Online Calibration of Account-Specific Risk Sensitivity from Local Behavioural Distances in Streaming Banking Transactions

Abstract Risk evaluation in digital banking can be affected by differences in the behavioural regularity of individual accounts. A single sensitivity parameter may respond differently to local behavioural distances across accounts. This prompted us to examine an online calibration procedure in which an account-specific risk sensitivity parameter is estimated from transaction-level nearest-neighbour distances within a local behavioural reference pool. During our exploratory analysis of IEEE-CIS fraud detection benchmark, 608 accounts (7.15%) were identified as eligible under criteria of at least 20 transactions and 30 days of history. Transactions were then represented through four behavioural deviation dimensions: recency, temporal, amount, and structural. Local behavioural distances of those deviation dimensions were used both for risk evaluation and for incremental calibration. In our experiment, a running mean was maintained from legitimate transaction observations without assuming a parametric distribution. Calibrated sensitivity varied substantially across behavioural strata, with mean values ranging from 0.7113 for high-activity accounts to 1.2534 for low-activity accounts, compared with the fixed reference of one. We also observed, at the transaction level, that an account with sensitivity of 0.6819 produced an adaptive risk of 0.7973 compared with 0.7284 under the fixed reference, while sensitivity above one reduced the corresponding risk. Interestingly, these results show that online calibration of risk sensitivity alters continuous risk while leaving the underlying local distances unchanged. The procedure provides a constant-memory mechanism for adapting risk sensitivity to account-specific behavioural regularity in streaming banking environments.