Bitcoin Forum

Bitcoin => Project Development => Topic started by: SheffieldCrypto on December 29, 2015, 02:37:19 PM



Title: Trading systems needed - machine learning forecasting start-up
Post by: SheffieldCrypto on December 29, 2015, 02:37:19 PM
I work with two PhD’s that develop machine learning models to forecast financial markets, through out this year we have been experimenting with applying them to bitcoin and other digital currency markets. One model we have developed seems find various inefficiencies seen in the markets that govern ‘alt coins’.

Here is a link to the model running on an AWS. http://ec2-52-27-84-167.us-west-2.compute.amazonaws.com:8000/webctrade/?modelname=TATraderMulti&coinname=DASH

The model is based on Hawkes processes and recurrent neural networks. It is currently set up on a 24hour time frame and makes a new prediction at 00:00 UTC every night. One innovation was not predicting price directly but to optimize for trading profit directly. The signal seen represents the percentage of your portfolio  (each coin trades it’s own portfolio) you should invest / the confidence and polarity of a price move on the next time step

Please view the backtest we ran across multiple coins at the end of 2014: http://sheffieldcrypto.com/wp-content/uploads/2015/06/backtest-1.txt

Scrolling to the bottom, one can observe the ‘pnl portfolio_value’. Starting with $5,401 the trader finished with $25,671 after a 3 months period, this is taking into consideration exchange fees however the trader assumes it could make a trade on the closing price, given the liquidity of some of these markets this may have not been possible, this requires some kind of strategy that optimizes for intra day trading during periods of high volume.

At times during this backtest you can see the trader made much more than the final profit however it was not able to anticipate the flash crashes seen on blackcoin.

Besides this model we are also developing a deep learning model and NLP (natural language processing) model, this will utilize various social media as its inputs.

What we require now is someone with experience in trading systems who can help us implement our model into a trading scenario (the model is currently configured to poloniex and written in python). www.genesismining.com have shown interest in partnering with us, they have an existing system that connects to many exchanges, the API to connect to their infrastructure can be seen here http://pastebin.com/zM8A3MeG. We would also be open to connecting to poloniex directly.

If you have any questions, please contact us directly at 
dave@sheffieldcrypto.com
 
Regards,
Dave
www.sheffieldcrypto.com