🍊BitaMin Project

• BitaMin, data analysis & data science assosiation, 12th and 13th joint project (2024.03.06 ~ 2024.06.05.)
• Time Series and Reinforcement Learning for Stock Trading

Team

송규헌	송휘종	서영우	이태경	정유진	정준우
Github	Github	Github	Github	Github	Github

✅Table of Contents

• 💼Project Introduction
  • Overview
  • 🔖Reference
• 🤗Environment
• 🦾Training
• 🛠️Test (Backtrading)
• 🛠️Test (System trading)

💼Project Introduction

Overview)

◾ Project Topic

Maximizing Portfolio Value by Time Series Forecasting and system trading using Reinforcement Learning.

◾ Goals

• Use Time Series Forecasting to predict the stock prices (high, low, maximum fluctuation rate, etc.) for the next 5 days and then select the 6 stocks with the highest (high-low) difference.
• Train Reinforcement Learning on the selected stocks and implement system trading

◾Flowchart

🔖Reference

◾Quantylab

We used baseline code of quantylab's rltrader for training model. Modified for real-time trading.

🤗Environment

• recommend making a conda virtual environment
• Python 3.7+
• PyTorch 1.13.1

conda install python==3.7
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
conda install pandas
conda install matplotlib
conda install -c conda-forge ta-lib
pip install -r trading_requirements.txt

🦾Training

• cd rltrader/src/
• refer to quantylab for detailed parameters descriptions.
  e.g.)

python main.py --mode train --ver v1 --name 001470_202404151325_202405241530 --stock_code 001470 --rl_method ppo --net cnn --backend pytorch --balance 500000000 --start_date 202404151325 --end_date 202405241530

• --mode : set 'train' when training model
• --ver : leave it with 'v1' for out code
• --name : set name of output directory and model file
• --start_date & --end_date : 'year-month-day-hour-minute'(%Y%m%d%H%M) format

When training procedure ends model parameters, output images and log are stored in /rltrader/models/ and /rltrader/output/

🛠️Test (Backtrading)

check the start_date and end_date of stock. start_date should be the 120 time steps before you want to start the test because of input size of CNN.
e.g.) if you want to backtrade from 2024.05.27 09:01 to 2024.05.31 15:30 then you should set the start_date as 202405241322 which is 120 time steps before the 2024.05.27 09:01.

Also you should set --name same as the train name in order to make sure rltrader use trained model for inference.

python main.py --mode test --ver v1 --name 001470_202404151325_202405241530 --stock_code 001470 --rl_method ppo --net cnn --backend pytorch --balance 500000000 --start_date 202405241320 --end_date 202405311530

🛠️Test (System Trading)

setting

You should place 'api.json' in order to use KIS API for trading.

• Make KoreaInvestment account.
• Follow the menu ‘트레이딩’ > ‘모의투자’ > ‘주식/선물옵션 모의투자 > 모의투자안내’ > ‘신청/재도전’
• Set initial balance and trading period as you wish.
• Go to API center KIS API.
• Apply for an API(url).
• Set your api.json file as below and place the file under /bitamin1213_trading/rltrader/src/quantylab/rltrader/

{
    "real_invest" : {
        "account" : "실전투자계좌번호8자리-01", 
        "app_key" : "실전투자계좌 APP Key 복사해서 붙여넣기",
        "app_secret" : "실전투자계좌 APP Secret 복사해서 붙여넣기",
        "access_token" : ""
    },
    "mock_invest" : {
        "account" : "모의투자계좌번호8자리-01",
        "app_key" : "모의투자계좌 APP Key 복사해서 붙여넣기",
        "app_secret" : "모의투자계좌 APP Secret 복사해서 붙여넣기",
        "access_token" : ""        
    }    

}

• Get access token by runnig utils.py

cd /bitamin1213_trading/rltrader/src/quantylab/rltrader
python utils.py

system trading

For real-time system trading, use 'predict' keyword. Setting start_date is same as backtrading, 120 time steps before.
You should run system trading code day by day. 'predict' method is just for a day trading.

python main.py --mode predict --ver v1 --name 001470_202404151325_202405241530 --stock_code 001470 --rl_method ppo --net cnn --backend pytorch --balance 500000000 --start_date 202405241320 --end_date 202405281530 --is_start_end 2 --is_mock 1

• --is_start_end : 0 for Monday, 1 for Tuesday~Thursday, 2 for Friday.
• --is_mock : 0 for real investment, 1 for mock investment.

Access_token is revoked after the system trading code ends. So you should get access_token before running predict code everyday.