Model Risk for Acceptable, but Imperfect, Discrimination and Calibration in Basel PD and LGD Models

Henry Penikas

The Basel Internal-Ratings-Based (IRB) approach allows banks to use sufficiently good credit risk models for the daily computation of their capital adequacy ratio. However, being sufficiently good does not naturally mean being perfect. Conventionally, risk managers increase the mean probability of default (PD) and loss given default (LGD) values by some margin when developing a model. They expect that it is sufficient to offset for potential model risk. This add-on, thought to be conservative enough, gave rise to the term ‘conservative margin’. The novelty of this paper is that we are the first to identify the cases when such a margin is not sufficient. The principal cause is the previously ignored requirement to “freeze” capital against the existing loans. This capital tie-up does not allow reallocating excessive capital requirements from actual non-defaults (false negatives) to unforeseen defaults (false positives). This is the first time when such a novel cause of model risk is discussed. The value the paper creates is severalfold. First, the revealed model risk has a material scale and can be part of the bank’s explicit or implicit risk-taking strategy. Therefore, it should be considered by researchers, as well as by validators and supervisors. Second, the paper offers an indicator to expand the risk indicator dashboard suggested by the Basel Committee, when designing risk-reward remuneration. This is especially true of contracts with risk model developers.

Model Risk for Acceptable, but Imperfect, Discrimination and Calibration in Basel PD and LGD Models