolations:
It is not - here is why:

The market behaved almost linearly in the past few days. Ofc the AI picks up on this.

The AI also models dynamics like overshoots etc. which a linear extrapolation does not.

But ofc AI is not magic. It is just guessing - slightly better than human guessing.

Listen carefully you fuckwits
I warned you guys a week ago.. and spread fud several times. Sometimes for the lulz but most of the time to save a few smart enough to listen.

Next support is around 5.5k , it'll break through that just as it has all the others.
3.3ish thousand will be another strong support area.... i am HOPING that it will bounce off that and gain momentum or stabalize there. In reality it will most likely have a small bounce then smash through that support.

1.5k - 2k is where we will be before we recover... .if we are following the normie chart. Do not dismiss this advice, sell if you haven't already.... buy in cheaper. Don't be a delusional bag holder

make 2 sets of data one organic and one manipulated
make it account for market manipulation

Can you elaborate how I incorporate market manipulation without tweaking tons of data by hand? Because I have no clue how I should do that. For that I would have to have a good model of market manipulation in the first place.

i`m scared

this this this and again this.

This analysis actually makes sense. Due to crashing to under 7k people will try to buy in thinking it's over (I *almost* did that last night), deadcat-ing the price. Others who bought in at 8k+ will try to cut losses, and that's a majority of people at the moment. This will bring the price all the way down.

I'd say your AI is correct in this case. Did you develop it yourself?

you will make it friend

I did using python and tensorflow. Only the future will tell if they are correct...

Let me just emphasize again: Please do not base your decisions on these predictions as they could be utterly wrong as well.

I'm laughing at all you retards who think you can predict the market.

According to OP wait until bitcoin hits 0 before you buy back in

crypto price is heavily manipulated
ai predicted the jump then a pump because theres always a pump that makes the price go up when its going down i.e price jump of the whole market when we hit 7.9k a couple days ago

feed separate learning data from stocks and crypto

I am not saying you should buy sell or whatever. Just to make that clear once again. I for my self am hodling.

Ofc you cannot predict the market. But you can make educated guesses. The AI models the dynamics of the market and tries to estimate a possible outcome.

Anyone who thinks that AIs can predict perfectly like magic is a total retard.

Any prediction on what comes after?
I Tethered/went into Fiat at ~9k
Think it’s a good idea to buy in when we hit the bottom or is crypto dead?

When did you start doing this? Would be interesting to see some examples from the top of the btc rally and mid january.

this AI is much worse than TA

at least TA can spot a trend reversal, I doubt this AI could

This i cannot tell you, and no one can for certain. It could have great buys all over again after floor, or sit there for months.... to a year before moving up again.

Me? I'm gonna save some coins and buy in where ever i see the floor happening. Crypto should not be dead, it just might be for a while though.

I started collecting data about half a year ago. I tried many approaches. None is anywhere near perfect. I can post a few examples.

This for ex. ends at the 1st of January at 0:00 UTC

This one ends on 16th of Jan. 23:00 UTC

Ad. trend reversal and TA:

Yes it will not be able to predict trend reversals too often.

TA: TA is bullshit - it only works for two reasons:
Many people believe it works and therefore they act accordingly --> The market behaves the same way.

If it does not work people just move around a few lines and say oh i forgot this or that, just so that it fits.

It seems very pessimistic desu. Why is that? As a compsci undergrad and a Veeky Forumsmen I find that fascinating.

AI is retarded. There's no wall at 8k with your imaginary majority. BTC are shooting to the moon for the last 2h. LMAOing at your life.

It's been oscilatting between 6-7k for the last 12h actually. No moons or hard drops at the moment.

I have a hard time explaining why the AI decides what to predict. It does not tell me either!

From my experiments I got the following impression:

It tries to extract some general trend (e.g. downwards) and then adds dynamic to it.

It is not always that pessimistic though. This example ends at 30th Dec. 2017 11:00 UTC

>uses the term AI at all

ai from walletinvestor predicting a nice bullrun

It is not a copypasta python script.

It is a deep neural network with two LSTM and one dense layer. (nodenumber is my secret ;))

kek

bruh as of right now the ai is correct

I work on related systems and the difficulty with using ML on financial problems as opposed to say, Go, is detailed fairly well in the below.

fooledbyrandomness.com/knowledge.pdf

Basically financial systems have non-stable, unbounded probability distributions which is why bots are fine when the regime is stable (mean-reverting or trending) but then you eventually get wrecked when the regime changes.

You mentioned you've only been keeping the data for half a year - I don't know what else you're using as indicators but historical price for btc is available up to minute granularity at kaggle/elsewhere.

is this an RNN? It’s most likely retarded but so is 90% of “technical analysis”. Might as well read tea leaves

What did your backtesting results look like?

Yes it is a kind of RNN.

I am keeping data like volume, bids, asks, and many more. I am collecting data from more than 250 coins simultaneously, which is fed to the AI.

>AI sees big downtrend
>AI predicts further big downtrend
Am i missing something here?

Ad. backtesting

I did not create an automated trading algo yet. This is just for prediction. I do not even plan to do automated trading - way too dangerous for my taste.

Otherwise I just compare the predictions to the real data. Mostly they look like shown above. Sometimes it is absolute stupid though.

Yes the AI sees a downtrend and also predicts one in many cases. But it monitors much more than just the price and therefore makes better guesses than I am able to.

But again this is not magic of any kind. That is why I posted the disclaimer.

i think i know the answer but... it seems like an interesting project, would you share the code?

>predicting the downtrend while the downtrend

what about the bitcoin deathspiral?
I thought once it dropped past 6k it was a guaranteed hard crash because it would no longer be profitable to mine?

You mean crypto market is non-ergodic? Statistical models fail, because system is partially controlled by external events and has internal state (different investment strategies and goals). Even with the same history outcome can change because of those variables. It doesn't mean it's impossible to predict but just price history would be not enough.

I predicted back in the first week of January that it would drop to either $8k, $3k or $750-1500.

I’ll probably buy some at $3k. This crash is basically an extended “dump” where all the whales who made billions during the 2017 run are trying to realize their gains and diversify their assets. Now if crypto was actually valuable and could just be traded for things like real estate, stocks, bonds, and commodities this wouldn’t be a big issue, but unfortunately that’s not happening. They need to first convert to fiat which means they need to sell billions worth of BTC without crashing the market in a week. It’s a bit of a race against time. Various whales are selling off at intervals and pumping the price in between and spreading nonsense about “strong support at X”.

Just look at any meme line thread over the last month. Someone is always saying “Every knows there is strong support here. The trend is reversing. Here are the magic red lines to prove it.” at all the bull traps we’ve hit on the way down.

The crash isn’t over until it becomes unprofitable for those who drove the price up 20x in 2017 to sell. In the $3000 range it’s probably worth it for certain investors to just hold and wait for the next bull run. But there is a chance that those big investors just think the whole thing is garbage and want out entirely. At that point it’ll be propped up by people who always cared about crypto and believed in it’s value and the floor will basically be where people feel good about their ability to buy whole Bitcoins at a price they can afford comfortably.

People just make weird emotional decisions like that. “Oh I don’t want only .21 BTC for $1200, but one whole BTC at $1200 sounds good!”

Difficulty will be adjusted down... It will still be profitable

thats fud
did you actually believe that
these million dollar mining companies aren't giving up after a week of blood in the streets

I do the same thing, well done user. LSTM networks and Backtrader mostly. Mine reckons we're going lower too, also says Stocks are FUCKED on opening. It's slightly pessimistic though - tends to overshoot, but on average still trades better than O do lol.

I recommend including some sentiment analysis. I have Veeky Forums hooked up and a few other things like leddit and Twitter. Helps a lot - basically whatever these faggots say, the bot does the opposite.

what is it say, ai??

Are you Monte-carlo nigger, or non-stochastic gentleman?

t. another AI-user

You just track the external events. Sentiment from social media and headlines. Transaction fees. Support % Resistance levels. Hell, just dump in the other market candles, eg SPY - surprising correlations show up. I found just having a column of which weekday it was helped the MSE a bit.

Even just on raw OHLCV my Bot has about a 52 to 55% accuracy on the next candle. 70% with the other data. I have several seperately trained random LSTM networks with distinct random seeds, and have that across a few different timeframes. So they compensate each other. I aggregate the estimates - done.

With careful position sizing, it's pretty much theft at this point lol.

Thanks for suggestions. I was wondering if anyone is finding transaction log on the chain useful? Like seeing what top wallets are doing? Or even comparing activity across several top blockchains?

>there are people replying to this conman

Actually this kind of stuff is EXACTLY what a neural network AI is good at doing. But it would require a lot of data to train it properly.

There will never better a better tool for predicting these kinds of patterns than a neural net though.

What is your validation accuracy? Build a net which got about 58% on buy vs sell decision. Planning on building a Q learning network next. Also didn't notice much difference between using LSTM and CNN, you tried both?

Trend is your friend

70% on test set? May I ask what candle size and time window you use? Ty user

Accuracy: I could only give you an accuracy comparable to yours if I made binary buy/sell decisions, which I don't.

I tried deep Q already - takes forever and does not give better results. In fact it turns out that learning nets to trade/predict stuff is much more efficient using supervised learning.

I have not tried CNN yet. I tried standard dense nets, which showed a slightly worse performance.

Also heard good things about decision tree classification and random forests, tried those?

You are predicting the next price point for a certain window. I tried this too, however just seemed like a very short MA. (Tried it with only close as a feature)

I tried SVMs and DNNs for binary buy/sell and partially for buy/hold/sell decisions. The hard part though is to prepare the data correctly. You need a good algo that classifies past positions in the right class - which sounds straightforward but isn't.

That is why I decided to switch over to price predictions. I btw. also predict other data like volume etc.

>Shorted at 7.1k over an hour ago
>Come back and the price has been moving sideways for the entire time
If I close my short with 2% gains I just know the price will tank without me.

Price went up as I was typing, now I'm at negative. I'll see myself out

user, you can indicate which post you are replying to by clicking on its number - top right of the post.

Don’t tell him it’s funnier this way.

So if i read it correctly your current program is not tied into an trading program,right?
Also why do you think the other kind of ANNs failed to produce better results assuming you used the same inputs for every one of them?

you should split it 50%/50% user. i'll write a model myself this week anyway