/Stock-Markets-Analytics-Zoomcamp-Project

Primary LanguagePython

Final Project for 'Stock-Markets-Analytics-Zoomcamp'

Introduction

High-frequency trading (HFT) is a complex financial task treated as a near real-time sequential decision problem. Traditional approaches often forecast equity trends and optimize weights via combinatorial optimization. However, these methods face challenges such as computational inefficiencies and limitations in handling a large number of assets with discrete action spaces.

An efficient DRL-based policy optimization (DRPO) method for HFT has been proposed to address these issues. This method models portfolio management as a Markov Decision Process, directly inferring equity weights to maximize accumulated returns. The environment is separated into "static" market states and "dynamic" portfolio weight states, simplifying agent interactions without losing interpretability. A reward expectation calculation algorithm using probabilistic dynamic programming enables agents to collect feedback without complex trajectory sampling.

Reference

  1. Han, L., Ding, N., Wang, G., Cheng, D., & Liang, Y. (2023, August). Efficient Continuous Space Policy Optimization for High-frequency Trading. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4112-4122). Source code

Dependencies

  • torch
  • numpy<2
  • pandas
  • python-dateutil
  • yfinance
  • tabulate
  • wandb (optional, for using validation data during training)

Environment Setting

Set up a Python virtual environment and install the dependencies:

pip install -r requirements.txt

Training Model and Simulation Trading

Training

  • Ensure your computer has an NVIDIA GPU to train your model.
  • You can modify train_config.json to customize your training configuration.
  • Run python train.py to start the training process.

Simulation Trading

  • Make sure you have at least one pre-trained model file in the model folder (I have provided mine).
  • You can modify trade_config.json to customize your trading configuration.
  • Run python main.py to simulate real-time trading.

Dataset

Market Num. of stocks Train Validation Test features
DJIA 30 30 2001-2021 2022 2023 open, close, high, low prices, true range ratio
  • OCHL prices are normalized by dividing them by the previous day's closing price and then taking the logarithm.
  • "DOW," "CRM," and "V" have been replaced by "XOM," "PFE," and "RTX" due to insufficient data.

Results

Start Date End Date CAGR Sharpe Ratio Maximum Drawdown Calmar Ratio
2023-01-03 2023-12-29 1.204 1.223 0.154 1.304