Improving the accuracy of credit scoring models using an innovative Bayesian informative prior specification method

Abstract

A new Bayesian informative prior specification method (BAF method – Bayesian priors using ARIMA forecasts) is proposed to introduce additional information into credit risk modelling and improve model predictive performance. We use logistic regression to model the probability of default of mortgage loans comparing the Bayesian approach with various priors and the frequentist approach. But unlike previous literature, we treat coefficient estimates in the probability of default model as stochastic time series variables. We build ARIMA models to forecast the coefficient values in future time periods and use these ARIMA forecasts as Bayesian informative priors. We find that the Bayesian models using this prior specification method produce more accurate predictions for the probability of default as compared to frequentist models and Bayesian models with other priors.

Keywords:

• Autoregressive time series
• Bayesian analysis
• probability of default
• credit risk

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

http://www.freddiemac.com/research/datasets/sf_loanlevel_dataset.page.

Notes

1 This is legitimate firstly because the accounts in the test sample and those used for prior specification are originated in different years, therefore the accounts used for forecasting are out of sample. Secondly by the time the test data is available, all information relating to accounts originated prior to that is available.

2 Estimation results are shown in Appendix D.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [grant number 72101087] and the Chinese Scholarship Council.