Bitcoin, EUR/USD, and AUD/USD short term predictions using machine learning

[K128kevin2]

I run a website (www.btcpredictions.com) which I posted about on here close to a year ago. I got helpful feedback from people, so I thought I would ask here again because I've just recently made some major changes.

This site uses artificial intelligence algorithms to try to predict future bitcoin prices over the next 24 hours and 5 days. The predictions are displayed on simple charts. The major changes that I've made in this newest version is that it now predicts aud/usd and eur/usd exchange rates in addition to bitcoin, and that it shows past performance (which won't start working until tomorrow). I'll be adding more currency pairs soon too.

Anyway I was hoping to see what people thought, and get some suggestions on what could improve. I know the ui is a bit crappy... I'm more of a back-end software kind of guy. There are some bugs in it too, but it seems to be functioning mostly for the moment.

Any constructive feedback would be much appreciated!

[talks_cheep]

Didn't you have a neural network or something like it? If I remember correctly, it didn't work very well. Have you made major strides since then?

[NorrisK]

I was wondering why you need people to create an account? To be honest, I don't really want to make accounts for every small service out there. I allready have way too many accounts.

If it is just for showing the price predictions, I think you could do without the sign up part.

[rebuilder]

The chart for past performance doesn't seem to load. Firefox on linux.

[WENGER]

The site doesn't look creepy but rather very simple interface and it seems it lacks functionality. I tried using IE,chrome and firefox but when I click charts, noting shows up, only next and past text. The website seems to only have some explanation as far as what I could see unless I have to sign up before some stuff show up. Aside from these the site loads fast.

[K128kevin2]

Didn't you have a neural network or something like it? If I remember correctly, it didn't work very well. Have you made major strides since then?

I'm not sure when you last saw the site, but I've made changes over time here and there. For a few months it was actually doing shockingly well, although the past week ish it's been not too great. I'm always working on improving it.

I was wondering why you need people to create an account? To be honest, I don't really want to make accounts for every small service out there. I allready have way too many accounts.

If it is just for showing the price predictions, I think you could do without the sign up part.

It won't send you any emails other than to sign up, and if you need to reset your password. There are a lot of reasons that it now requires sign up, one of which is so that I can get a better idea of site usage.

The chart for past performance doesn't seem to load. Firefox on linux.

The performance doesn't load because there is no past performance yet because the site has been up for less than 24 hours. Check again tomorrow night

The site doesn't look creepy but rather very simple interface and it seems it lacks functionality. I tried using IE,chrome and firefox but when I click charts, noting shows up, only next and past text. The website seems to only have some explanation as far as what I could see unless I have to sign up before some stuff show up. Aside from these the site loads fast.

Yes you need to sign up. It won't send you any emails other than to verify your account. The page to sign up is here: http://www.btcpredictions.com/signup

[n2004al]

I run a website (www.btcpredictions.com) which I posted about on here close to a year ago. I got helpful feedback from people, so I thought I would ask here again because I've just recently made some major changes.

This site uses artificial intelligence algorithms to try to predict future bitcoin prices over the next 24 hours and 5 days. The predictions are displayed on simple charts. The major changes that I've made in this newest version is that it now predicts aud/usd and eur/usd exchange rates in addition to bitcoin, and that it shows past performance (which won't start working until tomorrow). I'll be adding more currency pairs soon too.

Anyway I was hoping to see what people thought, and get some suggestions on what could improve. I know the ui is a bit crappy... I'm more of a back-end software kind of guy. There are some bugs in it too, but it seems to be functioning mostly for the moment.

Any constructive feedback would be much appreciated!

So you have created the site which predict the price (value) of bitcoin compared to one or other currencies. Good invention. But I have a question. You tell that you have used "artificial intelligence algorithms" to predict future prices of bitcoin. My first question is: since this kind of algorithms are created by the mind of an human (yours in this case) and, as I know, this kind of existence (named human brain) is not yet able to realize such kind of "job" that your site do, how is possible that the product be more intelligent that the creator? The human brain cannot give to your site that what is not able to have. At least this is what tell my brain. Maybe I'm wrong. If so can you correct me?

[rebuilder]

Could you draw the average error as upper and lower bounds to the predicted price? It might outline the nature of the predictions better.

[oda.krell]

Always interesting to see new methods (especially ML based) here. Thanks for the stimulus :)

A few remarks:

* While I'm guilty of 'black box' posting myself, it would of course be a lot more interesting to look a bit behind the veil of your algorithm. I gather it's an ANN-based predictor. Anything more you want to say about it?

What NN architecture specifically? And, maybe most importantly, a few words about the structure (and features) of your input, and what exactly you have been training here?

(EDIT: As in: Vector or scalar input? Price only, or other factors? Uniform time scale, or several? Any idea how to describe, qualitatively, what your input represents, and what the generalization is your network (ideally) learned?)

* The way it looks now, the network predicts a moderate rise over the next 5 days, to around $560/565. As it stands now, those predictions are a bit "isolated", so ideally, you would add the following:

(a) An additional output of your predictor that represent the certainty of your network wrt its prediction. Probably could be done as another error function you need to train (training aims to reduce the combined error then, of 'price deviation' and 'confidence deviation').

This suggestions is probably similar to rebuilder's request above, but possibly, to do what he asks, you'd first need to implement something like (a).

(b) Tracking how your method performed on the site. Even if (a) above is too much effort, it should be not too much effort to graph the history of prediction vs. actual price. Possibly, you were already planning to do so. There's probably more than one way to implement that one in detail, since I expect the output of your NN to get closer to the actual price as the prediction distance gets smaller.

Perhaps, even though it might be a bit of a hack, you could discretely color-code the prediction output vs. price for different distances, e.g. at time N, you plot the actual price of N, plus the price predicted by the NN at N-5 days (in red), plus the prediction of N-4 days (in orange), etc., up to maybe N-1 day.

(Apologies to the data scientists and statisticians :D ... as you can see, I have no damn clue how to properly plot error for a rolling window method)

[teonoaro]

Well,
First of all, Congrats for the site.
Second of all, I don't like the site. You need to put sometime to design the website professionally, if you have friend with good designing knowledge have him to help you. This is mandatory.
Third of all, As the site is up in a while, you would have asked our feedback after few days.
Then,

About page and Contack information are enough for this site.

You need to put some Cautionary information stating this is an machine learning prediction thing.

[americanpegasus]

Did your neural net predict that most people are not going to give you an e-mail address for a wonky price prediction? 
 
Because I'm not even going to give Wall Street Journal an e-mail to read articles; why would most people give you one? 
 
If you insist on this foolish scheme, show everyone the predictions from 48 hours ago and require an e-mail address to view up to the minute predictions.

[belmonty]

The site doesn't look creepy but rather very simple interface and it seems it lacks functionality. I tried using IE,chrome and firefox but when I click charts, noting shows up, only next and past text. The website seems to only have some explanation as far as what I could see unless I have to sign up before some stuff show up. Aside from these the site loads fast.

Yes you need to sign up. It won't send you any emails other than to verify your account. The page to sign up is here: http://www.btcpredictions.com/signup

What will you do with the email addresses you gather? Some sites sell them on to advertising companies. How do we know you are telling the truth if you say you will keep them private and only use them for sign up? Some companies make a living harvesting email addresses.

[oda.krell]

^, ^^

Just sign up with your spam mail address. I'm pretty sure that's what everyone does in cases like this, right?

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

[americanpegasus]

^, ^^

Just sign up with your spam mail address. I'm pretty sure that's what everyone does in cases like this, right?

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

 
 
It's like the homeless guy who comes into the local gas station and steals a cookie every day.  It's just a cookie right?  It's not really hurting anyone. 
 
But it begins the slow slide towards degeneracy.  Websites simply displaying information do not need an e-mail address. 
 
I *love* the idea of using neural networks to help predict asset prices.  This is awesome!  But I refuse to participate in nonsense.  How can I trust your models and methods if you are so clueless about visitor desires and behavior?  You are purporting to be able to predict human behavior with regards to price, but yet you show right out of the gate that you are clueless with regards to human behavior regarding how they want to view your information. 
 
This site needs to get rid of this, or provide a basic service and then an advanced service to registered members. 
 

[belmonty]

^, ^^

Just sign up with your spam mail address. I'm pretty sure that's what everyone does in cases like this, right?

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

 
 
It's like the homeless guy who comes into the local gas station and steals a cookie every day.  It's just a cookie right?  It's not really hurting anyone. 
 
But it begins the slow slide towards degeneracy.  Websites simply displaying information do not need an e-mail address. 
 
I *love* the idea of using neural networks to help predict asset prices.  This is awesome!  But I refuse to participate in nonsense.  How can I trust your models and methods if you are so clueless about visitor desires and behavior?  You are purporting to be able to predict human behavior with regards to price, but yet you show right out of the gate that you are clueless with regards to human behavior regarding how they want to view your information. 
 
This site needs to get rid of this, or provide a basic service and then an advanced service to registered members. 
 

There have been countless scams here and they often requested email addresses to participate. People get scammed, and no doubt had their email addresses sold. Most bitcointalk members are aware of this, and the OP should be. Asking for email addresses without good reason raises warning flags, and the OP shouldn't be doing it if he wants us to trust his site.

[oda.krell]

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

I already said everything I believe needs to be said here. /OT

[uki]

^, ^^

Just sign up with your spam mail address. I'm pretty sure that's what everyone does in cases like this, right?

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

 
  
It's like the homeless guy who comes into the local gas station and steals a cookie every day.  It's just a cookie right?  It's not really hurting anyone.  
  
But it begins the slow slide towards degeneracy.  Websites simply displaying information do not need an e-mail address.  
  
I *love* the idea of using neural networks to help predict asset prices.  This is awesome!  But I refuse to participate in nonsense.  How can I trust your models and methods if you are so clueless about visitor desires and behavior?  You are purporting to be able to predict human behavior with regards to price, but yet you show right out of the gate that you are clueless with regards to human behavior regarding how they want to view your information.  
  
This site needs to get rid of this, or provide a basic service and then an advanced service to registered members.  

I agree with that. I also wanted to have a look on the chart, and discuss it a bit, as this sounds as really an interesting approach. But, giving up my mail somehow puts me off. There is absolutely no need to ask for that. To the contrary, if I can see the charts and find them interesting I will be the first one to actually drop you a mail with questions. A pity, and sorry for yet another OT post.

[K128kevin2]

So you have created the site which predict the price (value) of bitcoin compared to one or other currencies. Good invention. But I have a question. You tell that you have used "artificial intelligence algorithms" to predict future prices of bitcoin. My first question is: since this kind of algorithms are created by the mind of an human (yours in this case) and, as I know, this kind of existence (named human brain) is not yet able to realize such kind of "job" that your site do, how is possible that the product be more intelligent that the creator? The human brain cannot give to your site that what is not able to have. At least this is what tell my brain. Maybe I'm wrong. If so can you correct me?

That's a fun an interesting point of discussion in my opinion. Actually, computers and AI often outperform humans at many tasks. The most obvious example (not AI) is a calculator. A calculator is created by humans and it can far outperform us at performing mathematical computations.

Neural networks can also outperform humans at some tasks. One surprising example is that they are very good at recognizing hand written digits. The first neural network I ever created performed this task, and it was better at recognizing the poorly written digits than I was!

The reason that neural networks are not limited by what humans can do is because we, as humans, don't directly program them to perform tasks. We program a neural network to be able to LEARN how to do a certain task, and they can sometimes eventually learn the task better than a human.

A computer can be more intelligent than its "creator" because you don't teach it what you know - you teach it how to learn.

[K128kevin2]

Could you draw the average error as upper and lower bounds to the predicted price? It might outline the nature of the predictions better.

Thanks for the suggestion, others have suggested this too in the past. I may do this eventually. It's tough though because the price will not always be within the average error range. I think people may misread it, but I may try it anyway.

[K128kevin2]

Always interesting to see new methods (especially ML based) here. Thanks for the stimulus

A few remarks:

* While I'm guilty of 'black box' posting myself, it would of course be a lot more interesting to look a bit behind the veil of your algorithm. I gather it's an ANN-based predictor. Anything more you want to say about it?

What NN architecture specifically? And, maybe most importantly, a few words about the structure (and features) of your input, and what exactly you have been training here?

(EDIT: As in: Vector or scalar input? Price only, or other factors? Uniform time scale, or several? Any idea how to describe, qualitatively, what your input represents, and what the generalization is your network (ideally) learned?)

* The way it looks now, the network predicts a moderate rise over the next 5 days, to around $560/565. As it stands now, those predictions are a bit "isolated", so ideally, you would add the following:

(a) An additional output of your predictor that represent the certainty of your network wrt its prediction. Probably could be done as another error function you need to train (training aims to reduce the combined error then, of 'price deviation' and 'confidence deviation').

This suggestions is probably similar to rebuilder's request above, but possibly, to do what he asks, you'd first need to implement something like (a).

(b) Tracking how your method performed on the site. Even if (a) above is too much effort, it should be not too much effort to graph the history of prediction vs. actual price. Possibly, you were already planning to do so. There's probably more than one way to implement that one in detail, since I expect the output of your NN to get closer to the actual price as the prediction distance gets smaller.

Perhaps, even though it might be a bit of a hack, you could discretely color-code the prediction output vs. price for different distances, e.g. at time N, you plot the actual price of N, plus the price predicted by the NN at N-5 days (in red), plus the prediction of N-4 days (in orange), etc., up to maybe N-1 day.

(Apologies to the data scientists and statisticians ... as you can see, I have no damn clue how to properly plot error for a rolling window method)

Thanks for the detailed feedback, I really appreciate you taking the time to write this stuff out!

So yeah I guess it is a bit of a black box post... sorry about that. I usually don't go into details in an initial post about something like this because the majority of people wouldn't really know what I'm talking about since you'd have to have a background in machine learning.

Anyway the predictions are created by feed forward neural networks with varying numbers of layers and sizes. Generally it's 4 or 5 layers, and sizes are mostly in the 100-200 node range, but they are all different. The inputs for bitcoin are change in both price and volume leading up to the point at which predictions are being made, and for currencies it is just price. Also I'm happy to go into more detail about stuff if you'd like to contact me directly at btcpredictions@gmail.com.

a. A confidence interval would be wonderful, and others have suggested that too. Unfortunately, the way the model works, that data isn't really available. I may experiment with more probabilistic models in the future because I do agree that a confidence interval would be extremely valuable and interesting to see.

b. This feature actually exists already. It hasn't shown up yet because there aren't 24 hours worth of predictions yet (at this point, the site went up about 21 hours ago. A little later tonight, you should start seeing those charts. I'm excited to see them actually because the euro/usd and aud/usd predictions have not been tested yet at all.

Thanks again for the feedback! Like I said, feel free to shoot me an email if you'd like to get into more detail about the actual neural network.

[K128kevin2]

Well,
First of all, Congrats for the site.
Second of all, I don't like the site. You need to put sometime to design the website professionally, if you have friend with good designing knowledge have him to help you. This is mandatory.
Third of all, As the site is up in a while, you would have asked our feedback after few days.
Then,

About page and Contack information are enough for this site.

You need to put some Cautionary information stating this is an machine learning prediction thing.

Thanks for the feedback. Yeah I know the site design is not professional. I am not much of a front end developer, I am much more of a back end guy. I am actually looking for someone to help me out with the front end. I just don't have the time, effort, or interest to get particularly good at that stuff. There is a TON of work involved in the creation and maintenance of this site, and I just put a lot more effort into the back end apis and prediction software. Hopefully I can find people to help and make it much more professional looking.

[K128kevin2]

^, ^^

Just sign up with your spam mail address. I'm pretty sure that's what everyone does in cases like this, right?

Really doesn't seem worth the effort to go into a fight about principles here - it's not like OP is asking for your CC info, or anything that's of actual importance.

 
  
It's like the homeless guy who comes into the local gas station and steals a cookie every day.  It's just a cookie right?  It's not really hurting anyone.  
  
But it begins the slow slide towards degeneracy.  Websites simply displaying information do not need an e-mail address.  
  
I *love* the idea of using neural networks to help predict asset prices.  This is awesome!  But I refuse to participate in nonsense.  How can I trust your models and methods if you are so clueless about visitor desires and behavior?  You are purporting to be able to predict human behavior with regards to price, but yet you show right out of the gate that you are clueless with regards to human behavior regarding how they want to view your information.  
  
This site needs to get rid of this, or provide a basic service and then an advanced service to registered members.  

I agree with that. I also wanted to have a look on the chart, and discuss it a bit, as this sounds as really an interesting approach. But, giving up my mail somehow puts me off. There is absolutely no need to ask for that. To the contrary, if I can see the charts and find them interesting I will be the first one to actually drop you a mail with questions. A pity, and sorry for yet another OT post.

I'm sorry, I didn't realize people would be so upset by this. First and foremost, I can promise you that I will not sell anybody's email addresses, and I will not send you any emails except for verifying your account (and if you need to reset your password you'll get an email then too). I understand people might be skeptical, but I don't know what I could possibly do or say to prove to you that I'm telling the truth. I already actually have a database with thousands of emails from a previous iteration of this site where I had a membership section. I didn't receive any complaints about it last time so I figured I wouldn't this time either.

Anyway, there are a lot of reasons that I have it set up this way. The primary reason is that it gives me a better idea of site usage. It enforces that the people hitting my site are real people and not robots. Furthermore, if I manage to get people helping me out on this project and develop it into something more robust and professional, I may start charging a small fee to cover costs (it's already costing me a lot out-of-pocket to run this site). This way I already have a infrastructure for that set up.

I figured that since the information I'm asking for is just an email address, which is about as far from personal information as you can get, that people wouldn't be bothered by that. I may add a chart to the main page though since a few people seem to dislike it.

[K128kevin2]

 
It's like the homeless guy who comes into the local gas station and steals a cookie every day.  It's just a cookie right?  It's not really hurting anyone. 
 
But it begins the slow slide towards degeneracy.  Websites simply displaying information do not need an e-mail address. 
 
I *love* the idea of using neural networks to help predict asset prices.  This is awesome!  But I refuse to participate in nonsense.  How can I trust your models and methods if you are so clueless about visitor desires and behavior?  You are purporting to be able to predict human behavior with regards to price, but yet you show right out of the gate that you are clueless with regards to human behavior regarding how they want to view your information. 
 
This site needs to get rid of this, or provide a basic service and then an advanced service to registered members. 

I almost forgot I wanted to reply to this specifically. I just wanted to mention that I am not claiming to be able to predict human behavior, but rather I am claiming to be able to ATTEMPT to predict currency exchange rate data, which is distinctly different from human behavior. It is extremely different from ux design, which is essentially what you are discussing.

[jasonjm]

I run a website (www.btcpredictions.com) which I posted about on here close to a year ago. I got helpful feedback from people, so I thought I would ask here again because I've just recently made some major changes.

This site uses artificial intelligence algorithms to try to predict future bitcoin prices over the next 24 hours and 5 days. The predictions are displayed on simple charts. The major changes that I've made in this newest version is that it now predicts aud/usd and eur/usd exchange rates in addition to bitcoin, and that it shows past performance (which won't start working until tomorrow). I'll be adding more currency pairs soon too.

Anyway I was hoping to see what people thought, and get some suggestions on what could improve. I know the ui is a bit crappy... I'm more of a back-end software kind of guy. There are some bugs in it too, but it seems to be functioning mostly for the moment.

Any constructive feedback would be much appreciated!

this is not a dig at you, but at AI per se......

I work all day with computers and they are pretty much retarded as far as I am concerned. The machine doesn't learn a thing, all it does is follow whatever rules you have created for it to follow.

[K128kevin2]

this is not a dig at you, but at AI per se......

I work all day with computers and they are pretty much retarded as far as I am concerned. The machine doesn't learn a thing, all it does is follow whatever rules you have created for it to follow.

I would recommend googling some studies on recurrent neural networks for image analysis and music composition. The things they can do are amazing. Computer scientists can train neural networks to compose brand new music, and it's often actually good! They can perform a lot of pattern recognition tasks a lot better than people can. It's very impressive.

[Morecoin Freeman]

It's very impressive.

It really is. The discussion in this thread actually inspired me to read a lot about artificial neural networks last night.
This is not my subject but I am amazed. The learning part is very interesting and something I did not know too much about.

Very interesting project you have here and I am looking forward to see how it performs.

[jasonjm]

this is not a dig at you, but at AI per se......

I work all day with computers and they are pretty much retarded as far as I am concerned. The machine doesn't learn a thing, all it does is follow whatever rules you have created for it to follow.

I would recommend googling some studies on recurrent neural networks for image analysis and music composition. The things they can do are amazing. Computer scientists can train neural networks to compose brand new music, and it's often actually good! They can perform a lot of pattern recognition tasks a lot better than people can. It's very impressive.

all its doing is creating rules based on previous patterns it spots (neural networks).

if the futures patterns don't match its worthless (which is what happens in markets).

I listened to the music a while back, and it seems impressive until you realize its just a rehash of other peoples previously original work mixed up with some modifications, like a rapper sampling other peoples music. It is not creating original music, its following rules it has created by sampling original music.

[K128kevin2]

all its doing is creating rules based on previous patterns it spots (neural networks).

if the futures patterns don't match its worthless (which is what happens in markets).

I listened to the music a while back, and it seems impressive until you realize its just a rehash of other peoples previously original work mixed up with some modifications, like a rapper sampling other peoples music. It is not creating original music, its following rules it has created by sampling original music.

But what you described is EXACTLY what humans do when we write music. There's a famous quote: "Good artists copy, great artists steal". A more accurate rewording of this would be "All artists steal". When a person composes music, it is a product of their musical experience. They aren't just throwing random notes onto a sheet. Every musical composition is a rehash of music that the composer has heard before. The best example of this is basically every blues song ever written. They all follow one of a couple possible chord structures, they generally use the pentatonic scale for melodies, and even the lyrics tend to be of relatively similar subjects most of the time. In short, musical compositions come from people's brains, and people's brains are formed by their experiences. Training a neural network on people's musical compositions and teaching it to write music itself is basically a simplified simulation of this.

To say that the neural network is creating "rules" is a bit of an oversimplification. They learn patterns, which you could argue are rules, but by that same argument everything is rule-based. Back to the blues music example, you could say that composing blues is rule-based. That actually wouldn't be terribly inaccurate. Like I said, you choose a chord progression and follow it... and then you take a bunch of short riffs from other blues songs and combine them into a melody... BAM, you have an original blues composition.

I would argue that the vast majority of high level tasks that human brains can perform can be boiled down to pattern recognition. Vision is pattern recognition - we recognize patterns in the information being sent through our optic nerve, and map those patterns to higher level neurological structures that represent things like "cat", "dog", "Obama", etc. Our brains receive information from our ears that can be mapped to higher level structures that correlate with higher level abstract representations of what we are hearing. We may classify a sound as a voice, and then at an even higher level, classify it as Morgan Freeman's voice (for example). And you keep extrapolating up to hear patterns that represent words, and patterns of words that represent sentences, and patterns of sentences that represent thoughts. Every step is just a pattern recognition task, that takes something like raw sound information represented by the binary behavior of neurons (they either fire, or don't fire), and extrapolates a high level coherent thought out of it.

Basically, human brains are simple. Or rather, the components of a human brain are simple. The way they interact with each other is obviously wildly complicated... but brains are basically just a bunch of binary units... a bunch of neurons that are connected to each other, but all they do is fire, or don't fire. It's how they are wired up that allows us to perform high level tasks like what I mentioned in the previous paragraph. There is no reason that computers can't theoretically simulate many of these tasks.

Sorry for the giant monolithic response haha... I just really enjoy this topic. I find it super fascinating and exciting!

[K128kevin2]

It's very impressive.

It really is. The discussion in this thread actually inspired me to read a lot about artificial neural networks last night.
This is not my subject but I am amazed. The learning part is very interesting and something I did not know too much about.

Very interesting project you have here and I am looking forward to see how it performs.

That's awesome that you are learning more about ANNs! They are super interesting!

[oda.krell]

Seriously no gloating here (I know how promising, fascinating, but also retardly difficult ML can be ), but I think your algorithm might need some tweaking...

For the past days, it's been predicting a strong upwards trends, basically: starting about *now* at each time step.

A few days ago:



Today:



By now, there's a wide gap between actual and predicted price, non surprisingly:




This doesn't per se invalidate your algo (could be that it aims for long-term correctness & profitability, but accepts short term deviations), but it leads to a few questions, mainly: did you properly *test* the network?

As in: did you keep training and testing data separate, and did you perform (several) runs on the testing data?

Assuming you did that, perhaps the current situation is an exception, and the predictions will 'snap back' to a more correct level (or price 'snaps back' to your predictions - though I doubt that a bit right now ).

Otherwise, it's possible your model needs to be made more 'reactive', i.e. when model and reality diverge, the model must be able to adjust faster and revise its previous output direction. I could think of a few ways to get there, but this likely requires changing the network architecture, which maybe you'd rather not do.