Skip to content
/ trials Public

Our codebase trials provide an implementation of the Select and Trade paper, which proposes a new paradigm for pair trading using hierarchical reinforcement learning. It includes the code for the proposed method and experimental results on real-world stock data to demonstrate its effectiveness.

License

MIT license
111 stars 27 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

The-FinAI/trials

BranchesTags

Repository files navigation

Trials

zh en python pytorch black pre-commit poetry license Discord

Automatic Pair Trading

Description

Our codebase trials provide an implementation of the Select and Trade paper, which proposes a new paradigm for pair trading using hierarchical reinforcement learning. It includes the code for the proposed method and experimental results on real-world stock data to demonstrate its effectiveness.

Maintainer

Boyi Zhang

Prerequisite

The requirements to develop the project.

  • Python ^3.8
  • Git
  • Poetry
  • Python IDE (Visual Studio Code is recommended!)

⚠️ Read the official manual or documentation of following projects before write any code ⚠️:

  • git, a version control system
  • Poetry, a python library to manage project dependencies
  • towncrier, a changelog generater
  • pre-commit, a git commit hook runner
  • commitizen, a semantic version manager

Data Preparation

  1. Download the symbols of target stocks and put the .csv file containing the symbols of the target stocks into selected_symbol_path； Download the processed data put all .csv files containing stock datas into stock_path

  2. You can download the data of US stocks from the following website：Tiingo stock marke

  3. Select and process eligible U.S.S&P 500 stocks by running(The processing steps of the Chinese stock dataset are the same as the following):

python trials/preprocess/U.S.SP500-selected.py selected_symbol_path stock_path store_path begin_time end_time
  • selected_symbol_path, Stock symbols to be selected are stored in selected_symbol_path
  • stock_path, All stocks are stored in stock_path
  • store_path, Eligible U.S.S&P 500 stocks that have been screened and processed are stored in store_path
  • begin_time, The start time of the stocks to be screened. For example: '2000-01-01'
  • end_time, The end time of the stocks to be screened. For example: '2020-12-31'
  1. Randomly screen stocks from all stocks to form .csv of symbols, and generate as many .csv as possible with disjoint symbols. Then rolling.py can use these .csv to generate rolling datasets
python trials/preprocess/random_stocks_selected.py symbols_num_each_rolling, stock_data_path, random_symbol_path
  • symbols_num_each_rolling, The number of stocks contained in each rolling
  • stock_data_path, All U.S.S&P 500 stocks data screened by U.S.SP500-selected.py are stored in stock_data_path
  • random_symbol_path, The stocks symbols in each rolling are stored in random_symbol_path
  1. Form datasets for each period of each rolling in the form of .csv. by running:
python trials/preprocess/rolling.py stock_data_path store_path training_month validation_month testing_month random_symbol_path csv_name1 csv_name2 csv_name3
  • stock_data_path, All U.S.S&P 500 stocks data screened by U.S.SP500-selected.py are stored in stock_data_path
  • store_path, The final .csv files are stored in store_path
  • training_month, The number of months included in the training
  • validation_month, The number of months included in the validation
  • testing_month,The number of months included in the testing
  • random_symbol_path, The stock symbols included in each rolling that are randomly generated by random_stocks_selected.py in advance are stored in random_symbol_path
  1. We provide a subset of the processed dataset, please check it in /trials/data/.

How to Run?

  1. Single Run
poetry run trials/scripts/train_trials.py

See all parameters and their illustrations by poetry run trials/scripts/train_trials.py --help

  1. Hyperparameter Search We recommend to use hyperparamete search by wandb sweep.

Start with a template sweep.yaml to define searching parameters and values, then simply run:

wandb sweep rl_sweep.yaml

After showing

wandb: Run sweep agent with: wandb agent xxx/xxxxxxx/[sweep-id]

start the agents with sweep-id:

wandb agent xxx/xxxxxxx/[sweep-id]

Or you can start multiple agents with the script:

bash train_trials.sh [sweep-id] [num-process-per-gpu]

License

MIT

Feel free to cite our paper

@misc{han2023select,
      title={Select and Trade: Towards Unified Pair Trading with Hierarchical Reinforcement Learning}, 
      author={Weiguang Han and Boyi Zhang and Qianqian Xie and Min Peng and Yanzhao Lai and Jimin Huang},
      year={2023},
      eprint={2301.10724},
      archivePrefix={arXiv},
      primaryClass={q-fin.CP}
}

Main Results

S&P 500:

GGR Cointegration Correlation Wang TRIALS TRIALS wo TR
Subset 1 SR↑ -0.390 -1.557 -1.263 1.323 1.573 -0.075
ED↓ 0.008 0.036 0.010 0.036 0.027 0.011
MDD↑ -0.146 -0.407 -0.227 -0.111 -0.093 -0.325
AR↑ -0.031 -0.392 -0.141 0.320 0.512 -0.039
AV↓ 0.118 0.303 0.130 0.210 0.272 0.284
Asset Pair PG-TGT SPG_SWKS LMT-UNH SPG-SWKS RTX-WY ED-WY
Subset 2 SR↑ -2.005 -1.836 -1.576 0.735 2.099 -0.078
ED↓ 0.015 0.008 0.006 0.008 0.037 0.010
MDD↑ -0.334 -0.249 -0.223 -0.085 -0.112 -0.212
AR↑ -0.290 -0.201 -0.189 0.139 0.641 -0.012
AV↓ 0.174 0.129 0.140 0.167 0.240 0.192
Asset Pair AMGN-JPM BLK-FMC NEE-ROP BLK-FMC ALB-EQR AMGN-MS
Subset 3 SR↑ -1.492 -1.739 -1.297 1.700 1.862 0.021
ED↓ 0.022 0.044 0.005 0.005 0.029 0.008
MDD↑ -0.245 -0.739 -0.127 -0.103 -0.080 -0.202
AR↑ -0.203 -0.736 -0.096 0.362 0.344 0.005
AV↓ 0.158 0.650 0.091 0.179 0.153 0.203
Asset Pair KO-UPS CCL-PLD ADBE-MSFT CCL-PLD CCL-SCHW AES-PEAK
Subset 4 SR↑ -0.552 -1.665 -1.160 1.511 2.102 0.153
ED↓ 0.016 0.006 0.006 0.006 0.060 0.020
MDD↑ -0.096 -0.196 -0.164 0 -0.100 -0.331
AR↑ -0.061 -0.181 -0.087 0.104 0.666 0.029
AV↓ 0.136 0.128 0.093 0.050 0.247 0.244
Asset Pair A-CSCO ADM-GS HD-LOW ADM-GS INTC-NVDA ADM-FCX
Subset 5 SR↑ -2.418 -2.346 -1.737 0.611 1.564 0.349
ED↓ 0.010 0.009 0.007 0.007 0.033 0.009
MDD↑ -0.188 -0.267 -0.245 -0.136 -0.089 -0.176
AR↑ -0.162 -0.267 -0.199 0.128 0.336 0.072
AV↓ 0.080 0.137 0.135 0.193 0.182 0.187
Asset Pair VZ-WMT FAST-HUM AON-HON FAST-HUM AMZN-VZ FAST-TER
Mean SR↑ -1.371 -1.829 -1.406 1.176 1.840 0.074
ED↓ 0.014 0.021 0.007 0.021 0.037 0.011
MDD↑ -0.202 -0.372 -0.198 -0.087 -0.095 -0.249
AR↑ -0.149 -0.355 -0.142 0.210 0.500 0.011
AV↓ 0.133 0.269 0.118 0.160 0.219 0.222
STD SR↓ 0.793 0.274 0.215 0.430 0.238 0.161
ED↓ 0.005 0.016 0.002 0.016 0.012 0.004
MDD↓ 0.082 0.197 0.045 0.047 0.011 0.065
AR↓ 0.095 0.204 0.046 0.108 0.141 0.038
AV↓ 0.033 0.202 0.021 0.057 0.044 0.037

CSI 300:

GGR Cointegration Correlation Wang TRIALS TRIALS wo TR
Subset 1 SR↑ 0.021 -2.259 -1.000 0.110 1.558 1.491
ED↓ 0.007 0.007 0.011 0.007 0.020 0.014
MDD↑ -0.309 -0.316 -0.161 -0.164 -0.057 -0.037
AR↑ 0.004 -0.326 -0.094 0.024 0.269 0.149
AV↓ 0.210 0.177 0.114 0.192 0.146 0.081
Asset Pair 600653-600744 600808-822 568-858 600808-822 503-623 29-600740
Subset 2 SR↑ -1.368 -1.659 -1.609 0.104 1.210 0.910
ED↓ 0.016 0.007 0.031 0.007 0.032 0.009
MDD↑ -0.323 -0.304 -0.238 -0.093 -0.162 -0.153
AR↑ -0.221 -0.250 -0.196 0.025 0.375 0.179
AV↓ 0.185 0.177 0.143 0.171 0.277 0.174
Asset Pair 600091-600839 600103-600871 31-667 600103-600871 600109-600886 600219-600655
Subset 3 SR↑ -1.763 -0.728 -1.222 0.802 1.352 -0.154
ED↓ 0.007 0.018 0.009 0.018 0.054 0.015
MDD↑ -0.178 -0.173 -0.257 -0.158 -0.257 -0.202
AR↑ -0.178 -0.129 -0.196 0.208 0.625 -0.027
AV↓ 0.119 0.193 0.182 0.247 0.404 0.195
Asset Pair 600642-89 607-959 600690-600887 607-959 24-636 600633-600726
Subset 4 SR↑ -2.041 -2.741 -1.368 0.774 3.614 2.288
ED↓ 0.013 0.028 0.014 0.028 0.078 0.033
MDD↑ -0.324 -0.515 -0.361 -0.365 -0.091 0.000
AR↑ -0.321 -0.525 -0.295 0.278 1.676 0.315
AV↓ 0.191 0.266 0.248 0.383 0.278 0.113
Asset Pair 571-600190 518-600221 600741-786 518-600221 600809-825 400-939
Subset 5 SR↑ -0.793 -0.099 -1.661 1.962 1.798 0.213
ED↓ 0.024 0.024 0.011 0.024 0.048 0.024
MDD↑ -0.326 -0.135 -0.252 -0.130 -0.129 -0.221
AR↑ -0.159 -0.011 -0.258 0.649 0.442 0.042
AV↓ 0.216 0.176 0.183 0.261 0.203 0.297
Asset Pair 600215-921 600006-600115 46-600639 600006-600115 559-600074 553-600621
Mean SR↑ -1.189 -1.497 -1.372 0.750 1.906 0.950
ED↓ 0.013 0.017 0.015 0.017 0.046 0.019
MDD↑ -0.292 -0.289 -0.254 -0.182 -0.139 -0.122
AR↑ -0.175 -0.248 -0.208 0.237 0.677 0.132
AV↓ 0.184 0.198 0.174 0.251 0.261 0.172
STD SR↓ 0.736 0.969 0.245 0.678 0.877 0.877
ED↓ 0.006 0.008 0.008 0.008 0.020 0.009
MDD↓ 0.057 0.133 0.064 0.095 0.069 0.089
AR↓ 0.106 0.175 0.068 0.229 0.513 0.118
AV↓ 0.035 0.035 0.045 0.074 0.087 0.075

About

Our codebase trials provide an implementation of the Select and Trade paper, which proposes a new paradigm for pair trading using hierarchical reinforcement learning. It includes the code for the proposed method and experimental results on real-world stock data to demonstrate its effectiveness.

Topics

finance deep-reinforcement-learning pytorch fintech stock-market hierarchical-reinforcement-learning chinese-stock pair-trading algorithm-trading drl-trading-agents drl-algorithms drl-framework trading-tasks stable-baselines3 federated-reinforcement-learning

Resources

Readme

License

MIT license
Activity
Custom properties

Stars

111 stars

Watchers

3 watching

Forks

27 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages