Contents:
-
What is Reinforcement Learning?
-
Choosing a policy
-
Calculating the value of an action
-
Practical consideration
-
“Gamification” of trading
-
How is the system trained (each game independent)?
-
Reward-function engineering
-
What features do we use for the neural network?
-
How to test the system?
-
What type of ANN should be used?
-
Demo and results
-
Lessons learned
-
RL can be very sample inefficient
-
Reward function design is hard
-
Rewards in trading are sparse
-
Local optima are difficult to escape
-
RL could just be overfitting peculiar chart patterns
-
Results are unstable and hard to reproduce
-
Why is it so hard?
-
Financial series are very noisy
-
Financial systems are dynamic - rules keep changing
-
Rules evolve by the very act of understanding them
-
Computing power is still limited
-
New algorithms are yet to be discovered
-
-
Noise = Unexplained returns
-
-
Future work - let the machine select the features
-
Questions & Answers