Binance future trader deals in multiple cryptocurrencies simultaneously. In default, the model predict 30min ahead, and trade every 1min.
- Note: If want to predict other time periods, you can change lookahead_window variable to what you want to predict.
.
├── Makefile
├── README.md
├── develop
│ ├── Makefile
│ ├── dockerfiles
│ │ ├── Dockerfile
│ │ └── requirements.txt
│ ├── src
│ │ ├── backtester
│ │ │ ├── __init__.py
│ │ │ ├── backtester_v1.py
│ │ │ ├── basic_backtester.py
│ │ │ └── utils.py
│ │ ├── common_utils_dev
│ │ │ ├── __init__.py
│ │ │ └── common_utils_dev.py
│ │ ├── dataset_builder
│ │ │ └── build_dataset.py
│ │ ├── rawdata_builder
│ │ │ ├── __init__.py
│ │ │ ├── build_rawdata.py
│ │ │ ├── candidate_assets.txt
│ │ │ └── download_kaggle_data.py
│ │ ├── reviewer
│ │ │ ├── __init__.py
│ │ │ ├── paramset.py
│ │ │ ├── reviewer_v1.py
│ │ │ └── utils.py
│ │ └── trainer
│ │ ├── datasets
│ │ │ └── dataset.py
│ │ ├── models
│ │ │ ├── __init__.py
│ │ │ ├── backbones
│ │ │ │ ├── __init__.py
│ │ │ │ └── backbone_v1.py
│ │ │ ├── basic_predictor.py
│ │ │ ├── criterions.py
│ │ │ ├── predictor_v1.py
│ │ │ └── utils.py
│ │ └── modules
│ │ ├── acts.py
│ │ └── block_1d
│ │ ├── __init__.py
│ │ ├── conv1d.py
│ │ ├── dense.py
│ │ ├── dense_block.py
│ │ ├── norms.py
│ │ ├── residual.py
│ │ ├── seblock.py
│ │ └── self_attention.py
│ └── storage
│ ├── dataset
│ └── experiments
└── services
├── Makefile
├── dockerfiles
│ ├── Dockerfile
│ └── requirements.txt
├── k8s
│ ├── admin
│ │ └── namespace.yml
│ ├── deployments
│ │ ├── data_collector.yml
│ │ ├── database.yml
│ │ └── trader.yml
│ ├── deployments-template
│ │ ├── data_collector-template.yml
│ │ ├── database-template.yml
│ │ └── trader-template.yml
│ └── service
│ └── service.yml
├── src
│ ├── common_utils_svc
│ │ ├── __init__.py
│ │ └── common_utils_svc.py
│ ├── config
│ │ └── __init__.py
│ ├── data_collector
│ │ └── data_collector.py
│ ├── database
│ │ ├── database.py
│ │ ├── models.py
│ │ └── usecase.py
│ ├── exchange
│ │ └── custom_client.py
│ ├── handler
│ │ ├── __init__.py
│ │ └── slack_handler.py
│ └── trader
│ ├── trader_v1.py
│ └── utils.py
└── storage
└── database
- Based on DenseNet.
- Add Squeeze and Excitation block before transition block
- Add Self-Attention block before transition block
- Use Activation function as tanhexp. selectable [Mish, SeLU]
- Use Dropout concept as Combine(Dropblock, SpatialDropout)
- Dropout not only for time-series, but also for channels.
We display the binary accuracy on levels. The levels show 10 quantiles which computed by insample-outputs of each factor(predictions, abs(predictions), probability).
The outputs of model are 2 factors, 1. abs(prediction), 2. probability of sign.
- abs(prediction) show the volatility like factor. We expect when abs(prediction) level is high, the performance is high.
- probability show the confidence score of positive and negative. We expect when probability level is high, the performance is high
The results below are after the commission was paid to the exchange.
Backtesting setting Commisions: {"entry": 0.04%, "exit": 0.02% | 0.04%, "spread": 0.04%}
trade_winning_ratio 0.704072
trade_sharpe_ratio 2.458127
trade_avg_return 0.002149
max_drawdown -0.175861
total_return 5.111693
Follow below steps.
- You can give additional variable to execute command with tag
make <command> ARGS="--variable1 <value> --variable2 <value>" - All code run on same container by reuse. Once if execute run container, it run continuously, and reused by other command.
✨make dev_run GPU_TYPE="<NVIDIA or AMD>": Run container to develop
✨make dev_download_kaggle_data ARGS="--username <kaggle_username> --key <kaggle_key>": Download historical pricing data
- In here, I implemented kaggle data downloader, but never checked. I recommend to download files manually through website.
- Download from https://www.kaggle.com/jorijnsmit/binance-full-history, then store to <pwd>/develop/storage/dataset/rawdata/raw/spot/
- Download from https://www.kaggle.com/nicolaes/binance-futures, then store to <pwd>/develop/storage/dataset/rawdata/raw/future/
✨make dev_build_rawdata: Build cleaned rawdata
✨make dev_build_dataset: Build features and labels to train model
✨make dev_train: Train model
✨make dev_generate: Generate predictions in test-periods
✨make dev_review: Check Performance and find best parameters by backtesting in virtual-env to trading.
✨make dev_display_review: Display performance plots, it should be run after make dev_review is done.
additional
✨make dev_rm: You can delete the container
✨make dev_bash: You can enter the container
✨make svc_install: Install requirements(kubectl, minikube), currently mac and linux environment is acceptable.
✨make svc_run: Run to trade in real environment.
It ask the params to set. You can use default setting by just enter without any input.
- LEVERAGE: How much use leverage
- EXP_NAME: If you set variable of exp_name when model train, you should give.
- REPORT_PREFIX: If you set varialbe of prefix when model review, you should give.
- REPORT_BASE_CURRENCY: Currently only USDT is acceptable. just skip.
- REPORT_BEST_INDEX: `You must set this, dont use default value! set which reviewer was best on model reviews`
- EXCHANGE_API_KEY: API_KEY of binance. You can set API_KEY of binance future test-net, when you want to check behavior in test-net.
- EXCHANGE_SECRET_KEY: SECRET_KEY of binance. You can set SECRET_KEY of binance future test-net, when you want to check behavior in test-net.
- TEST_MODE: Give True or False. Give False when test in test-net.
- WEBHOOK_URL: Give slack webhook url, the loggings will be sent to slack channel.
additional
✨make svc_rm: Delete only pods(containers)
✨make svc_reapply: Update changes without delete.
✨make svc_delete: Delete minikube. Clean-up way. If once delete, it takes time to re-run.
✨make svc_pods: Check status of pods.
✨make svc_db_bash: Enter database container.
✨make svc_dc_bash: Enter datacollector container.
✨make svc_td_bash: Enter trader container.