ivanproskuryakov / ta-predict Goto Github PK

ohlc time series training and forecasting of close price with keras & tf2

python -m venv .env
source .env/bin/activate

pip install -r requirements-mac.txt
pip install -r requirements-ubuntu.txt
pip install --force-reinstall -r requirements-mac.txt

Running application needs a database with the name ta_dev, as configured in src/parameters.py

psql -U postgres
create database ta_dev;

Once it's done sync db schema with the command, and after your good to go with predicting.

ENV=dev python db_flush_sync.py

In the current state of development models are trained to predict the closing price for ohlc pairs on USDT market.

By running model_predict.py application fetches time data for 1000 passed intervals, and guesses the only single value for next price . Available intervals: 5m, 15m, 30m, 1h, 2h, 4h, 8h, 12h, 1d, 3d. Smaller intervals are not well picked up by this type of prediction, listen.py to be used instead.

ENV=dev python predict.py 1h model/gru-g-50-1000-11-1m-BTCUSDT.keras

Realtime prediction is build on top of binance websocket api that provides realtime data for the last timeframe used in model, while the initial ohlc values are fetched though the http api which takes a couple of minutes. Downloading of ohlc data from 300+ assets takes around two minutes, meaning listen.py should be started ahead. Available intervals: 1m, 3m, 5m, 15m, 30m, ..., listener also works on larger intervals.

ENV=dev python listen.py 5m model/gru-g-50-1000-11-1m-BTCUSDT.keras

I suggest having a separate database ta_train for prevention of spoiling training dataset.

Train your models with command model_train.py

ENV=train python db_flush_sync.py
ENV=train python db_populate.py

ENV=train python model_train.py
ENV=train python model_plot.py 1m model/gru-g-50-1000-11-1m-BTCUSDT.keras

Same as with training, tests require ta_test db for prevention of spoiling training or dev databases.

psql -U postgres
create database ta_test;

Running tests

ENV=test python -m pytest test
ENV=test python -m pytest --log-cli-level DEBUG -s test/service/test_trader.py
ENV=test python -m pytest test/service/test_trader.py
ENV=test python -m pytest -s test/service/test_trader.py

In some cases binance may reject anonymous requests, for solving these add your api_key, api_secret to the environment to be read from klines service and library related.

open file with the editor of choice, ex nano .env/bin/activate, see https://stackoverflow.com/a/9554331 put your key in the end of the file

API_KEY="your key from binance here"
API_SECRET="your secret from binance"
export API_KEY
export API_SECRET

reactivate the environment with source .env/bin/activate so the keys would be picked up by the app.