royyang0714 / binance-prediction-pytorch

Binance Prediction Pytorch Model

Main Results

ETHUSDT from 2021-01-01 00:00:00 to 2021-12-01 00:00:00

Time interval	ROI
1d (Human)	2.74%
1d (Model)	304.96%
4h (Human)	36.86%
4h (Model)	405.03%
1h (Human)	37.55%
1h (Model)	290.01%

BTCUSDT from 2021-01-01 00:00:00 to 2021-12-01 00:00:00

Time interval	ROI
1d (Human)	3.11%
1d (Model)	12.19%
4h (Human)	18.30%
4h (Model)	31.64%
1h (Human)	19.79%
1h (Model)	27.58%

Getting started

Environment

• Test OS: Ubuntu 16.04 LTS
• Python version: 3.8

Preparation

• Create folders.

mkdir images
mkdir checkpoints

• Please run pip install –r requirements.txt to install the needed libraries.

Dataset

Binance Public Data

• Clone the repo.
• Follow the instruction to download required data.

# ETHUSDT
python download-kline.py -s ETHUSDT -startDate 2017-08-01 -endDate 2021-12-01

# BTCUSDT
python download-kline.py -s BTCUSDT -startDate 2017-08-01 -endDate 2021-12-01

• It will download the required data as below. Unzip the zip files under the 1h, 4h and 1d directories.

binance_prediction_pytorch
    `-- binance-public-data
        `-- data
            `-- data
                `-- spot
                    |-- daily
                    `-- monthly
                        `-- klines
                            |-- ETHUSDT
                            `-- BTCUSDT

• Then soft link the data directory to the repo root as below.

binance_prediction_pytorch
    |-- binance-public-data
    `-- data
        `-- spot
            |-- daily
            `-- monthly
                `-- klines
                    |-- ETHUSDT
                    `-- BTCUSDT

Experiments

Training

• Run training and evaluation on ETHUSDT. It will store the checkpoints under checkpoints with ticker name and time interval if don't specify the checkpoint path with --ckpt.

# 1d
./run.sh ETHUSDT 1d

# 4h
./run.sh ETHUSDT 4h

# 1h
./run.sh ETHUSDT 1h

• Run training and evaluation on BTCUSDT

# 1d
./run.sh BTCUSDT 1d

# 4h
./run.sh BTCUSDT 4h

# 1h
./run.sh BTCUSDT 1h

Inference

• Specify the checkpoint path with eval mode to only do the inference.

./run.sh ETHUSDT 1h --ckpt ${YOUR_CHECKPOINT_PATH} --eval

Jupyter Notebook

• Directly try with jupyter notebook file Final.ipynb.