Project Overview

In this project, we'll predict the future price of Bitcoin. We'll use historical data on the price of Bitcoin, along with data from Wikipedia about edits to the Bitcoin page. We'll merge and combine this data, then use it to train a random forest model that will tell us if Bitcoin prices will increase or decrease tomorrow. We'll then switch to an XGBoost model and better predictors to improve accuracy.

We'll develop a backtesting system and use a robust error metric so we can tell if the algorithm is performing well.

This project can be extended to other cryptocurrencies as well.

Project Steps

• Load in data
• Clean and merge data
• Create an initial machine learning model and estimate accuracy
• Switch to a more powerful model and improve our predictors

Code

You can find the code for this project here

File overview:

• prediction.ipynb - a Jupyter notebook that contains the code to predict Bitcoin prices
• sentiment.ipynb - a Jupyter notebook that creates our wikipedia edit dataset.

Local Setup

Installation

To follow this project, please install the following locally:

• JupyerLab
• Python 3.8+
• Python packages
  • pandas
  • yfinance
  • scikit-learn
  • xgboost
  • mwclient
  • transformers

Data

Computing the Wikipedia edit data takes time. It can be faster to use the version that's already been generated. It's in this repository, and called wikipedia_edits.csv. Feel free to download and use the file. We'll be downloading the Bitcoin price data using the yfinance package as part of the project.

Running

First, run the code in sentiment.ipynb to generate a new Wikipedia edits dataset. The dataset committed in the repo is old, and this will get the edits up to the present day.

Second, run the code in prediction.ipynb. By default, this will load data from an existing btc.csv file. Removing that code will ensure that it downloads the newest data from Yahoo Finance.