Mixtures of Experts for Scaling Up Neural Networks in Order Execution

We develop a methodology that employs mixture of experts to scale up the parameters of reinforcement learning (RL) models in optimal execution tasks. The innovation of our approach stems from a spectral clustering-driven methodology for customized training of RL agents. First, we cluster the data using domain-specific knowledge, creating distinct subsets for specialized learning. Second, for each cluster, we train individual customized RL experts. Third, we train a router layer which serves as a crucial branching mechanism for dynamically assigning the most appropriate expert for each step during an episode. We deploy and test this methodology on real data of the Amazon stock on NASDAQ, and provide evidence that our approach improves the profitability of execution tasks by approximately one basis point when compared to a single expert scenario.