Evaluating credit assignment in modular financial reinforcement learning is challenging because an attributed signal may lower a component loss without providing decision‑relevant information or improving economic outcomes. We therefore propose an evaluation framework that combines equal‑information reconstruction controls, fixed‑utility comparisons, and predefined downstream guardrails. The framework is instantiated in an attribution‑guided modular trading architecture and validated through three controlled studies:
- Update allocation – residual‑based allocation is exactly reproduced by a standard loss‑budget control under the disclosed interface.
- Supervision target representation – direct shortfall supervision improves component prediction but fails to preserve decoded‑outcome consistency.
- Portfolio risk control – under fixed utility, component‑wise critic supervision worsens risk‑condition certainty‑equivalent and drawdown compared with scalar supervision.
These experiments reveal three evaluation failures: redundant attribution signals, prediction trade‑offs induced by representation, and misalignment between component supervision and downstream risk outcomes. A preserved historical engineering case further demonstrates that module‑level learning claims must be accompanied by run‑level provenance. The contribution is a reusable methodology for assessing credit assignment in expert systems operating under partial observability, delayed feedback, and economic risk constraints.
Review