We were inspired by the current trends in Cryptocurrency and its extreme volatility due to various major social, political and economic events all over the world. A major barrier to newcomers becoming involved in Cryptocurrency is its perceived volatility and uncertainty. So, we attempt to show some rationale behind its massive fluctuations in price, to perhaps allow it more acceptance in the mainstream

What it does

Uses the massive power of large-scale Deep Learning systems today to accurately predict Bitcoin prices in the near (and not-so-near) future

How we built it

  • Used the Aylien API to collect relevant articles related to a keyword from all over the web, even rating sites based on their 'authority'
  • Used the Google Natural Language Processing API to classify these articles based on a sentiment value
  • Collected historical Bitcoin price info (using Coinbase), and stored it on a Microsoft SQL database running on Microsoft Azure, along with the sentiment values
  • Used a TensorFlow-based predictive model running on Google Cloud ML, trained on the above historical data, to predict future Bitcoin pricing
  • Tested this model by stepping back into various points in the past, and verifying that the predictions made by our predictive model correlate with what we now know to be reality

Challenges we ran into

  • Finding a (free, proper) news API
  • Connecting to a SQL server running on Microsoft Azure from multiple endpoints
  • Understanding TensorFlow
  • Creating a good machine learning model
  • Getting TensorFlow to run on Google Cloud

Accomplishments that we're proud of

  • The amazing amount of relevant information that we were able to scrape from the web
  • Just being able to do anything with TensorFlow
  • Manipulating massive amounts (along the order of 10^6) of data, and making some sense out of it

What we learned

  1. Make sure the thing you want to attempt is even possible before you want to do it
  2. Ensure that there are enough resources out there already that you can utilize and build upon
  3. Do lots of research first
  4. Buildings are really cold

What's next for CryptoCaster

  • Improving the algorithm for prediction
  • Including more historical data
  • Implementations for other cryptocurrencies
  • More powerful web-scraping for prediction

Built With

Share this project: