Question
How does a reinforcement-learning agent behave when its actions change a simulated portfolio and its reward depends on the resulting returns?
Approach
The experiments use proximal policy optimization with recurrent actor and critic networks. Stock and cryptocurrency data pipelines feed trading environments, training runs, and backtests.
Recurrent state gives the policy a way to use observations from earlier steps. The reward definition determines how the agent balances gains, losses, and action choices.
Limits
The project records sparse positive rewards, limited exploration, and difficulty balancing the reward terms. These issues make it hard to separate learning progress from choices in the environment.
Backtest output alone would not establish that a strategy transfers to live trading. This work concerns the behavior of the learning setup; it does not establish a profitable trading strategy.