Integrating External Credit Scores in Internal Credit Risk Models

Abstract

From both a legal and business perspective, it is important for credit institutes to accurately assess the risk of default for potential borrowers. This is traditionally done using statistical methods, such as logistic regression or decision trees, utilising sociodemographic and credit history information.This thesis investigates whether a bureau score, provided by a credit bureau, should be incorporated as an internal variable within the credit risk model, referred to as a model-in-model approach, or used as a standalone model and combined with an internal model through ensemble techniques. The two approaches are evaluated using logistic regression and XGBoost on a dataset of real loan applications from a specific market. The logistic regression models utilise Weight of Evidence transformed variables, and the XGBoost models use non-transformed features. Model performance is evaluated using the discriminatory power measures Gini and KS on 100 bootstrap samples. Both Gini and KS are consistently higher for the model-in-model approach across the bootstrap samples. However, the evidence is less conclusive as to whether this result generalises to future datasets.

Popular Abstract

The Medici and the Rothschilds are two
of the most powerful families to ever walk
among us, but what do they have in com-
mon? The answer is banking. Both dynasties built

their fortune through granting loans to the right
people. Today credit institutes evaluate potential
customers using Credit Scoring. Credit Scoring is
the application of statistical methods to estimate
risk. This begs for the philosophical question: what
is risk? Lucky for us, the Basel Committee on
Banking supervision has already pondered on this
question. In consumer credit risk it is defined as
default; default is said to have occurred when a
debtor is more than 90 days late with repayment.
To facilitate the task of estimating risk, banks can
use scores from credit bureaus (e.g., Upplysnings-
centralen) based on credit history information.
Maybe you have recently rented a car? In that
case, it is likely that the rental company used a
bureau score to asses your creditworthiness. This
thesis investigates how these scores should be in-
tegrated in internal credit risk models.

The bureau score can be incorporated in inter-
nal models using two architectures. The simplest
approach is to use the score inside the model as a
feature. This is referred to as the model-in-model
approach. The other architecture also use the bu-
reau score, but outside the internal model. The
internal model is built on all information except
bureau score. Then a model using only bureau
score is built. The output from these models are
then combined, this is referred to as the ensemble-
model approach.

In order to determine what model architectu-
re is preferable, we built four different models. Two
of the models used the industry standard: logistic
regression. This method can be viewed as a gene-
ralisation of the equation for a line y = kx + m to
probabilities (and more dimensions). That is, y is
bounded by 0 and 1 and we can use several inputs
x1, . . . , xn. The other two models were built using
the modern machine learning method XGBoost
(Extreme Gradient Boosting). This method uses
a myriad of small decision trees that are built
sequentially based on the error of the previous tree.
A decision tree can be viewed as a series of yes or
no questions that results in a final classification.
For example: Do you like ice cream? If yes, have
one! If no, eat a banana. For both model types,
both architectures were evaluated.

This project was carried out in close collaboration
with a Nordic consumer bank. The bank provided
us with real loan application data from one of
its markets. All models were built on a majority
subset of this data, and evaluated on the remaining
minority subset. But how can we be sure that the
results are not dependant on the subset we use to
build our model? We do the same thing again on
another subset, then we repeat this until we see
no point in wasting any more computing power.
In our case, 100 times.

How do we know that one architecture is better
than the other? We compare the two approaches
using two performance measures: Gini and KS.
Gini measures how well a model separates two
groups. KS measures the maximum separation
between the model implied distributions of two
groups. These two groups are the goods and the
bads, the goods are customers who have not entered
default within a year, the bads are those who have.

In both cases, logistic regression and XGBoost, the
model-in-model approach came out on top. The
former model-in-model even beat the ensemble-
model on all 100 subset variations when measured
by Gini. The evidence is clear for this dataset,
though we can not be sure the same holds for
others. Still, our gut feeling is that the model-
in-model approach would be preferable on future
datasets too.

So, do you want to build a banking dynasty of
Medici proportions? In that case, use the model-
in-model approach when you build your credit risk models...

• author: Karlsson, Fredrik and Malmberg, Simon
• supervisor: Magnus Wiktorsson
• organization: Mathematical Statistics
• course: FMSM01 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Credit Risk, Bureau Score, Scorecard, Default, Ensemble Model, Model-in-Model, Logistic Regression, XGBoost, Weight of Evidence, Feature Engineering, Feature Selection, Gini, KS
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3557-2026
• ISSN: 1404-6342
• other publication id: 2026:E60
• language: English
• id: 9233877
• date added to LUP: 2026-06-10 16:48:25
• date last changed: 2026-06-10 16:48:25