[Admin] Meeting on 15.01.18

[mjuchli]

Progress

• Framework
  • Orders (incl. Types)
  • Order book
  • Match Engine
  • Action State
  • Action
  • Actions Space
• Reward functions tested with Q-Learning
  • Cumulative Reward
  • Profit on backtest

Next steps

• Increase Action State, e.g. more features
• Use policy gradient, e.g. recurrent neural nets
• Imitation Learning

Questions

• [x] Feedback on reward function
• [x] Feedback on evaluation techniques
• [x] Currently learns on random timestamp in order book. Would sequential learning and testing make sense?

[mjuchli]

Remarks:

• Add features from the future in order to artificially make profit with “cheating” (e.g. snooping).
• Is random sampling worse than sequential sampling?
• Do not interfere in reward function and leave as is