💵 Using Twitter and word2vec to predict trends in cryptocurrency. Normalized human sentiment goes in, the future of Bitcoin comes out.
The application of deep learning to financial forecasting is nothing new. When fed raw historical pricing data, a neural network can propose future trends based on the time-variant fluctuations of the past. In particular, Long Short Term Memory Recurrent Networks have been proven to be unreasonably effective at this task.
But alas, Economics is the science of choice.
The vast complexity of financial forecasting simply cannot be expressed by discrete pricing data, since there is an overwhelming human element. Passion, regret, life and death stand at the prescipice of every peak, and slump at the upset of every trough.
We live in an unprecedented informational era, where the entire breadth of human expression can be freely accessed via an API call. Here, we leverage Twitter to continuously gather and extract ongoing public sentiment, then use this to stimulate a neural network which fights to determine the cross-correlation with the future value of a given crypto.
Just as fake news quantifiably diminishes the public perception of truth, tweets about crypto can and will transmute bears into bulls. After all, Twitter is the Wall Street of blockchain currency.
This project relies upon defining a .env file to define your access credentials to Twitter. It takes the following form:
TWITTER_CONSUMER_KEY=XXXXX
TWITTER_CONSUMER_SECRET=XXXXX
TWITTER_ACCESS_TOKEN_KEY=XXXXX
TWITTER_ACCESS_TOKEN_SECRET=XXXXXYou can find more info here.
After this, you need to gunzip the word2vec-slim model to the public/ directory of your local clone.
Lastly, use yarn or npm install to initialize the project dependencies. Then execute the start script to begin training. After every training window, a prediction of the next cryptocurrency price will be made for the future.
backlogMinutes: The amount of time tweets and values need to be collected to satisfy a training data window.predictionMinutes: The amount of time into the future the network should forecast.subscribeTo: Define the kind of hastags to listen out for.
Note: The network will not start making predictions until backlogMinutes + predictionMinutes has elapsed.
Q: Is this project just for predicting bitcoin?
A: Any cryptocurrency served via the Coindesk API can be used as a prediction target; you just have to specify the ISO (i.e. "ETH"), and define the hashtags you think will be in some way related to the rise and fall of the currency's value.
