A reproducible protocol for auditing XAI-based trading policies: A case study on the S&P/BVL Peru General Index

Explainability methods increasingly inform real-time trading interventions, but claims of “behavioral correction” in this literature are rarely tested under the chronological separation, ablation, and multiplicity control that such claims require. We introduce a reproducible audit protocol — explicit train/validation/test roles with next-label embargoes, validation-only hyperparameter selection, close- t to open- t + 1 execution, an explainability-ablated baseline, a temporal permutation control, and bootstrap/multiplicity-corrected uncertainty — and apply it as a stress test to a previously reported CXAI-Guided Dynamic Stop-Loss (CGDSL) case study (3244 daily S&P/BVL Peru General Index observations, 2011–2024). Under this protocol, an originally reported “topological proof” of behavioral correction (Hazard Ratio sign change, 90.6% bias-mass reduction, 0% false-alarm rate) does not survive: adjusted tests are null under both within-seed and pooled 30-test multiplicity control, paired intervals reach zero, and two of three matched interventions produce counterfactual harm. Two post-review extensions probe the explainability component directly, with mixed results: an ATR-only ablation that removes the attribution gate achieves a more disciplined descriptive BDI than the full policy in two of three seeds, yet CGDSL’s net PnL exceeds the ablated baseline’s in all three seeds; and a temporal permutation of the attribution signal places the observed BDI at the 0.4th–3.8th percentile of an unadjusted null distribution, a diagnostic we report as exploratory rather than confirmatory. We argue the protocol’s value lies precisely in surfacing such mixed, unglamorous findings rather than only favorable ones, and offer it as a template other XAI-trading audits can apply before, not after, publication.