Bayesian Logistic Regression Models for Credit Scoring by Gregg Webster

Chapter 5: Conclusions and Implications 
 
5.1 Summary 
This study provided an investigation into the use of Bayesian logistic regression models 
for credit scoring. The main aim was to determine whether the use of relevant prior 
information was useful for a financial institution when it was having data quantity issues.
The first step of the study was to review existing literature. It was found that there are a 
number of models which can be used to build a credit scoring model - there is, however, 
no “best” model. Bayesian logistic regression models with relevant prior information were 
shown to provide an improvement over other models when the amount of data available to 
train the model is small. The literature review was then followed by a theory section. This 
section provided the theory which was used in the study.
In the results chapter, models were fitted to the data. The data set was randomly split into 
four sets, 
viz
. the “old”, “new”, validation and test data sets. For a financial institution 
expanding into a new economic location, the “old” data were assumed to come from the 
home location, the “new” data from the new location, the validation data from the old 
location and the test data from the new location. The financial institution was looking to 
build a scoring model in the new economic location with a limited amount of data. A 
logistic regression model was fitted on the “old” data and the parameters from this model 
were used as prior information for a Bayesian logistic regression model with informative 
prior in the “new” data. A logistic regression and Bayesian logistic regression with non-
informative prior was also fitted on the “new” data. Using two different cut-off 
probabilities for classification, it was found that on the test set, the Bayesian logistic 
regression model with informative prior provided a lower total error rate.
The error rates of the models were also compared when there are different amounts of 
“new” data available to build the model. It was found that the Bayesian logistic regression 
with informative prior provided relatively constant error rates. The logistic regression 
model and Bayesian logistic regression with non-informative prior had error rates which 

---

105 
started high when the amount of “new” data was small, and subsequently as the size of the 
“new” data increased, these error rates decreased and levelled off. The Bayesian logistic 
regression with non-informative prior gave lower error rates than the logistic regression 
model. The pattern of the error rates as the amount of data increased for these two models 
was, however, similar.
The use of prior information is very useful when a financial institution is expanding into a 
new economic location and at first has limited data available. The usefulness of using 
relevant prior information decreases as the amount of data available increases.