Bitcoin Forum

Hey guys,

This is my first post here and I'd like to show you guys a project I've been working on that is pretty relevant. I've been working for a few months on writing software that analyzes the history of bitcoin transactions and finds patterns in the rise and fall of prices. The software then predicts future prices based on these patterns, and it has a surprisingly low average error (~1.35% for prices within 24 hours and ~2.62% for prices within 5 days).

Well anyway if you are interested you can learn a lot more about the project on this website, which displays charts showing the predicted prices:
www.btcpredictions.com

The data is automatically updated every hour on the hour. Let me know what you guys think, I would greatly appreciate any feedback!

Good idea, but you are basing your numbers on a timeframe where there has been barely any volatility. Look at how tight of a range there has been for the high and low in the past week or so, there hasn't been much movement. You should test your algorithm when things start rapidly moving.

I wonder what will happen on 3/17 around noon to make the price drop to $318.

I hope it does so I can buy more  ;D

Nice easy setup with the site. As a web designer, I'm looking forward to seeing how this works out. Good luck!

It is true that the price is pretty stable right now, but the figures for average error are based on predictions during tens of thousands of points throughout the history of the exchange (bitstamp). The software looks at sections of prices in the past and tries to predict what the price became after those sections. It compares its prediction to the actual price data, and averages the errors for all of the predictions at each point in the historic data. That is how I got the figures 1.35% and 2.62% for average errors.

So basically what I'm saying is that the figures for average error are based on predictions made both during stable time periods and volatile time periods. However, you are still right - it is true that this is much more accurate when the price is less volatile. When the price starts changing a lot, it could have larger errors in predictions. This is mostly the case when real-world events affect the price of bitcoin, since obviously there is no way the software can foresee something like mt gox getting hacked or china changing its policy on bitcoins.


Haha crap I think that's an error... gonna have to fix that. Thanks for pointing that out lol

Edit: It's been fixed - thanks again for pointing that out! Gonna have to watch out for problems like that.

Some real world events that swing prices are foreseeable to some degree.  It would be awesome if your bot could scan news articles and find patterns, example: find an article that was posted by atleast 5 different sources with a same or very similar title/keywords.  Then the bot would have to determine whether or not the news is good or bad, or the bot could alert you of the news and you'll tell the bot whether or not the news is good or bad. 

I hate to say this because it's going to sound like I'm bashing your project (and trust me, I'm not), but any price prediction that goes mostly or solely off past data is not going to do well.  Investors and small time buyers see things in the news/forums.  If your bot is not parsing news/forums then you're predictions are going to fail.  It's possible, I suppose, that you could make predictions on non-volatile periods(like right now) based on past/historic non-volatile data.  But as soon as a big event happens, I'd scrap that completely and go based on news/events.

Yeah, that would be sick if it could scan news articles. That would be extremely complicated though, I wouldn't even know how to begin lol.

And you are right about it being less accurate when prices start to change a lot due to real-world events. However, even during the course of a spike or crash that resulted from something in the news, there are still patterns that the prices follow. Spikes and crashes may tend to play out in similar manners, and if the software detects that a spike/crash is occurring it may be able to predict how it will play out. To be honest though, this is more what I'm hoping rather than what I know. The software learns how to predict based on the entire history of transactions on bitstamp, so it has seen spikes/crashes. I haven't had a huge number of opportunities to see how well this software performs during spikes/crashes, but we'll see how it works as soon as the prices start fluctuating more I guess. Hopefully it will perform reasonably well.

I agree with this, unless you were to try to remove periods of volatility from the equation which I suspect would just make things worse.

It's a pretty interesting idea though!

I guess around 700$+ in this month, i'm gonna invest some money on bitcoin in couple of days for sure.

Some time in the next week ish I'm going to try to write an algorithm that will try to trade optimally based on these predictions, and it will simulate buying/selling 1 bitcoin. I'll display it's progress on the site and we can see if it is profitable :D

How far is the data fetched in bitstamp? Since the creation of the exchange? Do you plan to incorporate other exchanges in the future? Project sounds good

Great tool! Can you use your software with Litecoin?

The software bases its predictions off of a giant csv file found here: http://api.bitcoincharts.com/v1/csv/

And then the software also updates the contents of the csv file with the lastest transactions every like 2 seconds or something, since those files are only updated every like 12-24 hours ish. So yes, it looks at the entire history of transactions on bitstamp. And yes, it could be used for other exchanges.

As far as predicting litecoin prices, I would need to find or create a similar csv file or data file containing exchange information of a similar format. But yeah, I don't see why it wouldn't be able to predict litecoin prices. Just might require a bit of work as far as how the program gathers data from the file. Should work for stocks too, which I'm thinking of doing at some point.

Ugh predictions are so boring right now... I want the price to start fluctuating so I can really test this out

This will be interesting to keep track of over a longer period of time.  How long has the model been up?

Today is the fourth day that it has been up. The problem is that with the price being so flat right now it's difficult to really see how well it does. I mean, it has been predicting for the price to pretty much remain the same, but the real test will be to see how well it does when the price fluctuates.

Also I've been trying to test a program that simulates trading automatically based on these predictions, but the price isn't changing enough for it to make or lose any money. I included a 0.5% transaction fee so it just won't do anything if the price isn't going to change enough, which seems to be the case right now.

What are the inputs to a prediction? I assume you input prices for a certain number of days leading up to the prediction. How many days?

What kind of neural net are you using?

When I was first designing this, my original idea was to do like you said - input prices for a certain number of days leading up to the prediction. However, I found that this did not yield very good results. What works a lot better (and what this does) is using the percent change in price as input/output.

For the short term graph it looks at 60 hours leading up to the current hour and for the long term graph it looks at 80 hours.

The neural network is forward feed, has three layers(input, hidden, output), and uses backpropogation

can you make a little longer period time?
maybe two weeks or a month?

Longer predictions are one of the things I would like to work on for expanding this project. I wrote a small blurb about it in the "Coming Soon" page. The challenge with making predictions much longer than 5 days is that they just become less accurate. If I remember correctly, I believe that the last time I tried to do a 10 day prediction it gave me something like a 7 or 8% average error, which makes the predictions not particularly useful. However I do plan on experimenting with ways of structuring and training the neural network to see if I can make some reasonably accurate longer-term predictions.

Well BTC is pretty constant lately if you ignore a few crashes/bubbles