Bitcoin Forum
May 21, 2024, 09:15:15 PM *
News: Latest Bitcoin Core release: 27.0 [Torrent]
 
   Home   Help Search Login Register More  
Pages: [1]
  Print  
Author Topic: Official BTCPredictions Thread - Neural Networks for Price Prediction  (Read 1794 times)
K128kevin2 (OP)
Sr. Member
****
Offline Offline

Activity: 322
Merit: 250


View Profile
May 05, 2014, 11:04:49 PM
 #1

Hello! I am going to use this as the official thread for a project that I've posted here before so that I don't have to make a new thread every time there is an update Smiley

http://www.btcpredictions.com

This is a website that shows predicted bitcoin prices for the next 24 hours, 5 days, and 20 days. The predictions are made by software that uses artificial neural networks to analyze past prices, and the data is updated on the website every hour!

I recently added a feature that allows you to view past predictions against actual prices, and I've just expanded a bit on that feature by allowing you to navigate through those charts hour by hour. Here is the link:

http://www.btcpredictions.com/Historic.php

The default chart shows the prediction made 24 hours ago against the actual prices over the past 24 hours. You can press the back and forward buttons to go back/forth one hour at a time and see the different predictions made! Let me know if you guys have any suggestions, I was thinking it might be helpful to put the date you are currently looking at on the page but I haven't gotten around to that yet.
K128kevin2 (OP)
Sr. Member
****
Offline Offline

Activity: 322
Merit: 250


View Profile
May 14, 2014, 08:47:53 PM
 #2

I've been working on improving the neural network that makes predictions on my site by using a more complicated and informative input scheme and increasing the number of layers from three to four. The new network also makes predictions for prices over 15 minute periods for 24 hours instead of 1 hour periods like the current one. So far I have implemented the new input scheme and 15-minute predictions, but it is still using a 4-layer perceptron.

I have been testing this NN with various parameters and have been able to achieve an average error of around 0.67% so far! We will see how it behaves when an additional layer is added. It is possible that there will be some over-fitting (meaning that it won't predict well in actual practice) with the additional layer, but hopefully there will be some ways to work around this.
Pages: [1]
  Print  
 
Jump to:  

Powered by MySQL Powered by PHP Powered by SMF 1.1.19 | SMF © 2006-2009, Simple Machines Valid XHTML 1.0! Valid CSS!