The Dynamics of Operational Loss Clustering

This paper investigates the characteristics of the operational loss data formation mechanism that takes place between the date of discovery of a new operational risk event and the final settlement date on which all losses are materialized. The first loss that characterizes the initial impact of a new operational risk event frequently triggers a sequence of related losses. Then, losses generated by the same event are not independent and follow a predictable scheme and the frequency of secondary losses is not homogeneous: both are functions of the initial loss amount and time. We model the arrival intensity and loss severities with a shot-noise stochastic process and derive its key properties. We then discuss implications of our model for the estimation of the regulatory capital charge for operational risk. In an empirical analysis, we find strong evidence of a shot-noise behavior in operational losses using the data of a major U.S. commercial bank