Post-Hoc Macroeconomic Adjustment of IFRS 9 Probability of Default Estimates: An Exploratory Study of Norion Bank’s Swedish Unsecured-Loan Portfolio

Abstract

This thesis examines whether post-hoc macroeconomic adjustments can explain - and potentially reduce - the spread between Observed 12-month Default rates (OD12m) and the Probability of Default estimates-model (PD12m) used by Norion bank for unsecured consumer loans in Sweden. The analysis cover June 2019 to February 2024, and consists of individual contracts that are aggregated to monthlyaverages, a total of 56 months. We analyse the total portfolio in its entirety, as well as segmented on borrower characteristics like age, mortgage status, and co-borrower. A Linear Regression model using ordinary least square (OLS) is employed to estimate the spread, defined as the difference between realized OD12m and predicted PD12m. Correlation filtering, ElasticNet regularization, VIF screening, are used as feature selection techniques. Where autocorrelation is detected in the residuals, a Cochrane-Orcutt AR(1) correction is optionally introduced. The model is assessed using standard metrics like MAE, R2, adjusted R2, and residual diagnostics. For the total portfolio, three macro variables—a 24-month change in the tradeweighted krona (KIX), a 6-month change in household confidence, and a 12-month change in unemployment - explains the spread. While out-of-sample test errors remain low, MAE comparison reveals overfitting. Segment analysis shows that models for young borrowers, older borrowers, and non-mortgage holders capture the patterns of observed spread, whereas middle-aged borrowers, mortgage holders, and single borrowers exhibit severe overfitting and unstable macroeconomic relationships. The results indicate that linear macroeconomic overlays can add interpretive value and support PD12m-model monitoring, but they are not yet stable enough for direct IFRS 9 adjustments. Key limitations include the short time series, structural changes in the macroeconomic landscape around COVID-19 time, and potential macro leakage already embedded in borrower-level PD12m-models . Future research can explore non-linear models, rolling training windows, and direct macroeconomic integration into PD12m estimation.