CAIS++ Spring 2019 Project: Building an Agent to Trade with Reinforcement Learning

• February 3rd:

  • In meeting:
    • First Meeting, Environment Set-Up, DQN Explanation, Project Planning
  • Homework:
    • Read first three chapter of Spinning Up
    • Watch the first half of Stanford RL Lecture
    • Code up your own LSTM on the AAPL data. Check each other's out for inspiration, find online resources, ask questions in the slack. Should be a useful exercise for everyone!
    • (Optional: Watch LSTM Intro Video)

• February 10th: Working LSTM Model

  • State:
    • Current Stock Price
    • percent change from (n-1) to n open
  • Action Space: Buy, Sell, Hold (3-dimensional continuous) as percentage of bankroll
  • Reward:
    • Life Span (define maximum length agent can interact with environment)
      • Receives reward based on profit/loss at the end
      • Sparse reward, harder to train
  • Punishment
    • Test extra severe punishment for losses
    • set thresh-hold time before it can trade again based on punishment
  • Architecture of model
    • One day of encoding LSTM
      • Observation, no actions taken
    • Second day of Policy Net
      • Based on methodology learned from encoding
      • Actions are taken
    • Two-day batches
  • Model Dimensions
    • Encoding LSTM
      • #layers of LSTM, #layers of FCs
      • input size, hidden units size, encoding vector size
    • Policy LSTM
      • input size (state space size)
      • output size (action space size: 3d continuous)
  • Homework:
    • Jincheng and Yang: Begin building Encoding / Policy Net Models
    • Chris: Look through Andrew's current LSTM model
    • Grant: Do the preprocessing data
    • Tomas: - Continue working on RL architecture - Make graph of prices + volume over batch - Visualize price gaps
  • Pre-Process Data
  • Visualization
  • Gym Trading Environment
  • Integrate LSTM into DDDQN

• February 18th: Working DQN

  • Done for homework
    • Built first policy gradient model (Jincheng)
    • Worked on data pre-processing (Tomas)
  • Today's plan
    • Data pre-processing
    • Use data as input into the gym
    • Finalize the model

• February 24th: Work day

  • Finish pre-processing
  • Finish trading gym
    • simulate.py
      • Change action 'quit' to quit when timeseries ends
      • change time series to remove seconds
    • series_env.py
      • in class seriesenv
        • do not need daily_start_time, daily_end_time
        • remove randomization of start index (in def 'seed')
  • Finish pipelining <<<<<<< HEAD

• February 28th: Hackathon

  • TODO
    • Review the current reward function in series_env
    • Finish building the dataset
    • Merge dataset with environment and test
    • Begin building the model
    • Create sine wave csv for testing

• March 3rd: Working Actor-Critic Model

• March 10th: Add details like trading costs, slippage, and ask-bid spread; compute performance statistics; data visualization

=======

• March 3rd:
  • Work on implementing LSTM (Chris & Caroline)
  • Create test datasets (Grant)
  • Integrate dataset with gym (Yang & Jincheng)
• March 31th:
  • Finish LSTM

• Working Actor-Critic Model
• Add details like trading costs, slippage, and ask-bid spread; compute performance statistics; data visualization

13bd296a9295db93c71581f958b043ea734bdeda

• Build back testing environment
• Integrate NLP sentiment analysis as feature
• Add more indicators to model
• Clean up README
• Do we hold positions overnight? I think initially no. There are also weird jumps over holidays and weekends.
• Take into account high, low, close, volume data

• OpenAI Spinning Up
• UCL + DeepMind Lecture Series
• Stanford CS234: Reinforcement Learning
• CS234: Reinforcement Learning | Winter 2019
• Deep-rl-bootcamp
• David Silver:Intro Reinforcement Learning
• Stanford CS330: Deep Multi-Task and Meta Learning

• Financial Trading as a Game: A Deep Reinforcement Learning Approach
• Practical Deep Reinforcement Learning Approach for Stock Trading

• Model-based Deep Reinforcement Learning for Dynamic Portfolio Optimization

• Learning to trade with RL