Mixture-of-Experts for Cryptocurrency Order Execution: Training Stability, Tail Risk, and Failure Modes
The study tested a Mixture-of-Experts (MoE) approach combining K-Means partitioning with Double Deep QLearning experts for cryptocurrency order execution. While this architecture did not improve mean execution cost over standard DDQL, it showed that training specifications drastically affect stability and tail risk. The findings suggest architectural gains are often artifacts of favorable training setups rather than intrinsic policy improvements.