DARC-OR: Dual-Signal Adaptive Recalibration and Concept-Drift Response for Capacity-Constrained Operational-Risk Early Warning

Abstract

Operational-risk early-warning models in financial institutions operate under nonstationarity, extreme class imbalance, delayed outcome confirmation, and finite investigation capacity. This paper proposes DARC-OR, a drift-response framework that combines an immediately available Jensen-Shannon divergence signal on model scores with a delayed supervised EWMA signal on balanced log loss. The controller distinguishes short-lived distribution change from persistent or performance-relevant change, using recalibration for the former and recency-weighted adaptive-window retraining for the latter. A controlled 120-day transaction stream benchmark was executed for five independent seeds, totaling 720,000 transactions and 6,839 positive events, with abrupt, gradual, and recurring concept changes, a three-day label delay, and a 0.5% daily review capacity. Across the five runs, DARC-OR achieved average precision 0.1719 +/- 0.0280, exceeding a delayed-loss trigger by 21.55% and a static model by 38.22%; recall at capacity was 0.1423 +/- 0.0175 and Brier score was 0.008374 +/- 0.000305. However, its normalized exposure scenario cost (0.7221) was worse than both delayed-loss (0.6888) and static (0.6540) policies, and recurrence exposed substantial forgetting. The results therefore support dual-signal adaptation as a ranking-and-calibration improvement under drift, while showing that drift response, action cost, and recurring-concept memory must remain separately governed.