Expected and unexpected losses: credit risk loss distribution

What it is about

Conceptual and methodological foundations for measuring credit risk, distinguishing expected losses, those anticipated under normal conditions, from unexpected losses, associated with adverse scenarios and extreme events.

It also analyzes the statistical distribution of losses in credit portfolios, with concepts such as probability of default, loss severity and credit conversion factors, their relevance for integrated risk management and the solvency of institutions, and the challenges of implementing them properly.

Session summary

The speaker, from the Superintendency of Banking, Insurance and Pension Funds of Peru, explains that credit risk remains the main risk in the region’s banking systems and that its portfolios are highly heterogeneous, so each type of loan and each segment needs its own model. He shows why provisioning tables based on days past due are neither forward-looking nor risk-sensitive, and why an expected-loss model fits what actually happens better. He presents the loss distribution as a single curve: the mean is expected loss, covered by provisions, and the tail is unexpected loss, covered by capital. Expected loss is the product of the probability of default, the loss severity and the exposure at default.

He describes how the parameters were calibrated with local data: internal ratings for large companies, statistical models for retail credit, a recovery cash-flow methodology for severity, and conversion factors for undrawn lines. The challenges he identifies are data quality, above all for severity, human capital, the need to use local information rather than regional averages, supervisor training, and the different incentives of supervisors and institutions. The questions covered the rate used to discount recovery flows, systems with poor data, how to recognize collateral, how to segment a portfolio and the use of generative artificial intelligence in models.

Topics covered

• Portfolio heterogeneity and the need to segment
• Limits of days-past-due provisioning compared with expected-loss models
• Expected and unexpected loss in a single distribution: provisions and capital
• Probability of default, loss severity, exposure and conversion factors
• Calibration with local data and the supervisor’s internal ratings
• Challenges: data, human capital, supervisor training and incentives
• Questions: discount rate, collateral, segmentation and artificial intelligence