CryptoAnalytics

CryptoAnalytics is a software artifact for the analysis and forecasting of financial time series and cryptocurrency price trends.

Installation

The Python version used in this project is 3.9.10. It is suggested to create a clean virtual environment with the correct version using pyenv.

Please refer to these instructions to install pyenv on your current OS. To install pyenv-virtualenv as well, please follow this documentation.

Once pyenv has been set up, you can install the required Python version and generate a new virtual environment as it follows.

pyenv install 3.9.10
pyenv virtualenv 3.9.10 venv

With the following you can activate your virtual environment.

pyenv activate venv

Use the package manager pip to install the project requirements as it follows.

pip install -r requirements.txt

Demo

A demo for the price prediction flow of Bitcoin (BTC), using a set of predefined args, is available on CodeOcean.

The demo uses a pre-built dataset (using the data_pull command) with the following characteristics:

1. coins: "btc", "eth", "usdt", "usdc", "xrp", "busd", "ada", "doge".

2. start: 15-08-2023.

3. end: 15-11-2023.

To launch the demo, start a reproducible run on our CodeOcean capsule.

Structure

This artifact is organized as it follows:

Software Deployment

Set of utilities to deploy CryptoAnalytics in a production environment using four different frameworks: Torchserve, BentoML, Mlflow (base) and Mlflow (with MLServer). For further instructions, please refer to the README file in the directory /deployment.

Price Prediction Flow

1. data_pull.py: command to generate a new dataset of OHLC cryptocoin prices from CoinMarketCap.
2. data_split.py: command to generate train and validation sets from the original data.
3. model_pretrain.py: command to pretrain ML models for cryptocoin prices forecast.
4. model_forecast.py: command to use the pretrained ML models to forecast cryptocoin prices.

Extra

1. correlation_analysis.py: command to analyze correlations among cryptocoin prices, can be useful for pre-selection of feature variables.

Usage

Price Prediction Flow

Data Pull

python data_pull.py -p "destination_path" -f "filename" -c "examples\coins.json" -s "01-11-2020" -e "01-11-2023"

Args:

• -p, --path: destination directory for dataset storage (OPTIONAL, defaults to current directory).
• -f, --filename: file name for pulled dataset (OPTIONAL, defaults to dataset_coinmarketcap_START_END).
• -c, --coins: path to .json file with list of coins to include in the dataset (OPTIONAL, defaults to top-20 coins for market cap).
• -s, --start: starting date for dataset pull, in format %d-%m-%Y (OPTIONAL, defaults to yesterday).
• -e, --end: ending date for dataset pull, in format %d-%m-%Y (OPTIONAL, defaults to today).

Data Split

python data_split.py -p "destination_path" -f "filename1" "filename2" -d "data_path" -v "avg_ohlc" -tr 0.8 -vd 0.2

Args:

• -p, --path: destination directory for train, validation and test dataset storage (OPTIONAL, defaults to current directory).
• -f, --filenames: file names for pulled dataset (OPTIONAL, defaults to [train_TODAY, valid_TODAY, test_TODAY]).
• -d, --data: path to original .csv file dataset (REQUIRED).
• -v, --variable: price variable to consider, either avg_ohlc or close (OPTIONAL, defaults to avg_ohlc).
• -tr, --train: ratio for train set split (REQUIRED).
• -vd, --valid: ratio for validation set split (REQUIRED).

Model Pretrain

python model_pretrain.py -p "destination_path" -f "filename" -tr "train_path" -vd "valid_path" -t "btc" -ft "examples\features.json" -m "lstm" -c "examples\config_nn.json"

Args:

• -p, --path: destination directory for pretrained model storage (OPTIONAL, defaults to current directory).
• -f, --filename: file name for pretrained model (OPTIONAL, defaults to model_TODAY).
• -tr, --train: path to the .csv file train dataset (REQUIRED).
• -vd, --valid: path to the .csv file validation dataset (REQUIRED).
• -t, --target: target coin to predict (REQUIRED).
• -ft, --features: path to .json file with list of coins to use as feature/predicting variables (REQUIRED).
• -m, --model: model to pretrain for inference, either lstm, gru, xgboost, lightgbm or catboost (REQUIRED).
• -c, --config: path to .json file with list of configs to use for pretraining (REQUIRED).

Model Forecast

python model_forecast.py -p "destination_path" -f "filename" -hz "horizon" -vd "valid_path" -pt "pretrained_path" -t "btc" -ft "examples\features.json" -m "lstm"

Args:

• -p, --path: destination directory for predictions storage (OPTIONAL, defaults to current directory).
• -f, --filename: file name for predictions (OPTIONAL, defaults to predictions_model_TODAY).
• -hz, --horizon: forecasting horizon (REQUIRED).
• -vd, --valid: path to the .csv file valid dataset (REQUIRED).
• -pt, --pretrained: path to the pretrained model (REQUIRED).
• -t, --target: target coin to predict, same as pretraining (REQUIRED).
• -ft, --features: path to .json file with list of coins to use as feature/predicting variables, same as pretraining (REQUIRED).
• -m, --model: model to use for inference, same as pretraining (REQUIRED).

Extra

Correlation Analysis

python correlation_analysis.py -p "destination_path" -f "filename" -d "data_path" -v "avg_ohlc" -w "daily" -m "pearson"

Args:

• -p, --path: destination directory for correlations storage (OPTIONAL, defaults to current directory).
• -f, --filename: file name for correlations dataset (OPTIONAL, defaults to correlations_TODAY).
• -d, --data: path to .csv file dataset (REQUIRED).
• -v, --variable: price variable on which computing correlations, either avg_ohlc or close (OPTIONAL, defaults to avg_ohlc).
• -w, --window: sliding time window to use for computations, either daily, monthly or weekly (OPTIONAL, defaults to daily).
• -m, --method: method to compute correlations, either pearson, kendall or spearman (OPTIONAL, defaults to pearson).

Examples

You can find examples of coin list (for data pull), configs (for pretraining/forecast) and feature variable list (for pretraining/forecast) in the directory /examples.

Risk Warning

Please keep in mind that no one can predict prices of cryptocurrencies with total certainty. Our price predictions can serve as guidelines and indications of future price developments, but are not intended to be used as investment advices.

Authors

Pasquale De Rosa, University of Neuchâtel, pasquale.derosa@unine.ch.
Pascal Felber, University of Neuchâtel, pascal.felber@unine.ch.
Valerio Schiavoni, University of Neuchâtel, valerio.schiavoni@unine.ch.

References

Pasquale De Rosa, Pascal Felber and Valerio Schiavoni. 2024. CryptoAnalytics: Cryptocoins price forecasting with machine learning techniques. In: SoftwareX, Volume 26, 101663. https://doi.org/10.1016/j.softx.2024.101663.

License

MIT