SciPy Proceedings
article
Jul 13, 2026
Avoiding Zero-Trade Policies in RL with a Decoupled MLOps Architecture
A direct Stable-Baselines3 DQN trading USD/JPY 5-minute bars completed zero trades on the held-out period. We describe a decoupled architecture in which the RL model emits a directional score and a deterministic Python execution engine owns position management. On the same data and split, the decoupled system completed 2,625 fully risk-managed trades.
Sho Tanaka
doi
none
For Editors
submission
Build Status
COMPLETED
Submission Status
PENDING
Timestamp
Oct 09, 2026 05:34:53
Submission
scipy-2026-sho_tanaka
Storage
01a11f28-3321-721b-ab7c-a2773cbde894
https://prv.curvenote.com/
private CDN, long term retention policy
Checks

29 checks passed

5 optional checks failed

DOI Exists (optional)
No DOI available for sutton2018 (skipped)
No DOI available for moody1998 (skipped)
No DOI available for gymnasium (skipped)
No DOI available for stable_baselines3 (skipped)
No DOI available for talib (skipped)