Mixture-of-Experts for Cryptocurrency Order Execution: Training Stability, Tail Risk, and Failure Modes

Battle Count: 51

Highly relevant for practitioners implementing RL-based execution algorithms. It highlights critical pitfalls in training stability, the importance of repeated-seed evaluation, and the danger of attributing performance gains to architecture when they may stem from training specification artifacts.

Moderate complexity. Requires implementing DDQL, K-Means clustering, and a custom LOB replay simulator. The core logic is standard, but the rigorous evaluation protocol (100 seeds, paired comparisons) adds computational overhead.

The paper provides detailed pseudocode for the fill mechanics, specific hyperparameters (Table 1), and states that code and configurations will be released on publication. It explicitly discusses device-dependent reproducibility issues (CPU vs CUDA) and provides a device-matched baseline for decomposition.