Awesome AI Agents › Financial & Trading Systems

samre12/deep-trading-agent

⭐ 795 Python added to this list on 2025-04-19 repository created 2017-09-21

The Deep Trading Agent project is a sophisticated implementation of a trading agent for Bitcoin that leverages deep reinforcement learning techniques. The core of the system is based on the DeepSense network architecture, which is used for Q function approximation to enable the agent to make informed trading decisions. The project focuses on training an agent to maximize accumulated rewards by choosing among three possible actions: neutral, long, or short positions on Bitcoin trades. The agent learns from historical Bitcoin price data, which is processed and preprocessed to create a suitable dataset for training. The dataset consists of per-minute Bitcoin price series obtained from Coinbase exchange transactions, with preprocessing steps to handle missing values and create continuous blocks of data for effective learning. The implementation uses Python 2.7 and TensorFlow 1.1.0, along with libraries like Pandas for data preprocessing and tqdm for progress visualization. The project supports running in a Docker container, which simplifies setup and execution by providing a prebuilt environment with all dependencies and tools like Tensorboard for monitoring training progress. The Docker setup also includes scripts to pull the latest Bitcoin transaction data and prepare it for training. The model architecture is inspired by Deep Q-Trading, with modifications to incorporate the DeepSense network for better Q function approximation. The training process involves using historical price data, including closing, highest, lowest prices, and trading volume, normalized and structured as input channels to the model. The project aims to improve model performance by refining the reward function and state representation, including the use of exponentially weighted unrealized profit and loss (PnL) to stabilize learning. Overall, this project provides a comprehensive framework for developing and training a deep reinforcement learning-based trading agent for Bitcoin, combining advanced neural network architectures with practical data preprocessing and training tools to facilitate research and experimentation in algorithmic trading.

https://github.com/samre12/deep-trading-agent

agentalgorithmic-tradingartificial-intelligencebitcoinbitcoin-price-seriescoinbase-exchangedata-preprocessingdeep-q-networkdeep-q-tradingdeep-reinforcement-learningdeep-sensedeepsense-networkdockerpython-2.7q-function-approximationrecurrent-neural-networksreward-functiontensorboardtensorflowtradingtrading-agentunrealized-pnl