Random Forest BTC

Uses a random forest (RF) regression model to predict bitcoin prices 3 days ahead. Trained on data from 2010-2019.

Features

• Prediction of:
  • BTC price in USD 3 days ahead of input BTC data
• Visualization of:
  • RF model fit to test and training data

Basic Usage

Installation

Create venv and install requirements:

python3 -m venv <venv-name>
source <venv-name>/bin/activate
cd front/
make install

Note: If mac users receive SSL: CERTIFCATE_VERIFY_FAILED URLError, please run the following commands in terminal to Install Certificates.command file: open /Applications/Python\ 3.7/Install\ Certificates.command

This will allow the scraping to work for current day.

Current predictions

Run the front/Main.py file or equivalently:

cd front/
make run

load the resulting http://127.0.0.1:5000/ address in your browser.
The resulting prediction is based on the current day's BTC data.
To input custom data point, click on the manual input tab and fill in the appropriate fields.

Retrain model

To train your own model, and view prediction results, edit this part of the script:

    RFregressor = RandomForestRegressor(20)
    RFregressor.build_forest(X_train, y_train)
    y_pred = regressor.predict(X_test)
    print(y_pred)

where 20 in this case represents number of trees in the forest, X_train and X_test lists of dictionaries, each data point as a dictionary and y_pred a list of numbers, representing predictions of the tree.

Command line

run: python random_forest_regressor.py -{flags} (default behavior is to train a new model)

• training:
  • --output file name where model will be saved, should end in .pkl (default: ra_model{utc_time}.pkl)
  • --days number of days ahead to predict (default: 3)
• loading previous model:
  • --model pkl file name of previous model to be loaded

Results

Without retraining our model, to see the results based on 20 trees, bootstrap sample of original training size, and considering all features, simply run the random_forest_regressor.py file in the main folder.

These results should match those shown in the performance.txt file

Tests

Run the test_forest_pytest.py file.

File explanations

random_forest_regressor.py: the main backend source code.
front: contains main front-end source code.
final_forest_drop0.pkl: our final saved model in pickle form (where we removed most of the the data points when bitcoin had 0 value).
hyperparam_scripts: contains our scripts used to find ideal parameters.
features: combined to create raw_csvs files.
other_models: other models used to compare with our own.
performance.txt: comparison of the errors for those models and our own.

Acknowledgements

We used the following for our code: flask, sklearn, collections, matplotlib, pandas, numpy, and blockchain

Authors

Ryan Shuey, Evan Truong, & Elle McFarlane