Paper: arXiv 2610.03369

Authors: Alexander Ardaiz, Varun Budati, Ali Habibnia

Abstract

Deep reinforcement-learning policies for order execution can vary substantially across training seeds, so apparent architectural gains may reflect favourable training realisations rather than reproducible properties of the architecture. We evaluate vanilla Double Deep Q-Learning (DDQL), K-means-partitioned mixtures of DDQL experts at $K \in {2, 4, 8}$, and dense networks parameter-matched to the $K{=}4$ and $K{=}8$ expert budgets on 5-minute mean-aggregated BTC/USDT limit order book data from Binance. No learned configuration significantly improves mean implementation shortfall over DDQL. Under the reported specification, all have higher mean shortfall than TWAP (0.39 bps) and immediate liquidation (0.21 bps) in an environment whose frictionless replay and terminal-urgency penalty make early liquidation nearly costless; 11/100 vanilla-DDQL runs, versus none in either MoE $K{\geq}4$ arm, converge to a policy that waits until forced liquidation. We then decompose this specification on a device-matched baseline. Annealed exploration alone eliminates observed collapses (12/100 to 0/100; exact McNemar $p{=}4.9{\times}10^{-4}$), matching the elimination under expert partitioning. Combining annealed exploration with the aligned reward restores collapse in 19/30 runs; with all three specification changes, it rises to 48/100. In this environment, expert partitioning is unnecessary to suppress collapse and appears to mask a training-specification failure rather than confer an intrinsic performance benefit. No MoE $K{=}8$ run collapses under any of the six specifications tested. Across-seed dispersion is lowest at $K{=}8$ but non-monotone and not robust to family-wise adjustment, while within-policy tail risk worsens monotonically with $K$. The apparent attribution of the failure mode reverses between 30 and 100 seeds, illustrating the importance of repeated-seed evaluation.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates strong empirical rigor through extensive testing across multiple seeds and configurations, coupled with a novel approach to evaluating MoE stability. While the underlying math is standard RL, the statistical analysis of training stability adds a layer of complexity. The findings are significant for practical application in crypto order execution.

Research Flowchart

  flowchart TD
    A[Research Goal: Evaluate MoE for Crypto Order Execution Stability & Tail Risk] --> B{Methodology: Compare DDQL, MoE DDQL (K=2,4,8), & Parameter-Matched Dense Nets};
    B --> C[Data: 5-min BTC/USDT Limit Order Book (Binance)];
    C --> D[Computational Process: 100 Repeated-Seed Evaluations per Policy, Ablation Studies on Training Specs];
    D --> E[Key Finding 1: No MoE Architecture Significantly Improves Mean Shortfall Over DDQL; All Inferior to TWAP/Immediate Liquidation];
    E --> F[Key Finding 2: MoE (K>=4) Suppresses "Collapse" Policies, But Attributed to Masking Training Spec Failures (e.g., Unannealed Exploration)];
    F --> G[Key Finding 3: Within-Policy Tail Risk Worsens Monotonically with K; K=8 Best for Across-Seed Dispersion (But Not Robust)];