/Deep-Reinforcement-Learning-in-Trading

Primary LanguageJupyter Notebook

Deep Reinforcement Learning for Trading

This repository provides the code for a Reinforcement Learning trading agent with its trading environment that works with both simulated and historical market data. This was inspired by OpenAI Gym framework.

This repository has the Keras implementation of

Code for agents

Article about this project

gif

Requirements

  • Python 3.5/3.6
  • Keras
  • Tensorflow

Workflow

Architecture

Features

  • 3 Reinforcement learning Agents (DQN, DDQN, DDDQN)
  • ADX and RSI technical indicator and extensible for more
  • Historical stock market data ingestion through CSV

Policy

State

[𝐴𝐷𝑋(𝑑), 𝑅𝑆𝐼(𝑑), 𝐢𝐢𝐼(𝑑), π‘π‘œπ‘ π‘–π‘‘π‘–π‘œπ‘›, π‘’π‘›π‘Ÿπ‘’π‘Žπ‘™π‘–π‘§π‘’π‘‘ π‘Ÿπ‘’π‘‘π‘’π‘Ÿπ‘›]

Action

The agent could take three actions – Buy, Sell or Hold

Reward

The reward objective is set to maximize realized PnL from a round trip trade. It also includes

  • Trading commision (penalty)
  • Holding fee (penalty), like the interest brokers charge for overnight levereg position. These rewards are what controls and optimize the agents during the training phase that determines the trading behavior. The latter 2 parameters control the trading frequncy of the agent.

What's next?

  • Prioritized Experience Replay
  • LSTM networks
  • Asynchronous Advantage Actor-Critic (A3C)
  • Multi-agent
  • Reward engineering