Umeå universitets logga

This thesis evaluates whether a Hidden Markov Model (HMM) can be used to obtain a compact representation of invoice payment behaviour for dynamic credit risk monitoring. The analysis is based on monthly payment histories from sole proprietorship credit accounts. In the first stage, the HMM summarises previous payment behaviour into filtered latent state probabilities. In the second stage, theseprobabilities are included as covariates in a flexible multi-state transition model for future delinquencyand default risk.

The HMM based model is evaluated out of time (OOT) against a richer benchmark model based on manually engineered observed-history variables. The evaluation focuses on one-step transition prediction, calibration, and propagated 3- and 6-month horizon risk. In addition, the study examines whether the HMM based model can be used for risk segmentation and whether portfolio group random effects for internal categories add predictive or diagnostic value.

The results show that the HMM based model does not outperform the richer observed history benchmark. In several evaluations the benchmark performs slightly better, while the HMM based model remainsclose in OOT transition risk, horizon risk and monitoring performance. The value of the HMM based representation is therefore that it provides a more compact and interpretable summary of payment history. The risk segmentation analysis further indicates that the fitted HMM based model score canconcentrate future adverse outcomes in highest risk groups, whereas the portfolio group random effect extension does not materially improve OOT performance.