This Reinforcement Learning Stock Trader uses a mix of human trading logic and Q-Learning to trade Equities found on Yahoo.com/finance in your terminal!
It works by running defined trading logic for a set of historical trades, and then hands over the torch to Q-Learning for the remaining set of historical data.
To use:
- Make sure all necessary libraries are installed: Numpy, Pandas, Scipy
- Clone this project into any directory on your machine
- CD into the project directory
- Pass the following commmand:
RL-Trader.py [EQUITY] [START DATE - DAY/MONTH/YEAR] [STARTING PORTFOLIO VALUE] [DAYS TO LEARN]
For example: RL-Trader.py F 1/1/2000 1000 100
^ This command runs the RL script against Ford's historical data and learns using our trading logic (under logic/logic.py) for 100 days before Reinforcement Learning kicks in with a starting portfolio of $1,000. The resulting Q-Table, as well as the profit, is then printed.
I have found that this script works especially well against times of economic contraction. That being said, results are contingent on the trading logic given to the RL agent, as well as the attributes of the RL agent itself.
Unlike other Reinforcement Learning scripts, it is better to keep the greedy factor (Epsilon) low (around .05-.5) as it increases the amount of analytical decisions the script makes. Because markets have a stochastic factor, it did not make sense to have the script choose a random 'buy' or 'sell' call, but instead use logic an analyist might use (this is under state_logic), only maximizing when there is enough data in the Q-Table (analagous to traders using trading strategies that have worked before).
- Calculation for Alpha compared to given Equity/Market
- Ability to add more than one equity
If you'd like to see anything added -- feel free to message me: krolo@wisc.edu