Reading the paper 'Reinforcement Learning for FX trading 'at https://stanford.edu/class/msande448/2019/Final_reports/gr2.pdf it states:
While our end goal is to be able to make decisions on a universal time scale, in order to apply a reinforcement learning approach to this problem with rewards that do not occur at each step, we formulate the problem with a series of episodes. In each episode, which we designate to be one hour long, the agent will learn the make decisions to maximize the reward (return) in that episode, given the time series features we have.
This may be a question for the authors but is it not better in RL to apply rewards at each time step instead of "rewards that do not occur at each step"? If apply rewards at each time step then the RL algorithm will achieve better convergence properties as a result of learning at smaller time intervals rather than waiting for "one hour". Why not apply rewards at each time step?