The trading framework for Huobi Exchange
The Application of my research projects RLQuant & DLQuant

• Can be used to trade a portfolio
• Generate orders based on the actions of agents
• Trace orders and adjust prices if exceed the waiting threshold
• Multi-Thread order managing system
• Sets of configurable parameters in json
• Three models are integrated (Policy Gradient & Recurrent Policy Gradient & Direct RL)
• Models are implemented with two libraries (TF & Torch), (recommend to use Torch version, faster!)
• Setting trade time by yourself
• Find more from the code :)

• Python 3.5+
• numpy
• tensorflow
• pandas
• talib
• statsmodels
• PyTorch

• Multi-Task Recurrent Policy Gradient
  • Improved based on Recurrent Reinforcement Learning: A Hybrid Approach
  • Integrate RNN to encode the temporal correlation
  • Predict the observation of the return rate of next step (supervised approach)
  • Feed the output of RNN into a vanilla Policy Gradient
  • Slow, but better than Vanilla PG
• Policy Gradient
  • The implementation of vanilla Policy Gradient Policy Gradient Methods for Reinforcement Learning with Function Approximation
  • Simple, fast, stable
• Direct RL
  • Based on Deep Direct Reinforcement Learning for Financial Signal Representation and Trading
  • Directly optimize the expect return
  • Able to output continuous actions (weights)
  • Stable, can make good performance as RPG

The interpretation of config file:

{
  "models": { //path of models
	"PG_TF": {
	  "model_path": "./model_backup/PG_TF"
	},
	"RPG_TF": {
	  "model_path": "./model_backup/RPG_TF"
	},
	"RPG_Torch": {
	  "model_path": "./model_backup/RPG_Torch"
	},
	"DRL_Torch": {
	  "model_path": "./model_backup/DRL_Torch"
	}
  },
  "data": {
	"trade_bar_count": 200,
	"train_bar_count": 2000,
	"tick_interval": "60min" //{"60min","30min","15min","5min"}
  },
  "train": {
	"fee": 1e-5,
	"normalize_length": 10, //longer: more stable, shorder: more sensitive  
	"batch_length": 64,
	"learning_rate": 1e-3,
	"reward_threshold": 0.3, // prevent overfitting, stop at this threshold
	"max_training_epoch": 30,
	"train_length": 1500
  },
  "test": {
	"test_length": 400
  },
  "trade": {
	"base_currency": "btc",
	"debug_mode": true,
	"portfolio_config": "./config/portfolio_config.json", //json file with a simple list of currency name: ["bat","soc","wicc"]
	"model_type": "DRL_Torch", //one of above
	"account_file": "./config/account.json", //{"ACCESS_KEY": "your access key","SECRET_KEY": "your secret key"}
	"order_type": "limit", // {"limit","market"} recommend to use limit order
	"price_discount": 5e-3, //the price will be adjusted if exceed order waiting time, 
	                            for buy order: new_price=price*(1+pd) i.e. use higher price to buy when the market lacks of liquidity. 
	                            for sell order: new_price=price*(1-pd) i.e. use lower price to sell
	"amount_discount": 0.09, // Avoid exceeding the maximum transaction amount caused by calculation errors  
	"trace_order": true, // recommend to use
	"trade_time": [ // The exact minute to trade, for example: on every hour's [15,30,45,59] to trade with 15min bars
	  55
	],
	"max_asset_percent": 0.4, //Max weight for a single asset. The holding percent of each currency in the portfolio cannot exceed this threshold.
	"max_order_waiting_time": 120 //When tracing orders, if the order cannot be full filled with in this value(second), 
	                                a new order will be generate with discounted price, until all actions are executed.
  },
  "system": {
	"log_file": "./log/portfolio_log.csv"
  }
}

To launch your AI trader:

1. Create a config file and modify it with your own parameters

cp ./config/config_template.json ./config/config.json

2. Create your account file with your access key and secret key

echo "{"ACCESS_KEY": "your access key","SECRET_KEY": "your secret key"}" >> ./config/account.json

3. Create your portfolio file with the asset you want to trade

echo "["bat","soc", "wicc","edu"]" >> ./config/portfolio.json

4. Build your model

python3 portfolio_manager.py build_model

When it passes the backtest with the target return rate you set, it will automatically save the model parameters, then, you can use it to do live trading.

5. Live-Trading!!!

If you want to test if the system works correctly, you can choose to let it trade for once.

python portfolio_manager.py trade_now

If all things are ready

• Set the debug_mode to false in your config file
• Make sure you have confidence in the market and agents
• Prepare to launch you AI trader

python portfolio_manager.py trade

Or you can use the shell trade.sh

6. Check the system periodically

There is always risk of loss in trading. All trading strategies are used at your own risk

The volumes of many cryptocurrency markets are still low. Market impact and slippage may badly affect the results during live trading.

• Use the information of order book when managing orders
• Implement a security selection model based on maybe, fundamental data
• Maybe More