The system for forecasting the course of cryptocurrency based on text analysis

Summary: The developed system allows predicting a bitcoin cryptocurrency rate. Machine learning was used for forecasting this rate. Any text arrays can be used as input parameters, for example, messages from Twitter social network are used now. In this system, it is possible to predict the course of a cryptocurrency pair for various time intervals (month, week, day, hour, minute). The messages used for training can be written in any language or in several ones at a time.
The system can be simply transferred to any other cryptocurrency pair. To do this, it’s enough to change the input parameters (messages related to another cryptocurrency pair). Tensor Flow, developed by Google as an open software library for machine learning, was used to train the neural network. The system has been learning during the specified period of time (over the past few years) and then it can predict the final rate by itself.

 

Technologies: Python, Tensor Flow, Natural Language Toolkit, Twitter/VK API, Git, Jenkins.

Screenshots: 

Project features:

  • Working with the Tensor Flow framework a large number of technical difficulties arose, due to the absence or too brief summary of the documentation for all used methods.
  • Full adhering to Scrum processes.
  • Using Git/Gerrit combination.

Project results:

  • A system for predicting cryptocurrency rates by the methods of predictive analytics has been created.
  • As a result of the system work, the neural network returns delta of rate change in percentage, for example, it returns 1.5%, that predicts the growth of the crypto currency rate by 1.5%.
  • The response of the neural network is correct, if the correct direction was selected and the predicted value deviated from the correct one by no more than 1% (up or down).
  • The prediction accuracy depends on the number of cryptocurrency messages. Approximately 7-8 thousand messages were used for training, that is a small amount for such a complex logic. As a result the following data on the accuracy of the predicted rate were received: by the time this document was being written, the accuracy of the predict direction (fall or grow) is about 60%. While the predict accuracy of the rate change delta is 25-30%.

Company's achievements during the project:

  • Project team proposed and implemented a big amount of suggestions for improving the rate value accuracy.