Replication package: regime-conditional prediction-decision coupling in uncertainty-aware deep reinforcement learning for trading systems

Abstract

This record contains the complete replication material for a study of how strongly adistributional prediction model should be coupled to a sequential decision policy in low-signal,non-stationary markets. The manuscript evaluates fifteen coupling-mechanism configurations acrossthree walk-forward windows on five indices in three markets (CSI 300, CSI 500, S&P 500, Nasdaq,Hang Seng) under a statistical audit protocol frozen in advance. Contents scripts/ - 37 analysis scripts: environment and feature construction, theConv1d-LSTM-Transformer encoder, preference-conditioned SAC training, the fifteen mechanismconfigurations, walk-forward evaluation, the tercile block-bootstrap audit, the four switching-rulebacktests, matched-dial and matched-exposure controls, and the revision analyses (Benjamini-Hochbergand Holm multiplicity control, two-level day-by-seed inference, regime-definition robustness). data/per_day_returns/ - per-day log returns of the W0 test window for everymechanism and seed in all five markets (250 files, transaction costs included, no futureinformation). audit/ - every audit result file behind the tables and figures, including the270-cell tercile grid with q-values, the two-level bootstrap output, the regime partitions and theregime-definition robustness results. configs/configs.json - environment, encoder, SAC and mechanism hyperparameters, thefrozen protocol blocks and the walk-forward window definitions. PROTOCOL_CHRONOLOGY.md - dated protocol history: which rules and parameters werefixed before the results they evaluate and which analyses were added later. TABLE_MAP.md and README_REPRODUCE.md - the mapping from every table andfigure to the script and artefact that produce it, with the commands. checksums.sha256 - SHA-256 of every file (POSIX paths). The material is anonymized for double-anonymized peer review; the author list and the link to thepublished article will be added on acceptance. Code is released under the MIT licence, data anddocumentation under CC BY 4.0.