Research features and DQN architecture for small 2-action model

[karlie15]

DoD:

• have a model that converges
• outperforms random guesses during back-testing

[karlie15]

While data backfills:

• [ ] Create backtest eval framework (random guesses, rules based, etc)
• [ ] Create a sample set of tickers and option windows to test on model per ticker or one master model? leaning towards master model...