20 Day Price Forecast (8.2% error)

[K128kevin]

Hey guys,

So I've posted this here before, but I'm working on a software project that uses neural networks to predict bitcoin prices. Predictions can be seen here:
http://www.btcpredictions.com

A few people were asking me about predictions beyond 5 days, which was the longest forecast I had until today. I was able to get the software to make "reasonable" 20 day predictions. I say "reasonable" but really I would say they are probably very far from reliable. The average error is generally between 8.1% and 8.3% (meaning that if it predicts $500, the price will usually be between $460 and $540). So the 20 day prediction probably is not useful for trading, but it is interesting because it reflects the general patterns that bitcoin prices are taking over a slightly longer period of time.

And remember, the 20 day prediction is new and still very much in a "beta" stage. I will try to improve on it, but I can't promise it will ever be significantly more accurate.

[1715656714]

1715656714

[1715656714]

1715656714

[1715656714]

1715656714

[Cassius]

Nice, great effort.

[Cassius]

Out of interest, what happens when you back-test with historical data? Specifically, does it predict the big dips and flash crashes as well as the trend? And do long long-term forecasts roughly stick to the exponential trend? It would be interesting to find out!

[segeln]

Hey guys,

So I've posted this here before, but I'm working on a software project that uses neural networks to predict bitcoin prices. Predictions can be seen here:
http://www.btcpredictions.com

A few people were asking me about predictions beyond 5 days, which was the longest forecast I had until today. I was able to get the software to make "reasonable" 20 day predictions. I say "reasonable" but really I would say they are probably very far from reliable. The average error is generally between 8.1% and 8.3% (meaning that if it predicts $500, the price will usually be between $460 and $540). So the 20 day prediction probably is not useful for trading, but it is interesting because it reflects the general patterns that bitcoin prices are taking over a slightly longer period of time.

And remember, the 20 day prediction is new and still very much in a "beta" stage. I will try to improve on it, but I can't promise it will ever be significantly more accurate.

Hi,nice you made a 20 day forecast. But I did`nt want it because I think it is very,very , very far from reliability same as a longterm weatherforecast.
Honestly I do not understand why you did it and neglecting the Need of an update-time-intervall shortage (to 5 Minutes)This would be much more useful,since th prediction get much more reliable.
Hope you are working on that

[K128kevin]

Cassius - I'm not sure what happens because I don't actually see the predictions, but maybe I'll alter it so that I can. That would be interesting to see. I have a feeling it would predict very mild spikes and crashes at certain times but nothing to the same magnitude as the actual fluctuations.

segeln - I did it because I thought it would be interesting to see and other people had expressed interest in seeing it as well. I will definitely work on having it update more often too, don't worry xD

[Cassius]

Cassius - I'm not sure what happens because I don't actually see the predictions, but maybe I'll alter it so that I can. That would be interesting to see. I have a feeling it would predict very mild spikes and crashes at certain times but nothing to the same magnitude as the actual fluctuations.

segeln - I did it because I thought it would be interesting to see and other people had expressed interest in seeing it as well. I will definitely work on having it update more often too, don't worry xD

Ah, I figured long-term previous data had been used to 'teach' the software.
Again, would be interesting to see if the dips were predicted, just less severe (which might make sense, as you suggest) - at which point, maybe there would be a fix to increase volatility... who knows.

[K128kevin]

Cassius - That actually is how it works... the long-term previous data is used to 'teach' the neural network. I just haven't been having it output the predictions it makes while it's learning. I would be curious to see that data too though, so I think I will have it do that. Maybe I'll put up a graph showing this compared to the real historic prices.

[Cassius]

That would be cool - especially if the lines converged as it learned with each passing week.

[rebuilder]

Cassius - That actually is how it works... the long-term previous data is used to 'teach' the neural network. I just haven't been having it output the predictions it makes while it's learning. I would be curious to see that data too though, so I think I will have it do that. Maybe I'll put up a graph showing this compared to the real historic prices.

I'm confused now. Isn't the average error a measure of how much the algorithm deviates from historical prices?

[segeln]

segeln - I did it because I thought it would be interesting to see and other people had expressed interest in seeing it as well. I will definitely work on having it update more often too, don't worry xD

so I don`t worry
I wish you all the best for your work,though I am sceptic about the 20-day longterm prediction
Thanks for replying

[akujin]

[Cassius]

Even 8% average error gives a solid UP from here. Suggest we return in 1/5/20 days to evaluate...
Hmm.. I wonder what the reliability is in very basic up/down terms. That would be cool, giving a 20-day bull/bear prediction that had a strong chance of accuracy.

[K128kevin]

rebuilder - Okay I think we're not communicating clearly here, let me try to explain how it works as clearly as possible:

The neural network trains on historic data. The training process involves trying to make predictions at certain points in history and then changing parts of the neural network based on how far off the predictions were from the actual prices. The average error is how far off the predictions were once the network was done training. So basically the average error represents how far off each predicted price was from the actual price for predicting prices at every hour (or 6 hours, or 12 hours) in the past 3 years. I hope this is more clear, but please let me know if it isn't.

Cassius - I think if it just tried to predict up or down that it would be very accurate. I would guess that it would be right at least like 90% of the time, but I'm just speculating.

[Cassius]

Cassius - I think if it just tried to predict up or down that it would be very accurate. I would guess that it would be right at least like 90% of the time, but I'm just speculating.

That's what I figured. And 90% is a lot better than 50%, which is roughly what you've otherwise got. Whilst that's not so much use on the day-to-day level due to the relatively small fluctuations, it is on the scale of a month.

[OgNasty]

This is great.  Would love to see a graph showing the back data as well. 

[segeln]

rebuilder - Okay I think we're not communicating clearly here, let me try to explain how it works as clearly as possible:

The neural network trains on historic data. The training process involves trying to make predictions at certain points in history and then changing parts of the neural network based on how far off the predictions were from the actual prices. The average error is how far off the predictions were once the network was done training. So basically the average error represents how far off each predicted price was from the actual price for predicting prices at every hour (or 6 hours, or 12 hours) in the past 3 years. I hope this is more clear, but please let me know if it isn't.

could you elaborate this more, please.Very interesting.
I thought you were checking and calculating the Bitstamp orderbook.
I put your link in the german forum of bitcoin.de = coinforum.de

[rebuilder]

K128kevin:
OK, so I had the right idea, you just haven't looked at the datapoint-by-datapoint comparison, only the average error rate output in the end.

Interestingly, the short-term prediction is now forecasting a return from the 570's to about 536 in 15 hours time, as a more or less continuous downward slope. Seems like a significant enough dip to watch for comparisons.

Edit: make that 13 hours, and 532 or so - the prediction just updated as I was typing.

[m3g4tr0n]

This is great.  Would love to see a graph showing the back data as well. 

Yes, and the ultimate way of showing it would be to let the 24 hour graph shift incrementally out the left side of it's box into a real time graph. That way viewers could see the prediction accuracy overlaid on top of what is actually happening. I would definitely donate something if that scenario came to pass.

good work

[razfaz]

Great work!
Having a major in CS, I was looking a bit into algorithmic trading too. Until now I just did it on a very basic level, using simple stuff to base my choices on. The results often were that the capital rose by some (2-3) percents in a month. Having hodl't and sodl't gave mostly better results. Due to the lack of time I did not investigate it in more detail.

Which programming language do you use? Are there any frameworks supporting neural network programming?

[K128kevin]

OgNasty/m3g4tr0n - yeah a lot of people have been asking about that, I think I'll try to make graphs showing this some time soon, or maybe extend the current ones to the left like you said.

segeln - Thanks for posting it to the German forum! Here's a bit of elaboration:

The part of a neural network that determines what results it will give (predictions, in this case) is "weights". A neural network consists of a bunch of nodes with edges between them, and the edges have certain weights. Initially these weights have random values. You teach a neural network by having it try to do calculations on data where you know the answer, and then adjust the weights based on how far off it was. So in this case, I have the neural network look at data at a few tens of thousands of points throughout the history of bitstamp and try to make a prediction. Initially the predictions are WAY off because the weights are just random. However, it sees how far off it was and adjusts the weights accordingly. Each time it trains on the historic data it is slightly more accurate. It takes about 25-30 ish training sessions for these neural networks to be able to predict with the average errors that I have listed on the site.

Let me know if you have any more questions!

rebuilder - Yeah I see it's predicting a pretty significant dip... I'm pretty nervous that it's going to be completely wrong lol but we'll see. I would not guess that it will dip down into the 530s or 540s but I suppose it's possible. I think what will happen is that it will dip but not by as much as it is predicting.

razfaz - Thanks! And that's pretty awesome if you were able to consistently earn 2-3% per month!

I use java for this project. I believe there are frameworks and libraries and stuff for neural networks, but I didn't use them. A few semesters ago I had to create a neural network for a class assignment, so I took the code from that and adjusted it accordingly to work for this problem (the old one was used for recognizing hand-written digits).

[Sindelar1938]

Sounds like a very promising prediction approach

Have bookmarked your website with a view to following

[K128kevin]

Sounds like a very promising prediction approach

Have bookmarked your website with a view to following

Thanks! I hope the predictions serve you well

[akujin]

[K128kevin]

This is interesting, thanks for tracking it's degree of accuracy like this!

I wouldn't say this is the neural network's best work but it's not too bad I guess. It missed that bump in the beginning which probably threw it off a bit, but it seems to have done a pretty good job at other points in the graph. We'll see how it continues in the future!

[Cassius]

Just curious - on the 20-day prediction the price drops to $556 on 27/3 (two days time) but this doesn't happen on the 5-day prediction. Is this a bug or an artefact of the way they are calculated?

[K128kevin]

Just curious - on the 20-day prediction the price drops to $556 on 27/3 (two days time) but this doesn't happen on the 5-day prediction. Is this a bug or an artefact of the way they are calculated?

It's probably a result of the fact that the 20 day prediction has such a high average error. I wouldn't count on the 20 day prediction being particularly accurate. In all probability there will be some real-life events that occur in that time period which unpredictably change bitcoin prices. Those predictions are just a lot less reliable.

[m3g4tr0n]

10 Hours ago, it was predicting ~535 for 9-10pm Eastern time, now it's 2 hours away and it's over 7% higher at ~575.

I want to see this thing be great, but that's a huge difference.

[solarflare]

Would be nice to have the error margin projected on the graph!

[K128kevin]

10 Hours ago, it was predicting ~535 for 9-10pm Eastern time, now it's 2 hours away and it's over 7% higher at ~575.

I want to see this thing be great, but that's a huge difference.

Yeah it did mess that up - that was a pretty big change. A 7% change in the 24 hour prediction is pretty rare and I think that's the first change that far outside the average margin of error that I've seen. That will happen sometimes though, and trading based on these predictions may cause people to lose money sometimes. However, I believe it is correct more often than incorrect and that in the long run, if someone were to trade based on these charts they would end up earning money. That's why I'm testing the trade simulator project I describe in the Coming Soon page - so far it's earned a small amount of money but it hasn't had a lot of opportunities yet since prices have been pretty stable.

[arepo]

10 Hours ago, it was predicting ~535 for 9-10pm Eastern time, now it's 2 hours away and it's over 7% higher at ~575.

I want to see this thing be great, but that's a huge difference.

Yeah it did mess that up - that was a pretty big change. A 7% change in the 24 hour prediction is pretty rare and I think that's the first change that far outside the average margin of error that I've seen. That will happen sometimes though, and trading based on these predictions may cause people to lose money sometimes. However, I believe it is correct more often than incorrect and that in the long run, if someone were to trade based on these charts they would end up earning money. That's why I'm testing the trade simulator project I describe in the Coming Soon page - so far it's earned a small amount of money but it hasn't had a lot of opportunities yet since prices have been pretty stable.

i'm not sure trading based on this alone is a very good idea. using it to inform trading decisions along with other inputs is probably a safer strategy. an obvious weakness is that although the average error may be single-digits, there could be individual instances where the actual error is quite large, though these are rare enough that the average error remains low. depending on the size of your margin, one such instance could easily destroy your trading profitability.

--arepo

[Twilight_Sparkle]

Would be nice to have the error margin projected on the graph!

+1

[K128kevin]

i'm not sure trading based on this alone is a very good idea. using it to inform trading decisions along with other inputs is probably a safer strategy. an obvious weakness is that although the average error may be single-digits, there could be individual instances where the actual error is quite large, though these are rare enough that the average error remains low. depending on the size of your margin, one such instance could easily destroy your trading profitability.

--arepo

I completely agree. I would use these predictions as a guideline but I definitely wouldn't follow them strictly 100% of the time. You definitely have to read up on the latest news and factor that into your decision as well. I believe that combining the neural network predictions with your own (presumably reasonable and somewhat intelligent) judgement could earn you a lot of money.

[Alley]

I think this is going to be a bad few days for bitcoin price.  This IRS news will confuse and frustrate people and buying will be low.  I don't see it going over $600 for at least a week.  Maybe a month.

[K128kevin]

I think this is going to be a bad few days for bitcoin price.  This IRS news will confuse and frustrate people and buying will be low.  I don't see it going over $600 for at least a week.  Maybe a month.

hmm maybe... it looks like the price is dipping down a bit. I think that the IRS taxing bitcoin is beneficial in the long run though. Bitcoin has to be taxed somehow for it to be viable as a currency, plus taxing it is a form of recognition and this gives it some more legitimacy.

[rebuilder]

The algorithm seems to disagree with current price action quite violently... Right now it's predicting a rise to the 580's within 24 hours.

[K128kevin]

Yeah I know, I'm curious to see if it shoots back up again. I think that it will, but it's hard to be sure. Like I wrote on the about page, it can have trouble when real-life events significantly affect bitcoin prices (like the China news in this case). So we'll see how well it does.

[segeln]

it can have trouble when real-life events significantly affect bitcoin prices (like the China news in this case).

what about data-updating more quickly ? So you are better reacting to real life Events.
I will repeat my proposal until you have experienced many times that you should have done it earlier.

[Cassius]

Today opened at 580 so it's not impossible. Seems unlikely but let's see what THE MACHINE says (Person of Interest, anyone?).

[K128kevin]

what about data-updating more quickly ? So you are better reacting to real life Events.
I will repeat my proposal until you have experienced many times that you should have done it earlier.

Haha yeah you really seem insistent on that upgrade lol. I do plan on doing that in addition to some other ways of possibly restructuring the way predictions are made in an attempt to increase accuracy. Right now I just got back to school from spring break though and I have a ton of work -_- haven't been able to do a lot with the predictions yet. I'll probably get started on this stuff around Wednesday next week... I should have much more free time by then.

[K128kevin]

Well this is not looking particularly promising at the moment lol

[porcupine87]

The average error is generally between 8.1% and 8.3% (meaning that if it predicts $500, the price will usually be between $460 and $540).

So you wanna say, that the in a 50% confidence intervall price will be in the intervall +-8% of the current price? OMG. I'll make millions with this information! I don't know. It seems to be nice but I need no neural network for this info.

Just out of curiousity. How far does your neural network go back in time? I just wonder why all lines go up all the time.

[K128kevin]

So you wanna say, that the in a 50% confidence intervall price will be in the intervall +-8% of the current price? OMG. I'll make millions with this information! I don't know. It seems to be nice but I need no neural network for this info.

Just out of curiousity. How far does your neural network go back in time? I just wonder why all lines go up all the time.

Well first of all, it doesn't always predict the price to rise. It predicts for the price to fall as well, though it would make sense if it predicted for the price to rise more than fall since that is what it has done more of throughout history.

Second, yeah the 20 day graph isn't very reliable. That's why I have the note there, and I also specify this in the about page. The 20 day chart only exists because it is interesting to see and a lot of people asked for it.

There is no confidence interval for where the price will fall in relation to the average error. The average error means that on average, predictions are off by that amount. If you have the neural network make predictions at every point in bitstamp's history and average how far off it was, you will get those numbers that I have put above the graphs. It probably comes more down to something like 90% of the time the price is in that interval, and 10% of the time it is a relatively significant amount off.

This tool isn't something you want to use exclusively and just blindly follow. However, I do think it is useful if you combine it with looking at the news and using your own judgement.

[porcupine87]

So you wanna say, that the in a 50% confidence intervall price will be in the intervall +-8% of the current price? OMG. I'll make millions with this information! I don't know. It seems to be nice but I need no neural network for this info.

Just out of curiousity. How far does your neural network go back in time? I just wonder why all lines go up all the time.

Well first of all, it doesn't always predict the price to rise. It predicts for the price to fall as well, though it would make sense if it predicted for the price to rise more than fall since that is what it has done more of throughout history.

Second, yeah the 20 day graph isn't very reliable. That's why I have the note there, and I also specify this in the about page. The 20 day chart only exists because it is interesting to see and a lot of people asked for it.

There is no confidence interval for where the price will fall in relation to the average error. The average error means that on average, predictions are off by that amount. If you have the neural network make predictions at every point in bitstamp's history and average how far off it was, you will get those numbers that I have put above the graphs. It probably comes more down to something like 90% of the time the price is in that interval, and 10% of the time it is a relatively significant amount off.

This tool isn't something you want to use exclusively and just blindly follow. However, I do think it is useful if you combine it with looking at the news and using your own judgement.

Ah ok, I read it wrong. I thought 8% is the average error for the daily prediction. 1.3% for the daily prediction and 2.59% for the 5 day is good. I chanllenced that with the prediction (price today = price tomorrow = price in 5 days) and the error was between 2.8 and 6% for the daily prediction and 4 to 16% for the 5day. I started at 4.feb and went back 1 month, 3m, 6m, 1y and 2y.

So when your network is still prediction higher prices although we have a down trend since 2 months I guess you go back a few months, right?

[porcupine87]

But:

The average error means that on average, predictions are off by that amount. If you have the neural network make predictions at every point in bitstamp's history and average how far off it was, you will get those numbers that I have put above the graphs.

There could be a logical flaw. Think about it. The more neurons you take, the better you can adjust it to the history. Take 1000 neurons and a big amount of computing power and you can lower the average error to 0.1%. But do you really think you can predict the future better?

Did you ever test your model? I don't mean the error you produce by taking the learning data. I mean the actual prediction for an at that point unknown future. For example: You are prediction 489.2$ for tomorrow 4pm GMT. Now save that prediction and compare with the real value tomorrow.

I think then you get a much larger number than 1.3%

[pedrosoft]

very very interesting

[Hpotsirc]

Did you ever test your model? I don't mean the error you produce by taking the learning data. I mean the actual prediction for an at that point unknown future. For example: You are prediction 489.2$ for tomorrow 4pm GMT. Now save that prediction and compare with the real value tomorrow.

I think then you get a much larger number than 1.3%

I think this is because the error adds up. So 1.3% only applies to the forecast of next hour and when you predict for example 12 hours ahead you could end up being 15.6% off.
It could propably be helpful to give additional information like average error of predicted price 6/12/18/24 hours ahead. That gives a better impression of the capabilities of this model.

[honorem]

These are terrible predictions