This was one of my projects at Apteo last year. A lot of applications of machine learning in finance deal with generating predictions from financial data but we tried to take this concept further - using these predictions to create an agent for automated portfolio management. The idea was to simulate a human financial analyst who would look at these predictions to actually buy or sell stocks and maintain his/her portfolio.
The Stock-Market Environment
A simulated stock-market environment was created where the agent can learn to “play” and train itself. Such an environment would mimic the real stockmarket and had historical stock data of around 6500 stocks in it.
The Trading Agent
A simple feed-forward neural network was used to take the financial features as input and produce the optimal actions to take.
Note: Due to an NDA with Apteo, the code is not available
The agent displays the trading decision for each stock on each date of test set.
The agent training itself in the environment.