The task is deliberately concrete: decide each day whether to be long, flat or short on Bitcoin and Tesla, using news alongside historical market data. Framing it as a discrete-action Markov decision process is what makes reinforcement learning applicable, since the agent picks one of three actions and lives with the consequence.

The paper compares four algorithms, Policy Gradient, PPO, Deep Q-Learning and DDPG, on the same problem. That comparison is the useful part. Financial results are notoriously easy to make look good by choosing the period, so putting several methods through an identical setup says more than any single impressive backtest.

The sentiment augmentation reflects a plain observation about these two assets in particular: their prices respond to what is being said about them, not only to what has already happened to them. Whether a model can extract that reliably from news is exactly the open question.