Execution algorithms are hierarchical decision frameworks designed to optimally plan and trade portfolios over a period of time. The planning process tracks a desired trade-off between transaction costs and uncertainty or risk, whereas the execution process implements the plan and focuses on efficiency.
In practice, the execution component of trading algorithms consists of complex heuristics or rules that encode the trading logic, which can become highly complex and difficult to optimize.
The topic of this presentation is to outline the application of reinforcement learning in trading algorithms to directly optimize down to the transaction level, over various problem configurations.
In this framework, the rules that govern the logic of a trading algorithm – historically written by humans, are instead encapsulated in the parameters of complex functions learned by machines.
This machine learned algorithm can outperform the hand-tuned algorithm in a variety of conditions to be discussed, and serves as a more integrated and efficient framework for the development of next generation of trading algorithms.