Model Risk Audit of a Barrier-Triggered Trading System: What Machine Learning Adds Under Coverage Adjustment

This paper proposes a diagnostic audit framework that extends perfect-foresight benchmarking. It places paired ex post oracles, for the original training label and the realized economic outcome, on a fixed population of trading opportunities with a common coverage-adjusted denominator. This construction was not identified within the scope of the search conducted. It is applied to a barrier-triggered FX contract (stop-loss 1.5, take-profit 2.0 times ATR, 336-bar time limit) on 15,476 opportunities across eleven major currency pairs, 2017–2022. The realized system’s win rate is 2.6439 percentage points below its breakeven threshold after cost, distinguishable from sampling noise across 192 tested settings of barriers, time limit and cost, narrowly at zero cost and with timed-out trades outside the threshold at short time limits. The ex post oracles reach a substantially higher value per opportunity, quantified rather than treated as a target for deployable models. A forward-evaluated model, tested on a later-starting subset of 10,268 opportunities, had negative value per opportunity under all three registered targets. The audit documents a measurable mismatch between this contract’s original training label and realized execution outcome. All results are diagnostic measurements within the training period. A registered confirmatory evaluation on held-out 2024 data has not been performed.