A Stochastic Uncertainty-Propagation Framework for Forward-Looking Expected Credit Loss Under IFRS 9: An Empirical Illustration from an Emerging-Market Banking Portfolio

Abstract

IFRS 9 requires expected credit loss (ECL) measurement to incorporate past events, current conditions and reasonable and supportable forecasts of future economic conditions. This study develops a stochastic uncertainty-propagation framework that maps uncertainty in credit-risk inputs and economic scenarios into a distribution of model-implied portfolio ECL. The framework retains the 12-month/lifetime distinction across IFRS 9 impairment stages, represents default timing through marginal default probabilities, and combines scenario-specific credit-risk quantities using probability weights and discounting. The empirical illustration uses an aggregated Kenyan banking portfolio comprising ABF Retail, Check Off, Non-Check Off and Retail Secured exposures, with total exposure of KES 286,404 million and a deterministic ECL benchmark of KES 15,245 million. The computational experiment follows a transparent design using explicit lognormal uncertainty factors, scenario probabilities, dependence assumptions, a fixed random seed and an independently reproducible simulation protocol. A 100,000-run Monte Carlo experiment produces a mean model-implied ECL of KES 15,193.40 million, standard deviation of KES 1,398.67 million and a central 90% interval of KES 13,025.71–17,587.48 million. The mean is 0.34% below the deterministic benchmark. Sensitivity analysis shows that alternative scenario weights and dependence assumptions change dispersion and tail outcomes while leaving the central estimate relatively stable. The results demonstrate uncertainty propagation rather than predictive superiority; account-level longitudinal data are required for calibration, back-testing and comparison with alternative IFRS 9 models. The analysis also documents numerical convergence and scenario-conditional behavior, providing a transparent basis for interpreting central and tail outcomes. The study therefore contributes a reproducible methodological architecture while identifying the empirical requirements for subsequent validation.