SSRN Id3553781

Machine Learning or Econometrics for Credit Scoring:
Let’s Get the Best of Both Worlds∗

Elena, Dumitrescu†, Sullivan, Hué‡, Christophe, Hurlin§, Sessi, Tokpavi¶
January 15, 2021
Abstract
In the context of credit scoring, ensemble methods based on decision trees, such
as the random forest method, provide better classification performance than standard
logistic regression models. However, logistic regression remains the benchmark in the
credit risk industry mainly because the lack of interpretability of ensemble methods
is incompatible with the requirements of financial regulators. In this paper, we pro-
pose to obtain the best of both worlds by introducing a high-performance and inter-
pretable credit scoring method called penalised logistic tree regression (PLTR), which
uses information from decision trees to improve the performance of logistic regression.
Formally, rules extracted from various short-depth decision trees built with pairs of
predictive variables are used as predictors in a penalised logistic regression model.
PLTR allows us to capture non-linear effects that can arise in credit scoring data
while preserving the intrinsic interpretability of the logistic regression model. Monte
Carlo simulations and empirical applications using four real credit default datasets
show that PLTR predicts credit risk significantly more accurately than logistic regres-
sion and compares competitively to the random forest method.
JEL Classification: G10 C25, C53

Keywords: Risk management; Credit scoring; Machine learning; Interpretability; Econo-
metrics.
∗
This paper has previously circulated under the title “Machine Learning for Credit Scoring: Improving
Logistic Regression with Non-Linear Decision-Tree Effects”. We are grateful to Emanuele Borgonovo (the
editor), two anonymous referees, Jérémy Leymarie, Thomas Raffinot, Benjamin Peeters, Alexandre Girard,
and Yannick Lucotte. We also thank the seminar participants at University of Orléans as well as the
participants of the 16th Conference “Développements Récents de l’Econométrie Appliquée la Finance”
(Université Paris Nanterre), 7th PhD Student Conference in International Macroeconomics and Financial
Econometrics, 35th Annual Conference of the French Finance Association and International Association
for Applied Econometrics for their comments. Finally, we thank the ANR programs MultiRisk (ANR-16-
CE26-0015-01), CaliBank (ANR-19-CE26-0002-02), and the Chair ACPR/Risk Foundation: Regulation and
Systemic Risk for supporting our research.
†
EconomiX-CNRS, University of Paris Nanterre, 200 Avenue de la République, 92000 Nanterre, France.
E-mail: elena.dumitrescu@parisnanterre.fr
‡
Corresponding author, Univ. Orléans, CNRS, LEO (FRE 2014), Rue de Blois, 45067 Orléans. E-mail:
sullivan.hue@univ-orleans.fr
§
Univ. Orléans, CNRS, LEO (FRE 2014), Rue de Blois, 45067 Orléans. E-mail: christophe.hurlin@univ-
orleans.fr
¶
Univ. Orléans, CNRS, LEO (FRE 2014), Rue de Blois, 45067 Orléans. E-mail: sessi.tokpavi@univ-
orleans.fr
1
Electronic copy available at: https://ssrn.com/abstract=3553781
1 Introduction
Many authors have highlighted the extraordinary power of machine learning allied with
economists’ knowledge base to address real-world business and policy problems. See, for
instance, Varian (2014), Mullainathan and Spiess (2017), Athey (2018), Charpentier et al.
(2018), and Athey and Imbens (2019). In this article, we propose to combine the best of
both worlds, namely, econometrics and machine learning, within the specific context of credit
scoring.1 Our objective is to improve the predictive power of logistic regression models via
feature engineering based on machine learning classifiers and penalisation techniques while
keeping the model easily interpretable. Thus, our approach aims to propose a credit scoring
methodology that avoids the perennial trade-off between interpretability and forecasting
performance.
The use of econometric models for credit scoring dates back to the 1960s, when the
credit card business arose and an automatised decision process was required.2 After a
period of rather slow acceptance, credit scoring had, by the 1970s, become widely used by
most banks and other lenders, and various econometric models were considered, including
discriminant analysis (Altman, 1968), proportional hazard (Stepanova and Thomas, 2001),
probit or logistic regression models (Steenackers and Goovaerts, 1989), among many others.
Then, logistic regression gradually became the standard scoring model in the credit industry,
mainly because of its simplicity and intrinsic interpretability. Most international banks still
use this statistical model, especially for regulatory scores used to estimate the probability of
default for capital requirements (Basel III) or for point-in-time estimates of expected credit
losses (IFRS9).
Credit scoring was one of the first fields of application of machine learning techniques
in economics. Some early examples are decision trees (Makowski, 1985; Coffman, 1986;
Srinivasan and Kim, 1987), k -nearest neighbours (Henley and Hand, 1996, 1997), neural
networks (NN) (Tam and Kiang, 1992; Desai et al., 1996; West, 2000; Yobas et al., 2000),
and support vector machines (SVMs) (Baesens et al., 2003). At that time, the accuracy
gains (compared to the standard logistic regression model) for creditworthiness assessment
appeared to be limited (see the early surveys of Thomas, 2000 and Baesens et al., 2003).
However, the performance of machine learning-based scoring models has been improved
substantially since the adoption of ensemble (aggregation) methods, especially bagging and
boosting methods (Finlay, 2011; Paleologo et al., 2010; Lessmann et al., 2015).3 In their
1
Broadly defined, credit scoring is a statistical analysis that quantifies the level of credit risk associated
with a prospective or current borrower.
2
In a working paper of the National Bureau of Economic Research (NBER), Durand (1941) was the first
to mention that the discriminant analysis technique, invented by Fisher in 1936, could be used to separate
entities with good and bad credit.
3
The ensemble or aggregation methods aim to improve the predictive performance of a given statistical
or machine learning algorithm (weak learner) by using a linear combination (through averaging or majority
vote) of predictions from many variants of this algorithm rather than a single prediction.
2
extensive benchmarking study, Lessmann et al. (2015) compared 41 algorithms with various
assessment criteria and several credit scoring datasets. They confirmed that the random
forest method, i.e., the randomised version of bagged decision trees (Breiman, 2001), largely
outperforms logistic regression and has progressively become one of the standard models in
the credit scoring industry (Grennepois et al., 2018). Over the last decade, machine learning
techniques have been increasingly used by banks and fintechs as challenger models (ACPR,
2020) or sometimes for credit production, generally associated with “new” data (social or
communication networks, digital footprint, etc.) and/or “big data” (Hurlin and Pérignon,
2019).4
However, one of the main limitations of machine learning methods in the credit scor-
ing industry comes from their lack of explainability and interpretability.5 Most of these
algorithms, in particular ensemble methods, are considered “black boxes” in the sense that
the corresponding scorecards and credit approval process cannot be easily explained to
customers and regulators. This is consistent with financial regulators’ current concerns
about the governance of AI and the need for interpretability, especially in the credit scoring
industry. See, for instance, the recent reports on this topic published by the French regu-
latory supervisor (ACPR, 2020), the Bank of England (Bracke et al., 2019), the European
Commission (EC, 2020), and the European Banking Authority (EBA, 2020), among many
others.
Within this context, we propose a hybrid approach called the penalised logistic tree
regression model (hereafter PLTR). PLTR aims to improve the predictive performance of
the logistic regression model through data pre-processing and feature engineering based on
short-depth decision trees and a penalised estimation method while preserving the intrinsic
interpretability of the scoring model. Formally, PLTR consists of a simple logistic regression
model (econometric model side) including predictors extracted from decision trees (machine
learning side). These predictors are binary rules (leaves) outputted by short-depth decision
trees built with pairs of original predictive variables. To handle a possibly large number of
decision-tree rules, we incorporate variable selection in the estimation through an adaptive
lasso logistic regression model (Zou, 2006; Friedman et al., 2010), i.e., a penalised version
of classic logistic regression.
The PLTR model has several advantages. First, it allows us to capture non-linear ef-
fects (i.e., thresholds and interactions between the features) that can arise in credit scoring
data. It is recognised that ensemble methods consistently outperform logistic regression be-
cause the latter fails to fit these non-linear effects. For instance, the random forest method
benefits from the recursive partitioning underlying decision trees and hence, by design, ac-
4
See Óskarsdóttir et al. (2019) or Frost et al. (2019) for a general discussion about the value of big data
for credit scoring. In the present article, we limit ourselves to the use of machine learning algorithms with
“traditional data” for credit risk analysis.
5
We do not distinguish explainability from interpretability.
3
commodates unobserved multivariate threshold effects. The notable aspect of our approach
consists of using these algorithms to pre-treat the predictors instead of modelling the de-
fault probability directly with machine learning classification algorithms. Second, PLTR
provides parsimonious and interpretable scoring rules (e.g., marginal effects or scorecards)
as recommended by the regulators, since it preserves the intrinsic interpretability of the
logistic regression model and is based on a simple feature selection method.
In this article, we propose several Monte Carlo experiments to illustrate the inability of
standard parametric models, i.e., standard logistic regression models with linear specifica-
tion of the index or with quadratic and interaction terms, to capture well the non-linear
effects (thresholds and interactions) that can arise in credit scoring data. Furthermore,
these simulations allow us to evaluate the relative performance of PLTR in the presence of
non-linear effects while controlling for the number of predictors. We show that PLTR outper-
forms standard logistic regression in terms of out-of-sample forecasting accuracy. Moreover,
it compares competitively to the random forest method while providing an interpretable
scoring function. We apply PLTR and six other benchmark credit scoring methodologies
(random forest, linear logistic regression, non-linear logistic regression, non-linear logistic
regression with adaptive lasso, an SVM and an NN) on four real credit scoring datasets.
The empirical results confirm those obtained through simulations, as PLTR yields good
forecasting performance for all the datasets. This conclusion is robust to the various pre-
dictive accuracy indicators considered by Lessmann et al. (2015) and to several diagnostic
tests. Finally, we show that PLTR also leads to more cost reductions than alternative credit
scoring models.
Our paper contributes to the literature on credit scoring on various issues. First, our
approach avoids the traditional trade-off between interpretability and forecasting perfor-
mance. We propose here to restrict the intrinsic complexity of credit-score models rather
than apply ex post interpretability methods to analyse the scoring model after training. In-
deed, many model-agnostic methods have been recently proposed to make the “black box”
machine learning models explainable and/or their decisions interpretable (see Molnar, 2019
for an overview). We can cite here among many others the partial dependence (PDP) or
individual conditional expectation (ICE) plots, global or local (such as the LIME) surrogate
models, feature interaction, Shapley values, Shapley additive explanations (SHAPE), etc.
In the context of credit scoring models, Bracke et al. (2019) and Grennepois and Robin
(2019) promoted the use of Shapley values.6 Bussman et al. (2019) recently proposed an
explainable machine learning model specifically designed for credit risk management. Their
model applies similarity networks to Shapley values so that the predictions are grouped
according to the similarity in the underlying explanatory variables. However, obtaining
6
This method assumes that each feature of an individual is a player in a game where its predictive
abilities determine the pay-out of each feature (Lundberg and Lee, 2017).
4
the Shapley values requires considerable computing time because the number of coalitions
grows exponentially with the number of predictive variables, and computational shortcuts
provide only approximate and unstable solutions. An alternative approach is the InTrees
method proposed by Deng (2019). That algorithm extracts, measures, prunes, selects, and
summarises rules from a tree ensemble and calculates frequent variable interactions. This
helps detect simple decision rules from the forest that are important in predicting the out-
come variable. Nevertheless, the algorithms underlying the extraction of these rules are not
easy to disclose. Finally, our contribution can also be related to the methods designed to
enable NNs and SVMs to be interpretable, especially the TREPAN (Thomas et al., 2017),
Re-RX (Setiono et al., 2008), or ALBA (Martens et al., 2008) algorithms. However, there
is a slight difference between these approaches and ours. While the latter aim to enable a
model (i.e., NNs or SVMs) to be explainable/interpretable, PLTR aims to improve the pre-
dictive performance of a simple model (i.e., the logistic regression model) that is inherently
interpretable.
Second, our approach can be viewed as a systematisation of common practices in the

credit industry, where standard logistic regression is still the standard scoring model, es-
pecially for regulatory purposes. Indeed, credit risk modellers usually introduce non-linear
effects in logistic regression by using ad hoc or heuristic pre-treatments and feature engi-
neering methods (Hurlin and Pérignon, 2019) such as discretisation of continuous variables,
merger of categories, and identification of non-linear effects with cross-product variables.
In contrast, we propose here a systematic and automatic approach for modelling such un-
observed non-linear effects based on short-depth decision trees. Thus, PLTR may allow
model developers to significantly reduce the time spent on data management and data
pre-processing steps.
More generally, our paper complements the literature devoted to hybrid classification
algorithms. The PLTR model differs from the so-called logit-tree models, i.e., trees that
contain logistic regressions at the leaf nodes such as the logistic tree with unbiased selection
(LOTUS) in Chan and Loh (2004) and the logistic model tree (LMT) in Landwehr et al.
(2005). Although similar in spirit, our PLTR method also contrasts with the hybrid CART-
logit model of Cardell and Steinberg (1998). Indeed, to introduce multivariate threshold
effects in logistic regression, Cardell and Steinberg (1998) used a single non-pruned decision
tree. However, the large depth of this unique tree complicates interpretability and may
lead to predictor inflation that is not controlled for (e.g., through penalisation, as in our
case). PLTR also shares similarities with the two-step classification algorithm recently
proposed by De Caigny et al. (2018) in the context of customer churn prediction. Their
initial analysis consisted of applying a decision tree to split customers into homogeneous
segments corresponding to the leaves of the decision tree, while the second step consisted of
estimating a logistic regression model for each segment. However, their method is based on
5
a single non-pruned decision tree as in the hybrid CART-logit model. Furthermore, their
objective was to improve the predictive performance of the logistic regression by identifying
homogeneous subsets of customers and not by introducing non-linear effects as in the PLTR
approach.
The rest of the article is structured as follows. Section 2 analyses the performance of lo-
gistic regression and random forest in the presence of univariate and multivariate threshold
effects through Monte Carlo simulations. In Section 3, we introduce the PLTR credit scor-
ing method and assess through Monte Carlo simulations its accuracy and interpretability
(parsimony) in the presence of threshold effects. Section 4 describes an empirical application
with a benchmark dataset. The robustness of the results to dataset choice is explored in
Section 5. Section 6 compares the models from an economic point of view, while the last
section concludes the paper.
2 Threshold effects in logistic regression

2.1 Non-linear effects and the logistic regression model
Let (xi , yi ), i = 1, ..., n be a sample of size n of independent and identically distributed
observations, where xi ∈ Rp is a p-dimensional vector of predictors and yi ∈ {0, 1} is a
binary variable taking the value one when the i-th borrower defaults and zero otherwise.
The goal of a credit scoring model is to provide an estimate of the posterior probability
Pr (yi = 1 |xi ) that borrower i defaults given its attributes xi . The relevant characteristics
of the borrower vary according to its status: household or company. For corporate credit
risk scoring, the candidate predictive variables xi,j , j = 1, ..., p, may include balance-sheet
financial variables that cover various aspects of the financial strength of the firm, such as
the firm’s operational performance, its liquidity, and capital structure (Altman, 1968). For
instance, using a sample of 4, 796 Belgian firms, Bauweraerts (2016) shows the importance of
taking into account the level of liquidity, solvency and profitability of the firm in forecasting
its bankruptcy risk. For small and medium enterprises (SMEs), specific variables related to
the financial strength of the firm’s owner are also shown to be important (Wang, 2012). For
retail loans, financial variables such as the number and amount of personal loans, normal
repayment frequency of loans, the number of credit cards, the average overdue duration of
credit cards and the amount of housing loans are combined with socio-demographic factors.
A typical example is the FICO score, which is widely used in the US financial industry to
assess the creditworthiness of individual customers.
Regardless of the type of borrower, the conditional probability of default is generally
modelled using a logistic regression with the following specification:
1
Pr (yi = 1|xi ) = F (η (xi ; β)) = , (1)
1 + exp (−η (xi ; β))
6
where F (.) is the logistic cumulative distribution function and η (xi ; β) is the so-called index
function defined as
p
X
η (xi ; β) = β0 + βj xi,j ,
j=1
where β = (β0 , β1 , ..., βp ) ∈ Rp+1 is an unknown vector of parameters. The estimator βb is

obtained by maximizing the log-likelihood function
n n
X o
L(yi ; β) = yi log {F (η (xi ; β))} + (1 − yi ) log {1 − F (η (xi ; β))} .
i=1
The main advantage of the logistic regression model is its simple interpretation. Indeed,
this model searches for a single linear decision boundary in the predictors’ space. The core
assumption for finding this boundary is that the index η (xi ; β) is linearly related to the
predictive variables. In this framework, it is easy to evaluate the relative contribution of
each predictor to the probability of default. This is achieved by computing marginal effects
as
∂ Pr (yi = 1 |xi ) exp (η (xi ; β))
= βj ,
∂xi,j [1 + exp (η (xi ; β))]2
with estimates obtained by replacing β with β.

b Thus, a predictive variable with a positive
(negative) significant coefficient has a positive (negative) impact on the borrower’s default
probability.
This simplicity comes at a cost when significant non-linear relationships exist between the
default indicator, yi , and the predictive variables, xi . A very common type of non-linearity
can arise from the existence of a univariate threshold effect on a single predictive variable,
but it can also be generalised to a combination of such effects (multivariate threshold effects)
across variables. A typical example of the former case in the context of credit scoring is the
income “threshold effect”, which implies the existence of an endogenous income threshold
below (above) which default probability is more (less) prominent. The income threshold
effect can obviously interact with other threshold effects, leading to highly non-linear mul-
tivariate threshold effects. The common practice to approximate non-linear effects in credit
scoring applications is to introduce quadratic and interaction terms in the index function
η (xi ; β). However, such a practice is not successful when unobserved threshold effects are
at stake.
To illustrate the inability of standard parametric models, i.e., standard logistic regression
model or logistic regression with quadratic and interaction terms, to capture accurately the
non-linear effects (thresholds and interactions) that can arise in credit scoring data, we
propose a Monte Carlo simulation experiment. In a first step (simulation step), we generate
p predictive variables xi,j , j = 1, ..., p, i = 1, ..., n, where the sample size is set to n = 5, 000.
Each predictive variable xi,j is assumed to follow the standard Gaussian distribution. The
7
index function η (xi ; Θ) is simulated as follows:
p p−1 p
X X X
η (xi ; Θ) = β0 + βj 1 (xi,j ≤ γj ) + βj,k 1 (xi,j ≤ δj ) 1 (xi,k ≤ δk ) , (2)
j=1 j=1 k=j+1
where 1 (.) is the indicator function and Θ = (β0 , β1 , ..., βp , β1,2 , ..., βp−1,p )0 is the vector of
parameters, with each component randomly drawn from a uniform [−1, 1] distribution, and
(γ1 , ..., γp , δ1 , ..., δp )0 are some threshold parameters, whose values are randomly selected from
the support of each predictive variable generated while excluding data below (above) the
first (last) decile. The default probability is then obtained for each individual by plugging
(2) into (1). Subsequently, the simulated target binary variable yi is obtained as

1 if Pr (yi = 1 |xi ) > π
yi = (3)
0 otherwise,
where π stands for the median value of the generated probabilities.
In a second step (estimation step), we estimate by maximum likelihood two logistic
regressions on the simulated data {yi , xi }ni=1 : (i ) a standard logistic regression model and
(ii ) a (non-linear) logistic regression with quadratic and interaction terms. For the standard
logistic regression model, the conditional probability is based on a linear index defined as
p
X
η (xi ; β) = β0 + βj xi,j .
j=1
For non-linear logistic regression, we also include quadratic and interaction terms
p p p−1 p
X X X X
η (nl) xi ; Θ(nl) = α0 + ξj x2i,j +

αj xi,j + ζj,k xi,j xi,k .
j=1 j=1 j=1 k=j+1
0
where Θ(nl) = (α0 , α1 , ..., αp , ξ1 , ..., ξp , ζ1,2 , ..., ζp−1,p ) is the unknown vector of parameters.
Figure 1 displays the average value of the percent of correct classification (PCC) values of
these two models over 100 simulations and for different numbers of predictors p = 4, ..., 20.7
We observe that their accuracy decreases with the number of predictors. This result arises
because the two logistic regression models are misspecified because they do not control for
threshold and interaction effects, and their degree of misspecification increases with addi-
tional predictors. Indeed, in our DGP (i.e., Equation 2), the number of regressors controls
for the degree of non-linearity of the data generating process: more predictors correspond
to more threshold and interaction effects. These results suggest that in the presence of
univariate and bivariate threshold effects involving many variables, logistic regression with
a linear index function, eventually augmented with quadratic and interaction terms, fails
to discriminate between good and bad loans. In the case where p = 20, the PCCs of the
logistic regression models are equal to only 72.30% and 75.19%.
7
We divide the simulated sample into two sub-samples of equal size at each replication. The training
sample is used to estimate the parameters of the logistic regression model, while the classification perfor-
mance is evaluated on the test sample. To compute the PCC, we estimate yi by comparing the estimated
probabilities of default, pbi , to an endogenous threshold π
b. As usual in the literature, we set π
b to a value such
that the number of predicted defaults in the learning sample is equal to the observed number of defaults.
8
0.86
Linear Logistic regression
Non-Linear Logistic regression
0.84
Proportion of Correct Classification (PCC)
0.82
0.8
0.78
0.76
0.74
0.72
4 6 8 10 12 14 16 18 20
Number of Predictors
Figure 1: Comparison of performances under univariate and bivariate threshold effects:

linear and non-linear logistic regressions
2.2 Machine learning for non-linear effects

In the context of credit scoring, ensemble methods based on decision trees, such as the
random forest method, provide better classification performance than standard logistic re-
gression models (Paleologo et al., 2010; Finlay, 2011; Lessmann et al., 2015). The out-
performance of the random forest method arises from the non-linear “if-then-else” rules
underlying decision trees.8 Formally, for a given tree, l, the algorithm proceeds as follows.
Let Dm,l be the data (sub)set at a given node m of this tree. We denote by θm,l = (jm,l , tm,l,j )
a candidate split, where jm,l = 1, ..., p indicates a given predictor and tm,l,j is a threshold
value in the support of this variable. The algorithm partitions the data Dm,l into two subsets
Dm,l,1 (θm,l ) and Dm,l,2 (θm,l ), with9
Dm,l,1 (θm,l ) = (xi , yi ) | xi,j < tm,l,j ,
Dm,l,2 (θm,l ) = (xi , yi ) | xi,j ≥ tm,l,j ,
where the parameter estimates θbm,l satisfy

1
θm,l = (jm,l , tm,l,j ) = arg max H (Dm,l ) −
b b b H(Dm,l,1 (θm,l )) + H(Dm,l,2 (θm,l )) ,
θm,l 2
with H (.) a measure of diversity, e.g., the Gini criterion, applied to the full sample and
averaged across the two sub-samples. Hence, θbm,l appears as the value of θm,l that reduces
8
Indeed, the latter is a non-parametric supervised learning method based on a divide-and-conquer greedy
algorithm that recursively partitions the training sample into smaller subsets to group together as accurately
as possible individuals with the same behaviour, i.e., the same value of the binary target variable “yi ”.
9
To simplify the description of the algorithm, we focus only on quantitative predictors. A similar proce-
dure is available for qualitative predictors.
9
diversity the most within each subset resulting from the split. The splitting process is
repeated until the terminal sub-samples, also known as leaf nodes, contain homogeneous
individuals according to a predefined homogeneity rule. We denote by Ml the total number
of splits in tree l and by |Tl | the corresponding number of leaf nodes.
Figure 2: Example of a decision tree for credit scoring
An illustrative example of a decision tree is given below in Figure 2. At the first iteration
(or split), m = 1, θbm,l is defined by (b
jm,l , b
tm,l,1 ), with b
jm,l the index of the variable “income”
and b
tm,l,1 = 33270.53. The other iterations also include “age” and “education” for further
refinements. The process ends with a total number of 5 splits and 6 leaf nodes labelled
10, 11, 12, 13, 4 and 7. Each leaf node Rt , t = 1, ..., |Tl | includes a specific proportion
of individuals belonging to each class of borrowers (1=“default”, 0=“no default”). For
instance, leaf node “7” contains 89 individuals, 93.3% of them having experienced a default
event. Note that each of these individuals has an income lower than 33270.53 and is less
than 28.5 years old. The predominant class in each leaf defines the predicted value of yi for
individuals i that belong to that particular leaf. Formally, the predicted default value for
the ith individual is
|Tl |
X
hl (xi ; Θ
b l) = ct Ri,t ,
t=1
where Θl = (θm,l , m = 1, ..., Ml ) is the parameter vector for tree l, Ri,t = 1(i∈Rt ) indicates
whether individual i belongs to leaf Rt , and ct is the dominant class of borrowers in that
leaf node. For example, in leaf node 7, the “default” class is dominant; hence, the predicted
value hl (xi ) is equal to 1 for all the individuals that belong to this leaf node. Note that
this simple tree allows us to identify both interaction and threshold effects. For instance, in
the simple example of Figure 2, the predicted value can be viewed as the result of a kind
10
of linear regression10 on the product of two binary variables that take a value of one if the
income is lower than 33270.53 and the age is less than 28.5.
The random forest method is a bagging procedure that aggregates many uncorrelated
decision trees. It exploits decision-tree power to detect univariate and multivariate threshold
effects while reducing their instability. Its superior predictive performance springs from the
variance reduction effect of bootstrap aggregation for non-correlated predictions (Breiman,
1996). Let L trees be constructed from bootstrap samples (with replacement) of fixed size
drawn from the original sample. To ensure a low level of correlation among those trees,
the random forest algorithm chooses the candidate variable for each split in every tree, jm,l
with m ∈ {1, . . . , Ml } and l ∈ {1, . . . , L}, from a restricted number of randomly selected
predictors among the p available ones. The default prediction of the random forest for each
borrower, h (xi ), is obtained by the principle of majority vote; that is, h (xi ) corresponds to
the mode of the empirical distribution of hl (xi ; Θ
b l ), l = 1, ..., L.
1
Random Forest
0.95
0.9
0.85
0.8
0.75
0.7
4 6 8 10 12 14 16 18 20

linear and non-linear logistic regressions and the random forest method
To illustrate the ability of the random forest method to capture the non-linear effects
that can arise in credit scoring data well, we consider the same Monte Carlo framework
as in Section 2.1. The proportion of correct classification for the random forest algorithm,
displayed as a yellow line in Figure 3, is computed over the same test samples of length
2, 500 as the PCCs of the logistic regressions previously discussed. The optimal number of
trees in the forest, L, is tuned using the out-of-bag error. Our results confirm the empirical
findings of the literature: in the presence of non-linear effects, random forest outperforms
not only linear logistic regression (as expected) but also non-linear logistic regression. This
10
This equivalence is true only in the case of a regression tree when the target variable y is continuous.
11
illustrates the ability of random forests to capture both threshold and interaction effects
between the predictors well. These findings are valid regardless of the number of predictors,
even if the differences in classification performance between the three models are decreasing
in the number of predictors (see e.g. Vapnik and Chervonenkis, 1971, for a theoretical result
on the fact that theoretical risk or generalisation error for any machine learning algorithm
decreases with the number of observations and increases with the complexity given by the
number of predictors, or more generally the so-called VC dimension). Indeed, as the number
of predictors increases, the complexity and the non-linearity of the DGP also increases,
which diminishes the performance of all the classifiers. For instance, the PPCs are equal to
99.18% (resp. 84.50%) for the random forest (resp. logistic regression with quadratic and
interaction terms) in the case with 4 predictors, against 81.20% (resp. 75.19%) in the case
with 20 predictors.
Despite ensuring good performance, the aggregation rule (majority vote) underlying the
random forest method leads to a prediction rule that lacks interpretation. This opacity is
harmful for credit scoring applications, where decision makers and regulators need simple
and interpretable scores (see ACPR, 2020 and EC, 2020, among many others). The key
question here is how to find a suitable trade-off between predictive performance and in-
terpretability. To address this issue, two lines of research can be explored. First, one can
try to diminish the complexity of the random forest method’s aggregation rule by selecting
(via an objective criterion) only some trees or decision rules in the forest.11 Second, we can
preserve the simplicity of logistic regression while improving its predictive performance with
univariate and bivariate endogenous threshold effects. We opt here for the second line of
research, with the PLTR hybrid scoring approach.
3 Penalised logistic tree regression

3.1 Description of the methodology
PLTR aims to improve the predictive performance of the logistic regression model through
new predictors based on short-depth decision trees and a penalised estimation method while
preserving the intrinsic interpretability of the scoring model. The algorithm proceeds in two
steps.
The objective of the first step is to identify threshold effects from trees with two splits.
For illustration, take the income and age as the j th and k th explanatory variables, respec-
tively, and assume that income is more informative than age in explaining credit default.
For each individual i, the corresponding decision tree generates three binary variables, each
(j)
associated with a terminal node. The first binary variable Vi,1 accounts for univariate
11
Note that this is the approach underlying the so-called InTrees method of Deng (2019), who proposed
a methodology to render the random forest outputs interpretable by extracting simple rules from a tree
ensemble.
12
threshold effects and takes the value of one when the income of individual i is higher than
(j,k)
an estimated income threshold and zero otherwise. The second (third) binary variable Vi,2
(j,k)
(Vi,3 ), representing bivariate threshold effects, is equal to one when the person’s income is
lower than its threshold and at the same time his/her age is lower (higher) than an estimated
age threshold and zero otherwise.12 Note that this particular form of splitting should arise
when both variables are informative, i.e., each of them is selected in the iterative process
of splitting. If the second variable is non-informative (age), the tree relies twice on the first
informative variable (income). Figure 4 gives an illustration of the splitting process.
Figure 4: Illustration of the two-stage splitting process
One leaf of each of the two branches originating from the root of the tree is retained to
(j) (j,k)
cover both one and two splits, i.e., the first two binary variables Vi,1 and Vi,2 in the example
above. We count at most p+q threshold effects for inclusion in our logistic regression, where
p represents the number of predictive variables and q denotes the total number of couples of
predictors, with q ≤ p × (p − 1) /2. This is the case because the univariate threshold effects
(j)
Vi,1 are generated only by the variables retained in the first split irrespective of the variables
retained in the second split. Some predictive variables may be selected in the first split of
several trees, while others may never be retained. The latter group does not produce any
univariate threshold effects, while the former group delivers identical univariate threshold
(j)
effects, Vi,1 , out of which only one is included in the logistic regression.13
12 (j)
It is also possible that the univariate threshold variable Vi,1 takes the value of one when the income is
lower than an estimated income threshold, and zero otherwise. In that case, the bivariate threshold effect
(j,k) (j,k)
Vi,2 (Vi,3 ) is equal to one when the individual’s income is higher than its threshold and at the same time
his/her age is lower (higher) than an estimated age threshold, and zero otherwise.
13
Note that one could also go beyond two splits by analysing triplets or quadruplets of predictive variables.
Such a procedure would allow the inclusion of more complex non-linear relationships in the logistic regression.
13
In the second step, the endogenous univariate and bivariate threshold effects previously
obtained are plugged in the logistic regression

(j) (j,k)
1
Pr yi = 1|Vi,1 , Vi,2 ; Θ = h i, (4)
(j) (j,k)
1 + exp −η(Vi,1 , Vi,2 ; Θ)
with
p p p−1 p
(j) (j,k)
X X (j)
X X (j,k)
η(Vi,1 , Vi,2 ; Θ) = β0 + α j xi + βj Vi,1 + γj,k Vi,2
j=1 j=1 j=1 k=j+1
the index and Θ = (β0 , α1 , ..., αp , β1 , ..., βp , γ1,2 , ..., γp−1,p )0 the set of parameters to be esti-
mated. The corresponding log-likelihood is
n
(j) (j,k) 1 Xh h
(j) (j,k)
i
L(Vi,1 , Vi,2 ; Θ) = yi log F η(Vi,1 , Vi,2 ; Θ)
n i=1
h i i
(j) (j,k)
+ (1 − yi ) log 1 − F η(Vi,1 , Vi,2 ; Θ) ,
(j) (j,k)
where F (η(Vi,1 , Vi,2 ; Θ)) is the logistic cdf. The estimate Θ
b is obtained by maximizing
the above log-likelihood with respect to the unknown parameters Θ. Note that the length
of Θ depends on the number of predictive variables, p, which can be relatively high. For
instance, there are 45 couples of variables when p = 10; this leads to a maximum number of
55 univariate and bivariate threshold effects that play the role of predictors in our logistic
regression.
To prevent overfitting issues in this context with a large number of predictors, a common
approach is to rely on penalisation (regularisation) for both estimation and variable selection.
In our case, this method consists of adding a penalty term to the negative value of the log-
likelihood function, such that
(j) (j,k) (j) (j,k)
Lp (Vi,1 , Vi,2 ; Θ) = −L(Vi,1 , Vi,2 ; Θ) + λP (Θ), (5)
where P (Θ) is the additional penalty term and λ is a tuning parameter that controls the
intensity of the regularisation and which is selected in such a way that the resulting model
minimises the out-of-sample error. The optimal value of the tuning parameter λ is usually
obtained by relying on a grid search with cross-validation or by using some information
criteria. In addition, several penalty terms P (Θ) have been proposed in the related literature
(Tibshirani, 1996; Zou and Hastie, 2005; Zou, 2006). Here, we consider the adaptive lasso
estimator of Zou (2006). Note that the adaptive lasso satisfies the oracle property; i.e., the
probability of excluding relevant variables and selecting irrelevant variables is zero, contrary
to the standard lasso penalisation (Fan and Li, 2001). The corresponding penalty term
(0) (0)
is P (Θ) = Vv=1 wv |θv | with wv = |θbv |−ν , where θbv , v = 1, ..., V , are consistent initial
P
Nevertheless, the expected uprise in performance would come at the cost of increased complexity of the model
toward that of random forests, which would plunge its level of interpretability. For this reason, in our PLTR
model, we use only short-depth decision trees involving two splits.
14
estimators of the parameters and ν is a positive constant. The adaptive lasso estimators
are obtained as
V
X
(j) (j,k)
Θalasso (λ) = arg min − L Vi,1 , Vi,2 ; Θ + λ
b wv |θv |. (6)
Θ
v=1
(0)
In practice, we set the parameter ν to 1 and the initial estimator θbj to the value obtained
from the logistic-ridge regression (Hoerl and Kennard, 1970), and the only free tuning pa-
rameter, λ, is found via 10-fold cross-validation.14 Note also that since decision tree algo-
rithms are immune to collinearity by nature and the adaptive lasso regression is consistent
in variable selection for Generalized Linear Models even when the irrepresentable condition
is violated, the PLTR method is robust to collinearity issues.
In summary, PLTR is a hybrid classification model designed to increase the predictive
power of the logistic regression model via feature engineering. Its first step consists of
creating additional binary predictors based on short-depth decision trees built with couples
of predictive variables. These binary variables are then introduced, in a second step, in a
penalised logistic regression model, where the adaptive lasso is used for both estimation and
variable selection.
3.2 PLTR under threshold effects: Monte Carlo evidence

In this subsection, we assess the accuracy and interpretability of the PLTR method in the
presence of threshold effects. For that, we consider the same Monte Carlo experiment as
that defined in Section 2.
Figure 5 displays the PCC for our PLTR method computed over the same test samples of
length 2, 500 that were generated with the DGP in (2)-(3). The main conclusion is that the
PLTR method outperforms the two versions of the logistic regression, i.e., with and without
quadratic and interaction terms. Equally important, when there are few predictors, i.e., p
is small, the PCC of PLTR is lower than that of random forest. However, as p increases,
the performance of PLTR approaches that of the random forest method, and both models
have approximately the same classification performance. For example, the PCCs are equal
to 94.81 for our new method and 99.18 for the random forest with p = 4, against 83.65
and 81.20 for p = 20, respectively. Note that the latter case seems more realistic, as credit
scoring applications generally rely on a large set of predictors in practice.
Performance is not the only essential criterion for credit scoring managers. The other
fundamental characteristic of a good scoring model is interpretability. Interpretability and
14
Different estimation algorithms have been developed in the literature to estimate regression models with
the adaptive lasso penalty (for a given value of λ): the quadratic programming technique (Shewchuk et al.,
1994), the shooting algorithm (Zhang and Lu, 2007), the coordinate-descent algorithm (Friedman et al.,
2010), and the Fisher scoring algorithm (Park and Hastie, 2007). Most of them are implemented in software
such as MATLAB and R, and we rely here on the algorithm based on Fisher scoring. See McIlhagga (2016)
for more details on this optimisation algorithm.
15
1
Random Forest
PLTR
0.95
0.9
0.85
0.8
0.75
0.7
4 6 8 10 12 14 16 18 20

linear and non-linear logistic regressions, the random forest method and PLTR
accuracy are generally two competing objectives: the first is favoured by simple models,
the latter by complex models. Moreover, the degree of interpretability of a credit scoring
model is difficult to measure. As discussed in Molnar (2019), there is no real consensus
in the literature about what is interpretable for machine learning, nor is it clear how to
measure this factor. Doshi-Velez and Kim (2017) distinguishes three levels of evaluation
of interpretability: the application level, the human level, and the function level. While
the application and human levels are related to the understanding of the conclusion of
a model (from an expert or a layperson, respectively), the function level corresponds to
the evaluation of decision rules from a statistical viewpoint (for example, the depth of a
decision tree). In the specific context of credit scoring, Bracke et al. (2019) distinguishes six
different types of stakeholders (developers, 1st- and 2nd-line model checkers, management,
regulators, etc. ).15 Each of them has its own definition of what interpretability should be
and how to measure it. For instance, the developer and 1st-line checkers may be interested
in individual predictions when they obtain customer queries and in better understanding
outliers. In contrast, second-line model checkers, management, and prudential regulators are
likely to adopt a more general viewpoint and may be less interested in individual predictions.
In the credit scoring context, interpretability can be measured from at least two perspec-
15
Bracke et al. (2019) distinguished the (i) developers, i.e., those developing or implementing an ML
application; (ii) 1st-line model checkers, i.e., those directly responsible for ensuring that model development
is of sufficient quality; (iii) management responsible for the application; (iv) 2nd-line model checkers, i.e.,
staff that, as part of a firm’s control functions, independently check the quality of model development
and deployment; (v) conduct regulators that are interested in deployed models being in line with conduct
rules and (vi) prudential regulators that are interested in deployed models being in line with prudential
requirements.
16
tives. First, one can consider simple metrics such as the size of the set of decision rules. This
indicator allows us to compare models in terms of ease of interpretation: the fewer the rules
in a decision set, the easier it is for a user to understand all the conditions that correspond
to a particular class label. The size of a given rule in a decision set is a complementary
measure. Indeed, if the number of predicates in a rule is too large, it will lose its natural
interpretability. This perspective corresponds to the function level evaluation mentioned by
Doshi-Velez and Kim (2017). Second, one can interpret the decision rules through marginal
effects, elasticities, or scorecards. This second perspective corresponds to the human-level
evaluation evoked by Doshi-Velez and Kim (2017) or to the global model interpretability
defined by Molnar (2019). Which features are important and what kind of interactions take
place between them?
In this paper, we confirm this trade-off between interpretability and classification per-
formance. The less accurate model, i.e., the logistic regression model, is intrinsically in-
terpretable through marginal effects or explicit scorecard. In contrast, the model with the
highest classification performance among our competing models, i.e., the random forest
model, is not interpretable for two reasons. First, the forest relies on many trees with many
splits, which involves many complicated if-then-else rules. Second, the rules obtained from
the trees are aggregated via the majority vote.
Within this context, our PLTR method is a parsimonious solution to the trade-off be-
tween performance and interpretability. The scoring decision rules are simple to interpret
through marginal effects (as well as elasticities and scorecards) similar to those of tradi-
tional logistic regression. This is facilitated by the simple decision rules obtained in the first
step of the procedure from short-depth decision trees. Indeed, the skeleton of our PLTR
is actually a logistic regression model with binary indicators that account for endogenous
univariate and bivariate threshold effects. The complete loan-decision process based on the
PLTR method is illustrated in Figure 6. The input of the method includes all the predictive
variables from the loan applicant, while the output is fundamentally the decision to accept
or to reject the credit application based on the default risk of the person. Additionally, the
mapping from the inputs to the output allows one to transform the internal set of rules of
PLTR into transparent feedback about the weaknesses and strengths of the application.
To provide more insights into interpretability, we compare our PLTR model and the
random forest in the same Monte Carlo setup as in Section 2, with p fixed to 20, using
simple metrics. We consider the two metrics previously defined, i.e., the size of the set of
decision rules and the size of a given rule in the decision set. Across the 100 simulations, the
random forest registers an average number of 160.9 trees, each with an average number of
410.5 terminal nodes. This leads to a decision set of 410.5 × 160.9 binary decision variables
or rules that can be used for prediction with this method. Across the same simulations,
the average number of active binary decision variables in our penalised logistic regression
17
Figure 6: PLTR inference process
is equal to 146.9.16 Moreover, the number of predicates involved in each of these binary
decision variables for our PLTR method varies between 1 and 2 by construction, whereas
the maximum number of predicates in a rule of the random forest model is 14.5 on average.
Hence, the PLTR model appears to be easier to interpret than the random forest model and
comparable to non-linear logistic regression in this sense.17
Furthermore, marginal effects and elasticities can be easily obtained in PLTR due to
the linearity of the link function (cf. Equation 4). On the one hand, this greatly simplifies
significance testing as well as the implementation of out-of-sample exercises. On the other
hand, this allows credit institutions to easily explain, in a transparent way, the main reasons
behind a loan decision.
4 Model performance with a benchmark dataset

As a complement to Monte Carlo simulations, we now consider an empirical application
based on a benchmark credit default dataset to assess the practical usefulness of PLTR.
16
Note that for p = 20 predictors, the maximum number of binary variables is equal to 20 + 20×19
2 = 210.
This result illustrates the selection processed through adaptive lasso regression.
17
The major difference between these two methods is the endogenous character of the thresholds that
characterise variable interactions in our framework.
18
4.1 Data description and processing
To gauge the out-of-sample performance of the PLTR method and to illustrate its inter-
pretability, we use a popular dataset provided by a financial institution for the Kaggle
competition “Give me some credit”, which is often used in credit scoring applications (Bae-
sens et al., 2003). The dataset includes several predictive variables and a binary response
variable measuring default. The predictive variables provide information about the cus-
tomers (age, monthly income, the number of dependents in the family) and the application
form (number of mortgage and real estate loans, the monthly debt payments, the total bal-
ance on credit cards, etc.). The dataset contains 10 quantitative predictors. See Table A.3
in Appendix A for a description of the variables in the dataset.
The number of instances in the dataset is equal to 150, 000 loans out of which 10, 026
defaults, leading to a prior default rate of 0.067.18 All the missing values have been replaced
by the mean of the predictive variable. Finally, regarding data partitioning, we use the so-
called N × 2-fold cross-validation of Dietterich (1998), which involves randomly dividing the
dataset into two sub-samples of equal size. The first (second) part is used to build the model,
while the second (first) part is used for evaluation. This procedure is repeated N times,
and the evaluation metrics are averaged. This method of evaluation produces more robust
results compared to classical single data partitioning. We set N = 5 for computational
reasons.
4.2 Statistical measures of performance and interpretability

To evaluate the performance of each classifier, we use five accuracy measures considered by
Lessmann et al. (2015) in their benchmarking study: the area under the ROC curve (AUC),
the Brier score (BS), the Kolmogorov-Smirnov statistic (KS), the percentage of correctly
classified (PCC) cases, and the partial Gini index (PGI). These indicators are related to
different facets of the predictive performance of scorecards, namely, the accuracy of the
scores as measured by the BS statistics, the quality of the classification given by the PCC
and KS statistics, and the discriminatory power assessed through the AUC and the PGI
statistics. By using several statistics instead of a single one, we expect to obtain a robust
and complete evaluation of the relative performances of the competing models.
The AUC tool evaluates the overall discriminatory performance of each model or classi-
fier. It is a measure of the link between the false positive rate (FPR) and the true positive
rate (TPR), each computed for every threshold between 0 and 1. The FPR (TPR) is the per-
centage of non-defaulted (defaulted) loans misclassified as defaulted (non-defaulted). Thus,
18
It is well known that imbalanced classes impede classification: some classifiers may focus too much
on the majority class and neglect the minority group (of interest). They can hence exhibit good overall
performance despite poorly identifying the minority group, i.e., the borrowers that default. A common
solution consists of using an under-sampling or over-sampling method, such as SMOTE. Nonetheless, here,
we choose not to resample the data, as the prior default rate is larger than 6%.
19
AUC reflects the probability that the occurrence of a randomly chosen bad loan is higher
than the occurrence of a randomly chosen good loan.
The Gini index is equal to twice the area between the ROC curve and the diagonal.
Hence, similar to the AUC metric, it evaluates the discriminatory power of a classifier
across several thresholds, with values close to one corresponding to perfect classifications.
However, in credit scoring applications, it is not realistic to study all possible thresholds.
Informative thresholds are those located in the lower tail of the distribution of default
probabilities (Hand, 2005). Indeed, only applications below a threshold in the lower tail can
be granted a credit, which excludes high thresholds. The partial Gini index solves this issue
by focusing on thresholds in the lower tail (Pundir and Seshadri, 2012). With x denoting a
given threshold and L(x) denoting the function describing the ROC curve, the PGI is then
defined as19
Rb
2 a L(x)dx
P GI = − 1.
(a + b)(b − a)
The PCC is the proportion of loans that are correctly classified by the model. Its
computation requires discretisation of the continuous variable of estimated probabilities of
default. Formally, we need to choose a threshold π above (below) which a loan is classified
as bad (good). In practice, the threshold π is fixed by comparing the costs of rejecting good
customers/granting credits to bad customers. Since we do not have such information, we set
this threshold to a value such that the predicted number of defaults in the learning sample
is equal to the observed number of defaults.
The Kolmogorov-Smirnov statistic is defined as the maximum distance between the
estimated cumulative distribution functions of two random variables. In credit scoring
applications, these two random variables measure the scores of good loans and bad loans
(Thomas et al., 2002).
Lastly, the Brier score (Brier, 1950) is defined as
n
1X c
BS = (Pr(yi = 1|xi ) − yi )2 ,
n i=1
where Pr(y
c i = 1|xi ) is the estimated probability of default and yi is the target binary default
variable. Note that it is the equivalent of the mean-square error but designed for the case
of discrete-choice models. Overall, the higher these indicators are, the better the model is,
except for the Brier score, for which a smaller value is better.
Regarding the interpretability of the scoring models, the criteria retained to compare
PLTR and the random forest method are the size of the decision set and the average size of
rules in a decision set (see also Subsection 3.2).
19
The PGI within bounds a = 0 and b = 1 is equivalent to the Gini Index. In the empirical applications,
we evaluate the PGI within the (0, 0.4) bounds as in Lessmann et al. (2015).
20
4.3 Statistical evaluation results
Table 1 presents the average value of each statistic across the 5 × 2 cross-validation test
samples. We compare the out-of-sample performance of PLTR to that of traditional logistic
regression and the random forest method. Three different versions of the logistic regression
are implemented: simple linear logistic regression, its non-linear version, which includes as
additional variables quadratic and interaction terms,20 , and a penalised version of this last
model to avoid overfitting due to the large number of predictors. We use the adaptive lasso
penalty as described above. These augmented logistic regression models are used to assess
the importance of non-linear effects of the features. We also include an SVM and NN in
the comparison, as they are widely used for credit scoring applications in the literature
(Thomas, 2000; Baesens et al., 2003; Lessmann et al., 2015).
Table 1: Statistical performance indicators: Kaggle dataset
Methods AUC PGI PCC KS BS

Linear Logistic Regression 0.6983 0.3964 0.9082 0.3168 0.0576
Non-Linear Logistic Regression 0.7660 0.5255 0.9127 0.4173 0.0649
Non-Linear Logistic Regression + ALasso 0.8062 0.6102 0.9208 0.4751 0.0535
Random Forest 0.8529 0.6990 0.9260 0.5563 0.0500
PLTR 0.8568 0.7076 0.9247 0.5647 0.0496
Support Vector Machine 0.7418 0.4830 0.9117 0.3723 0.0619
Neural Network 0.7517 0.5006 0.9074 0.3895 0.0552
Note: Non-linear logistic regression includes linear, quadratic and interaction terms. The method labelled
“Non-Linear Logistic Regression + ALasso” corresponds to a penalised version of non-linear logistic regres-
sion with an adaptive lasso penalty.
The results displayed in Table 1 show that the random forest method performs better
than the three versions of the logistic regression, and this holds for all statistical measures
considered. In particular, the differences are more pronounced for the AUC, PGI and
KS statistics. Our PLTR method also performs better than the three versions of logistic
regression irrespective of the performance measure. This is particularly applicable for the
AUC, PGI and KS metrics, for which the dominance is stronger. The take-away message here
is that combining decision trees with a standard model such as logistic regression provides
a valuable statistical modelling solution for credit scoring. In other words, the non-linearity
captured by univariate and bivariate threshold effects obtained from short-depth decision
trees can improve the out-of-sample performance of traditional logistic regression. The SVM
and NN results are consistent with those in the literature (Thomas, 2000; Baesens et al.,
2003; Lessmann et al., 2015; Grennepois et al., 2018). They are slightly better than those of
the logistic regression model, but these methods generally perform less well than ensemble
20
As already stressed, this non-linear model is the one that is generally used to capture non-linear effects
in the framework of logistic regression.
21
learning methods such as the random forest method. Most importantly, these models also
perform less well than PLTR.
The results in Table 1 also show that PLTR compares competitively to the random forest
method. All statistical performance measures are of the same order. Therefore, the two
methods exhibit similar statistical performance, and neither of them should be preferred
over the other based on these criteria. However, the parsimony of PLTR contrasts with
the complexity underlying the prediction rule of the random forest method. To illustrate
this point, Table 2 displays the interpretability measures for the random forest method and
PLTR, as well as that of linear logistic regression for comparison purposes. The average
number of trees in the random forest method across the 5 × 2 cross-validation test samples
is equal to 173.9. These trees have on average 5, 571.1 terminal nodes, with a total of
5, 571.1 × 173.9 binary variables for prediction (via the majority vote). By contrast, the
average number of bivariate threshold effects selected by our penalised logistic regression is
only 40. More importantly, these bivariate threshold effects are easily interpretable because
they arise from short-depth decision trees. In addition, the PLTR rules are built from only
2 predicates at most, whereas the rules from the random forest method are built from an
average number of 32.2 predicates at most. Overall, both criteria confirm that PLTR is
easier to interpret than the random forest method. These differences in terms of the size of
the decision set and size of the rules between both models are the penalty of capturing more
non-linear effects, although such effects do not seem to play a significant role in this dataset.
For comparison, the average number of predictors is 11 for linear logistic regression, each
of them relying on a single predicate. The PLTR results are not very different from those
of linear logistic regression, with the gap corresponding to the non-linear effects included in
our model to improve the performance of the benchmark linear logistic regression method.
Table 2: Measures of interpretability: Kaggle dataset

Methods Size of the decision set Maximal number of predicates
Linear Logistic Regression 11 1
Random Forest 5,571.1 × 173.9 32.2
PLTR 40 2
Note: This table displays the average values of interpretability measures for linear logistic regression, the
random forest method and PLTR.
Lastly, to highlight the interpretability advantage of our method, we report in Table 3

the 10 most important decision rules from short-depth decision trees, which are selected
by an adaptive lasso in the implementation of our PLTR method. These decision rules are
associated with the largest absolute values of the marginal effects (averaged across individ-
uals). A positive (negative) value of a given marginal effect provides information about the
strength of an increase (decrease) of the probability of default. We observe that three uni-
22
variate threshold variables are selected, i.e., “NumberOfTime60-89DaysPastDueNotWorse
< 0.5”, “NumberOfTimes90DaysLate<0.5” and
“RevolvingUtilizationOfUnsecuredLines<0.69814”, the first one appearing as the most im-
portant in terms of marginal effect. Referring to the description of this variable in Table
A.3, we can infer that the probability of default is 3.92% less important when the number
of times a borrower has been between 60 and 89 days past due (but not worse in the last 2
years) is lower than 0.5 compared to the reference case when this number is higher than 0.5.
Moreover, seven bivariate threshold effects are selected by the models as being important in
explaining credit default. This kind of analysis that helps measure through marginal effects
the importance of the decision rules from the short-depth decision trees is an important
added value of our PLTR model in terms of interpretability.
5 Robustness across datasets

In this section, we evaluate the out-of-sample robustness of the above empirical results across
datasets. To this end, we consider three popular additional datasets. The first one, named
“Housing”, is available in an SAS library and has been used by many authors for illustrative
examples (Matignon, 2007). The second dataset, labelled the “Australian dataset”, concerns
credit card applications and is a University of California at Irvine (UCI) dataset provided by
Quinlan, and it was used as a credit approval database in the Statlog project.21 Lastly, the
third dataset, labelled the “Taiwan dataset”, is also a UCI dataset that collects information
about default payments in Taiwan.
The Housing dataset includes 5, 960 loans, 1, 189 of which defaulted. Therefore, the prior
default rate is 19.95%. In the Australian (Taiwan) dataset, there are 690 (30, 000) instances
out of which 307 (6, 636) defaulted, leading to a prior default rate of 44.49% (22.12%). In
the Housing dataset, there are 12 explanatory variables, two of which are nominal. The
Australian dataset includes 6 numerical and 8 nominal predictors. For the Taiwan dataset,
there are 23 predictors, nine of which are nominal. Tables A.4 and A.5 display the list of
predictive variables for the Housing and Taiwan datasets, respectively. We do not provide
this information for the Australian dataset, as all attribute names and values have been
changed to meaningless symbols to maintain the confidentiality of the data.
We rely on the same (N ×2) comparison setup as for the benchmark Kaggle dataset, with
N = 5. Table 4 displays the values of the five statistics retained for the comparison of the
alternative models. For the Australian dataset, the two best performing models are PLTR
and the random forest method, with similar values for all five statistics.22 This finding once
21
StatLog is an international project that aims to compare the performances of machine learning, statis-
tical, and NN algorithms on datasets from real-world industrial areas, including medicine, finance, image
analysis, and engineering design.
22
For the non-linear logistic regression results, we find that all fitted probabilities are higher than 0.6.
Therefore, as we compute the PGI within (0, 0.4), this statistic cannot be computed. Unlike in practice,
23
Table 3: Decision rules and average marginal effects: full sample Kaggle dataset
# Decision rules Average marginal effects
1 “NumberOfTime60-89DaysPastDueNotWorse < 0.5” -0.0392
2 “NumberOfTimes90DaysLate<0.5” & “RevolvingUtilizationOfUnsecuredLines<0.59907” -0.0389
3 “NumberOfTimes90DaysLate<0.5” & “NumberOfTime60-89DaysPastDueNotWorse<0.5” -0.0342
4 “NumberOfTime60-89DaysPastDueNotWorse<0.5” & “NumberOfTime30-59DaysPastDueNotWorse<0.5” -0.0326
5 “NumberOfTimes90DaysLate<0.5” -0.0326
24
6 “NumberOfTime60-89DaysPastDueNotWorse>=0.5” & “NumberOfTime60-89DaysPastDueNotWorse<1.5” -0.0300
7 “RevolvingUtilizationOfUnsecuredLines>=0.69814” & “RevolvingUtilizationOfUnsecuredLines<1.001” -0.0285
8 “RevolvingUtilizationOfUnsecuredLines<0.69814” -0.0281
Note: The table provides the list of the decision rules associated with the 10 largest absolute values of the marginal effects (with respect to the probability of
defaulting) derived from the PLTR model estimated using the full sample. See Table A.3 in Appendix A for a precise description of the variables.
again confirms the relevance of our approach in terms of statistical performance. The same
picture is observed for the Taiwan dataset with the PLTR model appearing as efficient as
the random forest method.
Lastly, for the Housing dataset, the random forest method and PLTR are once again
the best performing models. However, in contrast to the results obtained for the other
datasets, the random forest model outperforms our method. Table 5 displays the inter-
pretability performance for these three additional datasets. Using the same arguments as
above, the average number of active variables (univariate and bivariate threshold effects) in
our penalised logistic regression is equal to 47.6, while the random forest method relies on
an average of 343.8 × 110.5 binary variables for prediction.23 Moreover, the results of PLTR
are close to those of linear logistic regression for both criteria, indicating that the PLTR
model remains interpretable despite including non-linear effects.
Other results, available upon request, show that by relaxing the constraint of parsi-
mony via the inclusion of trivariate and quadrivariate threshold effects, the performance
of our penalised logistic regression increases and reaches that of the random forest model.
This suggests that complex non-linear relationships that go beyond univariate and bivariate
threshold effects are present in this dataset. In view of this result, it is important to empha-
sise that our method offers a highly flexible framework to credit risk managers, as they can
tune their model according to the desired level of parsimony. The predictive performance
can be significantly improved but at the cost of less interpretable results.
Additional robustness results consists in out-of-sample forecasting performance com-
parison tests of the three main competing models, i.e. linear logistic regression, random
forest, and PLTR. We rely on Diebold-Mariano (Diebold and Mariano, 1995) and AUC
tests (Candelon et al., 2012) to perform pairwise comparisons and on the Model Confidence
Set (Hansen et al., 2011) to identify the bucket of models that are superior to the remaining
ones and which exhibit similar performance. They are all well known model comparison
approaches, the second being specific to the case with binary dependent variables. The
pairwise tests are two-sided, the null hypothesis corresponds to equal performance and its
rejection indicates that the model with smaller average loss is better. At the same time,
the Model Confidence Set identifies the subset of models that exhibit similar forecasting
abilities and outperform the remaining approaches.
Tables A.1 and A.2 in the Appendix A display these results for the four datasets under
analysis. They take the form of percentage of rejection of each null hypothesis in the 5 × 2
cross-validation test samples and the outperforming model under the alternative hypothesis
this bad performance can also be observed through the high value of the BS statistic compared to those of
the other methods.
23
In this dataset, we identify on average of 110.5 trees in the forest, with an average number of terminal
nodes equal to 343.8 for each tree. Furthermore, at most, 18.8 predicates are used on average in the rules of
the random forest method against 2 at most for the PLTR model. Hence, PLTR is once again better from
the interpretability point of view.
25
is displayed below in parentheses. Two different loss functions are used for the general tests
(Diebold-Mariano and Model Confidence Set), namely the Brier Score and the opposite of
the log-likelihood, in the spirit of a robustness check.
All findings are consistent with those already obtained with statistical performance in-
dicators. Namely, the pairwise comparisons reveal that the PLTR method is superior to
standard logistic regression, and its performance is far better than that of random forests
in two datasets, in the other two the results being more mitigated. Additionally, the Model
Confidence Set identifies most often the PLTR method as that belonging to the subset of
outperforming models.
Table 4: Statistical performance indicators: robustness check
Methods AUC PGI PCC KS BS

Australian dataset
Non-Linear Logistic Regression 0.6090 0.6067 0.2266 0.3921
Random Forest 0.9344 0.6246 0.8603 0.7523 0.0999
PLTR 0.9299 0.6370 0.8606 0.7425 0.1029
Neural Network 0.9141 0.5799 0.8539 0.7366 0.1102
Taiwan dataset
Random Forest 0.7722 0.4924 0.8102 0.4177 0.1362
PLTR 0.7780 0.5156 0.7959 0.4257 0.1352
Neural Network 0.7304 0.4226 0.7879 0.3885 0.1401
Housing dataset
Random Forest 0.9387 0.8157 0.9036 0.7455 0.0736
PLTR 0.9011 0.7341 0.8818 0.6694 0.0844
Neural Network 0.7910 0.5478 0.8132 0.4470 0.1208
sion with the adaptive lasso penalty.
6 Economic evaluation
An important question for a credit risk manager is to what extent these out-of-sample statis-
tical performance gains have a positive impact at a financial level for a credit company. An
26
Table 5: Measures of interpretability: robustness check
Methods Size of the decision set Maximal number of predicates

Australian dataset
Linear Logistic Regression 34.4 1
Random Forest 52.4 × 69.6 8
PLTR 25.4 2
Taiwan dataset
Linear Logistic Regression 78.7 1
Random Forest 2,378.7 × 174.7 29.9
PLTR 79.9 2
Housing dataset
Linear Logistic Regression 17 1
Random Forest 343.8 × 110.5 18.8
PLTR 47.6 2
Note: This table displays the average values of interpretability measures for linear logistic regression, the
random forest method and PLTR.
economic evaluation method consists of estimating the amount of regulatory capital induced
by the estimated probabilities of default. A similar comparison approach was proposed by
Hurlin et al. (2018) for loss-given-default (LGD) models. However, this approach requires
computing other Basel risk parameters, in particular the LGD and the exposure at default
(EAD), and hence needs specific information about the consumers and the terms of the
loans, which is not publicly available.
An alternative approach consists of comparing the misclassification costs (see Viaene
and Dedene, 2004). This cost is estimated from Type 1 and Type 2 errors weighted by their
probability of occurrence. Formally, let CF N be the cost associated with a Type 1 error (the
cost of granting credit to a bad customer) and CF P be the cost associated with a Type 2
error (e.g., the cost of rejecting a good customer). Thus, the misclassification error cost is
defined as
M C = CF P F P R + CF N F N R,
where FPR is the false positive rate and FNR is the false negative rate. There is no
consensus in the literature about how to determine CF N and CF P . Two alternatives have
been proposed. The first method fixes these costs by calibration based on previous studies
(Akkoc, 2012). For example, West (2000) set CF N to 5 and CF P to 1. The second method
evaluates misclassification costs for different values of CF N to test as many scenarios as
possible (Lessmann et al., 2015). Although there is no consensus on how to determine
these costs, it is generally acknowledged that the cost of granting credit to a bad customer
is higher than the opportunity cost of rejecting a good customer (see Thomas et al., 2002;
West, 2000; Baesens et al., 2003, among others). We choose to follow the second approach to
assess the performance of the competing models. We fix CF P at 1 without loss of generality
27
(Hernández-Orallo et al., 2011) and consider values of CF N between 2 and 50. Once these
misclassification costs are computed, we set the linear logistic regression as the reference,
and we compute the financial gain or cost reduction associated with an alternative scoring
model relative to this reference.24
Figures A.1-A.4 in Appendix A display the average cost reduction or financial gains over
the test samples for the four datasets considered above. First, all methods deliver positive
cost reductions, except in three cases. This means that financial institutions relying on each
of these methods rather than on the benchmark linear logistic regression are expected to
save an amount equivalent to the cost of rejecting (accepting) good (bad) applicants. In
view of the large number of credits in bank credit portfolios, these gains could represent
substantial savings for credit institutions. The fact that non-linear logistic regression leads to
an increase in costs compared to the linear logistic regression comes from the relatively high
number of variables in the two datasets (14 and 23 in the Australian and Taiwan datasets,
respectively). This leads to a proliferation of predictors (squares of the variables, cross-
products of the variables) and therefore to overfitting. The penalised version of the non-
linear logistic regression succeeds in dealing with this issue, which materialises in positive
values of cost reductions in all cases except for the Australian dataset. The NN and SVM
both reduce the misclassification costs compared to the logistic regression. This result is
once again consistent with the results of the literature.
Second, across all datasets, the PLTR method is among the most efficient in terms of cost
reduction. For the Kaggle dataset, the cost reduction relative to the linear logistic regression
is equal to 18.06% on average. This result also holds in the Taiwan dataset, with an average
cost reduction of 22.29%. Note that the random forest method leads to lower cost reduction
for these two datasets, with an average cost reduction of 13.09% (11.51%) for the Kaggle
(Taiwan) dataset. This means that although the random forest method has high global
predictive accuracy, as given by the proportion of correct classification (see Tables 1 and 4),
it fails to some extent to detect bad customers, which leads to a relative increase in costs due
to more false negatives.For the other two datasets (Australian and Housing), the random
forest method performs well. With the Australian dataset, the average cost reduction of the
random forest (PLTR) method is equal to 22.71% (14.89%). For the Housing dataset, the
average values are equal to 44.56% and 38.69% for the random forest method and PLTR,
respectively.
We also consider a second measure of performance, namely, the expected maximum profit
(EMP) introduced by Verbraken et al. (2014), to compare the models from an economic
viewpoint. The EMP takes into account the profits received by the non-defaulters and the
losses caused by the defaulters. This allows us to compute an EMP value that is expressed
as a percentage of the total loan amount and measures the incremental profit relative to not
24
The misclassification costs are computed from test samples.
28
building a credit scoring model. The EMP is based on the following utility function of the
decision maker:
P (t; b, c, c∗ ) = (b − c∗ ) π0 F0 (t) − (c + c∗ ) π1 F1 (t)
where t is a cutoff; b is the benefit associated with a true positive; c is the cost associated
with a false positive; c∗ is the cost associated with an individual case; π0 and π1 are the
prior probabilities of non-default and default, respectively; and F0 (t) and F1 (t) are the
corresponding cumulative density functions. The parameters b and c are calibrated using
the LGD and return on investment (ROI, see Verbraken et al., 2014 for more details).25 Since
our datasets do not include information on the LGD or the ROI, to calculate the EMP, we
assume that the LGD distribution is bimodal, with a probability of complete recovery equal
to 0.55 and a probability of complete loss of 0.1, and that the ROI per granted loan is fixed
to 26.44% for all the credits, which corresponds to the value considered by Verbraken et al.
(2014) for their illustrations.
Tables 6 and 7 report the results obtained for the Kaggle dataset and for the three
datasets considered in the robustness analysis. For the Kaggle dataset, the EMP analysis
confirms that the PLTR method generates more profit (0.4387%) than the different versions
of logistic regression (from 0.1910% to 0.3730%). Furthermore, PLTR exhibits similar eco-
nomic performance to random forest (0.4169%), while keeping the intrinsic interpretability
of logistic regression. Similar qualitative results are obtained for the three other datasets of
our robustness analysis. These results confirm those previously obtained with the misclas-
sification cost analysis (Viaene and Dedene, 2004).
Table 6: Economic performance indicator: Kaggle dataset
Methods Expected Maximum Profit (in %)

Linear Logistic Regression 0.1910
Non-Linear Logistic Regression 0.2925
Non-Linear Logistic Regression + ALasso 0.3730
Random Forest 0.4169
PLTR 0.4387
Support Vector Machine 0.2846
Neural Network 0.2384
To conclude, all results show that the PLTR model may generate important cost reduc-
tions compared to the standard logistic regression model generally used by the credit risk
industry while preserving its intrinsic interpretability.
25
To implement the EMP measure, we use the R package EMP (July 2019).
29
Table 7: Economic performance indicator: robustness check
Methods Expected Maximum Profit (in %)

Australian dataset
PLTR 10.1842
Taiwan dataset
PLTR 2.3075
Housing dataset
PLTR 3.3660
30
7 Conclusion
Despite the development and dissemination of many efficient machine learning classification
algorithms, the benchmark scoring model in the credit industry remains logistic regression.
This current state is caused mainly by the stability and robustness of the logistic regression
model and also its intrinsic interpretability. Many academic papers advocate the use of
more sophisticated ensemble methods, such as the random forest method. These black-
box models are not interpretable, but many agnostic methods can be used to make their
forecasting rules interpretable ex post for the various stakeholders (risk modellers, model
checkers, clients, management, regulators, etc.). Nevertheless, these alternative models are
still generally considered challenger models and rarely used in the credit granting process
or for regulatory purposes.
Recognising that traditional logistic regression underperforms random forest due to its
pitfalls in modelling non-linear (threshold and interaction) effects, this article introduces
penalised logistic tree regression (PLTR) with predictive variables given by easy-to-interpret
endogenous univariate and bivariate threshold effects. These effects are quantified by dummy
variables associated with leaf nodes of short-depth decision trees built with couples of the
original predictive variables. Our main objective is to combine decision trees (from the
field of machine learning) and a logistic regression model (from the field of econometrics) to
obtain the best of both worlds: a performing and interpretable hybrid credit scoring model.
Monte Carlo simulations and an empirical application based on four real-life credit scor-
ing datasets show that PLTR has good predictive power while remaining easily interpretable.
More precisely, using several metrics and diagnostic tests to evaluate the accuracy and the
interpretability of credit models, we show that it performs better in out-of-sample than
traditional linear and non-linear logistic regression, while being competitive relative to the
random forest method. We also evaluate the economic benefit of using our PLTR method
through misclassification costs and expected maximum profit analysis. We find that beyond
parsimony, the PLTR method leads to a significant reduction in misclassification costs.
31
A Appendix A: Additional Figures and Tables
20
15
Cost reduction (in %)
10
5
Non Linear Logistic regression
Non Linear Logistic regression+ALasso
Random Forest
PLTR
Support Vector Machine
0 Neural Network
-5
0 5 10 15 20 25 30 35 40 45 50
C FN
Figure A.1: Economic evaluation for the Kaggle dataset
25
20
15
10
-5
Non Linear Logistic regression
-10 Random Forest
PLTR
Neural Network
-15
-20
0 5 10 15 20 25 30 35 40 45 50
C FN
Figure A.2: Economic evaluation for the Taiwan dataset
32
40
20
-20 Non Linear Logistic regression

Random Forest
PLTR
-40
Neural Network
-60
-80
-100
-120
-140
-160
0 5 10 15 20 25 30 35 40 45 50
C FN
Figure A.3: Economic evaluation for the Australian dataset
50
45
40
35
30 Non Linear Logistic regression

Random Forest
25
PLTR
20 Neural Network
15
10
0
0 5 10 15 20 25 30 35 40 45 50
C FN
Figure A.4: Economic evaluation for the Housing dataset
33
Table A.1: Comparison of accuracy tests: Kaggle dataset
Inference procedures LLR vs RF LLR vs PLTR RF vs PLTR

AUC 100 100 70
(RF) (PLTR) (PLTR)
Diebold-Mariano (BS) 100 100 100
(RF) (PLTR) (PLTR)
Diebold-Mariano (Minus Log-Likelihood) 50 100 100
(RF) (PLTR) (PLTR)
LLR RF PLTR
Model Confidence Set (BS) 0 0 100
Model Confidence Set (Minus Log-Likelihood) 0 0 100
Note: The table displays the rejection frequencies of the inference procedures obtained over the N × 2 cross-
validation test samples. The label of the best performing model in each pair is presented below between
parentheses. The inference procedure labelled “AUC” corresponds to a test of AUCs comparison, “Diebold-
Mariano” to the test of Diebold and Mariano (1995), and “Model Confidence set” to the approach of Hansen
et al. (2011). The labels “BS” and “Minus Log-Likelihood” refer to the loss functions used in the tests.
The method labelled “LLR” is the linear logistic regression, “RF” is the random forest, and “PLTR” is the
penalised logistic tree regression model. A significance level of 5% is used for the pairwise tests and of 10%
for the “Model Confidence set”.
Bibliography
ACPR (2020). Governance of artificial intelligence in finance. Discussion papers publication,
November, 2020.
Akkoc, S. (2012). An empirical comparison of conventional techniques, neural networks and

the three stage hybrid adaptive neuro fuzzy inference system (ANFIS) model for credit
scoring analysis: The case of Turkish credit card data. European Journal of Operational
Research, 222(1):168–178.
Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of corporate
bankruptcy. The Journal of Finance, 23(4):589–609.
Athey, S. (2018). The impact of machine learning on economics. In The economics of

artificial intelligence: An agenda, pages 507–547. University of Chicago Press.
Athey, S. and Imbens, G. W. (2019). Machine learning methods that economists should
know about. Annual Review of Economics, 11:685–725.
Baesens, B., Gestel, T. V., Viaene, S., Stepanova, M., Suykens, J., and Vanthienen, J.
(2003). Benchmarking state-of-the-art classification algorithms for credit scoring. Journal
of the Operational Research Society, 54:627–635.
34
Table A.2: Comparison of accuracy tests: robustness check
Australian dataset
Inference procedures LLR vs RF LLR vs PLTR RF vs PLTR
AUC 60 60 10
(RF) (PLTR) (RF)
(RF) (PLTR) (RF)
(RF) (PLTR) (RF)
LLR RF PLTR
Taiwan dataset
LLR vs RF LLR vs PLTR RF vs PLTR
AUC 80 100 70
(RF) (PLTR) (PLTR)
(RF) (PLTR) (PLTR)
(RF) (PLTR) (PLTR)
LLR RF PLTR
Housing dataset
LLR vs RF LLR vs PLTR RF vs PLTR
AUC 100 100 100
(RF) (PLTR) (RF)
(RF) (PLTR) (RF)
(RF) (PLTR) (RF)
LLR RF PLTR
Note: The table displays the rejection frequencies of the inference procedures obtained over the N × 2 cross-
validation test samples. The label of the best performing model in each pair is presented below between
parentheses. The inference procedure labelled “AUC” corresponds to a test of AUCs comparison, “Diebold-
Mariano” to the test of Diebold and Mariano (1995), and “Model Confidence set” to the approach of Hansen
et al. (2011). The labels “BS” and “Minus Log-Likelihood” refer to the loss functions used in the tests.
The method labelled “LLR” is the linear logistic regression, “RF” is the random forest, and “PLTR” is the
penalised logistic tree regression model. A significance level of 5% is used for the pairwise tests and of 10%
for the “Model Confidence set”.
35
Table A.3: Description of the variables in the Kaggle dataset “Give me some credit”
Variable Type Description

SeriousDlqin2yrs Binary The person experienced 90 days
past due delinquency or worse
(Yes/No)
RevolvingUtilizationOfUnsecuredLines Percentage Total balance on credit cards and
personal lines of credit except
real estate and no instalment
debt such as car loans divided by
the sum of credit limits
Age Interval Age of the borrower (in years)
NumberOfTime30-59DaysPastDueNotWorse Interval Number of times a borrower has
been between 30 and 59 days
past due but not worse in the
last 2 years
DebtRatio Percentage Monthly debt payments, alimony
and living costs over the monthly
gross income
MonthlyIncome Interval Monthly Income
NumberOfOpenCreditLinesAndLoans Interval Number of open loans (like car
loan or mortgage) and credit
lines (credit cards)
NumberOfTimes90DaysLate Interval Number of times a borrower has
been 90 days or more past due
NumberRealEstateLoansOrLines Interval Number of mortgage and real
estate loans including home
equity lines of credit
NumberOfTimes60-89DaysPastDueNotWorse Interval Number of times a borrower has
been between 60 and 89 days
past due but not worse in the
last 2 years
NumberOfDependents Interval Number of dependents in family
excluding themselves (spouse,
children, etc.)
36
Table A.4: Description of the variables in the Housing dataset

Bad Binary Whether the consumer had a default on the loan (1) or not (0)
Clage Interval Age of the oldest trade (in months)
Clno Interval Number of trades
Debtinc Interval Ratio of debt to income
Delinq Interval Number of neglectful trades
Derog Interval Number of major derogatory reports
Job Nominal Professional categories
Loan Interval Amount of the loan
Mortdue Interval Amount due on the mortgage
Ninq Interval Number of recent credits enquired
Reason Binary Whether the loan is for debt consolidation (DebtCon) or home
improvement (HomeImp)
Value Interval Current property value
Yoj Interval Number of years at the present job
Table A.5: Description of the variables in the Taiwan dataset

Y Binary default payment (Yes = 1, No = 0)
X1 Quantitative Amount of the given credit (NT dollar)
X2 Binary Gender (1 = male; 2 = female)
X3 Nominal Education (1 = graduate school; 2 = university; 3 = high school;
4 = others)
X4 Nominal Marital status (1 = married; 2 = single; 3 = others)
X5 Quantitative Age (year)
X6-X11 Nominal X6 - X11: History of past payment. We tracked the past monthly
payment records (from April to September, 2005) as follows:
X6 = the repayment status in September, 2005; X7 = the
repayment status in August, 2005; . . .;X11 = the repayment
status in April, 2005. The measurement scale for the
repayment status is: -1 = pay duly; 1 = payment delay for one
month; 2 = payment delay for two months; . . .; 8 = payment
delay for eight months; 9 = payment delay for nine months and above
X12-X17 Quantitative Amount of bill statement (NT dollars). X12 = amount of bill state-
ment in September, 2005; X13 = amount of bill statement in August,
2005; . . .; X17 = amount of bill statement in April, 2005
X18-X23 Quantitative Amount of previous payment (NT dollars). X18 = amount paid in
September, 2005; X19 = amount paid in August, 2005; . . .;
X23 = amount paid in April, 2005
37
Bauweraerts, J. (2016). Predicting bankruptcy in private firms: Towards a stepwise regres-
sion procedure. International Journal of Financial Research, 7(2):147–153.
Bracke, P., Datta, A., Jung, C., and Sen, S. (2019). Machine learning explainability in
finance: an application to default risk analysis. Bank of England, Staff Working Paper
No. 816.
Breiman, L. (1996). Bagging predictors. Machine Learning, 26:123–140.
Breiman, L. (2001). Random forest. Machine Learning, 45:5–32.
Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly

weather review, 78(1).
Bussman, N., Giudici, P., Marinelli, D., and Papenbrock, J. (2019). Explainable
machine learning in credit risk management. Working paper SSRN, available at
http://dx.doi.org/10.2139/ssrn.3506274.
Candelon, B., Dumitrescu, E.-I., and Hurlin, C. (2012). How to evaluate an early-warning
system: Toward a unified statistical framework for assessing financial crises forecasting
methods. IMF Economic Review, 60(1):75–113.
Cardell, N. S. and Steinberg, D. (1998). The hybrid-CART logit model in classification and
data mining. Working paper, Salford-System.
Chan, K. Y. and Loh, W. Y. (2004). Lotus: An algorithm for building accurate and com-
prehensible logistic regression trees. Journal of Computational and Graphical Statistics,
13(4):826–852.
Charpentier, A., Flachaire, E., and Ly, A. (2018). Econometrics and machine learning.
Economics and Statistics, 505(1):147–169.
Coffman, J. (1986). The proper role of tree analysis in forecasting the risk behavior of
borrowers. Management Decision Systems, Atlanta, MDS Reports, 3(4):7.
De Caigny, A., Coussement, K., and De Bock, K. W. (2018). A new hybrid classification
algorithm for customer churn prediction based on logistic regression and decision trees.
European Journal of Operational Research, 269(2):760–772.
Deng, H. (2019). Interpreting tree ensemble with intrees. International Journal of Data
Science and Analytics, 7(4):277–287.
Desai, V. S., Crook, J. N., and Overstreet Jr, G. A. (1996). A comparison of neural
networks and linear scoring models in the credit union environment. European Journal
of Operational Research, 95(1):24–37.
38
Diebold, F. and Mariano, R. (1995). Comparing predictive accuracy. Journal of Business
and Economic Statistics, 13:253–263.
Dietterich, T. G. (1998). Approximate statistical tests for comparing supervised classifica-

tion learning algorithms. Neural Computation, 10(7):1895–1923.
Doshi-Velez, F. and Kim, B. (2017). Towards a rigorous science of interpretable machine

learning. arXiv preprint arXiv:1702.08608.
Durand, D. (1941). Risk elements in consumer installment financing. National Bureau of

Economic Research, New York.
EBA (2020). Report on big data and advanced analytics. European Banking Authority,
January, 2020.
EC (2020). White paper on artificial intelligence: A european approach to excellence and

trust. European Commission, February, 2020.
Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its
oracle properties. Journal of the American Statistical Association, 96(456):1348–1360.
Finlay, S. (2011). Multiple classifier architectures and their application to credit risk assess-
ment. European Journal of Operational Research, 210(2):368–378.
Friedman, J., Hastie, T., and Tibshirani, R. (2010). Regularization paths for generalized
linear models via coordinate descent. Journal of Statistical Software, 33(1):1–22.
Frost, J., Gambacorta, L., Huang, Y., Shin, H. S., and Zbinden, P. (2019). Bigtech and the
changing structure of financial intermediation. Economic Policy.
Grennepois, N., Alvirescu, M. A., and Bombail, M. (2018). Using random forest for credit
risk models. Deloitte Risk Advisory, September, 2018.
Grennepois, N. and Robin, E. (2019). Explain artificial intelligence for credit risk manage-
ment. Deloitte Risk Advisory, July, 2019.
Hand, D. J. (2005). Good practice in retail credit scorecard assessment. Journal of the
Operational Research Society, 56(9):1109–1117.
Hansen, P. R., Lunde, A., and Nason, J. M. (2011). The model confidence set. Econometrica,
79(2):453–497.
Henley, W. and Hand, D. (1996). A k-nearest-neighbour classifier for assessing consumer

credit risk. The Statistician, 45(1):77–95.
39
Henley, W. E. and Hand, D. J. (1997). Construction of a k-nearest neighbour credit-scoring
system. IMA Journal of Mathematics Applied in Business and Industry, 8:305–321.
Hernández-Orallo, J., Flach, P. A., and Ramirez, C. F. (2011). Brier curves: a new cost-
based visualisation of classifier performance. In ICML, pages 585–592.
Hoerl, A. E. and Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthog-
onal problems. Technometrics, 12(1):55–67.
Hurlin, C., Leymarie, J., and Patin, A. (2018). Loss functions for loss given default model
comparison. European Journal of Operational Research, 268(1):348–360.
Hurlin, C. and Pérignon, C. (2019). Machine learning et nouvelles sources de données pour
le scoring de crédit. Revue d’économie financière, (3):21–50.
Landwehr, N., Hall, M., and Frank, E. (2005). Logistic model trees. Machine Learning,
59:161–205.
Lessmann, S., Baesens, B., Seow, H.-V., and Thomas, L. C. (2015). Benchmarking state-
of-the-art classification algorithms for credit scoring: An update of research. European
Journal of Operational Research, 247:124–136.
Lundberg, S. M. and Lee, S.-I. (2017). A unified approach to interpreting model predictions.
Advances in Neural Information Processing Systems, pages 4765–4774.
Makowski, P. (1985). Credit scoring branches out. Credit World, 75(1):30–37.
Martens, D., Baesens, B., and Van Gestel, T. (2008). Decompositional rule extraction from
support vector machines by active learning. IEEE Transactions on Knowledge and Data
Engineering, 21(2):178–191.
Matignon, R. (2007). Data Mining Using SAS Enterprise Miner.
McIlhagga, W. H. (2016). penalized: A matlab toolbox for fitting generalized linear models
with penalties.
Molnar, C. (2019). Interpretable machine learning. A Guide for Making Black Box Models
Explainable. https://christophm.github.io/interpretable-ml-book/.
Mullainathan, S. and Spiess, J. (2017). Machine learning: an applied econometric approach.

Journal of Economic Perspectives, 31(2):87–106.
Óskarsdóttir, M., Bravo, C., Sarraute, C., Vanthienen, J., and Baesens, B. (2019). The value
of big data for credit scoring: Enhancing financial inclusion using mobile phone data and
social network analytics. Applied Soft Computing, 74:26–39.
40
Paleologo, G., Elisseeff, A., and Antonini, G. (2010). Subagging for credit scoring models.
European Journal of Operational Research, 201(2):490–499.
Park, M. Y. and Hastie, T. (2007). L1-regularization path algorithm for generalized linear
models. Journal of the Royal Statistical Society, 69(4):659–677.
Pundir, S. and Seshadri, R. (2012). A novel concept of partial Lorenz curve and partial
Gini index. International Journal of Engineering, Science and Innovative Technology,
1:296–301.
Setiono, R., Baesens, B., and Mues, C. (2008). Recursive neural network rule extraction for
data with mixed attributes. IEEE Transactions on Neural Networks, 19(2):299–307.
Shewchuk, J. R. et al. (1994). An introduction to the conjugate gradient method without

the agonizing pain. Carnegie-Mellon University. Department of Computer Science.
Srinivasan, V. and Kim, Y. H. (1987). Credit granting: A comparative analysis of classifi-

cation procedures. The Journal of Finance, 42(3):665–681.
Steenackers, M. and Goovaerts, J. (1989). A credit scoring model for personal loans. Insur-
ance: Mathematics and Economics, 8(1):31–34.
Stepanova, M. and Thomas, L. C. (2001). Phab scores: Proportional hazards analysis

behavioural scores. The Journal of the Operational Research Society, 52(9):1007–1016.
Tam, K. Y. and Kiang, M. Y. (1992). Managerial applications of neural networks: the case
of bank failure predictions. Management Science, 38(7):926–947.
Thomas, L., Crook, J., and Edelman, D. (2017). Credit scoring and its applications. SIAM.
Thomas, L. C. (2000). A survey of credit and behavioural scoring: forecasting financial risk
of lending to consumers. International journal of forecasting, 16(2):149–172.
Thomas, L. C., Edelman, D. B., and Crook, J. N. (2002). Credit scoring and its application.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society, 58(1):267–288.
Vapnik, V. N. and Chervonenkis, A. Y. (1971). On the uniform convergence of relative

frequencies of events to their probabilities. Theory of Probability and its Applications,
2(16):264–280.
Varian, H. R. (2014). Big data: New tricks for econometrics. Journal of Economic Perspec-
tives, 28(2):3–28.
41
Verbraken, T., Bravo, C., Weber, R., and Baesens, B. (2014). Development and application
of consumer credit scoring models using profit-based classification measures. European
Journal of Operational Research, 238(2):505–513.
Viaene, S. and Dedene, G. (2004). Cost-sensitive learning and decision making revisited.
European Journal of Operational Research, 166:212–220.
Wang, W. (2012). How the small and medium-sized enterprises’ owners’ credit features
affect the enterprises’ credit default behavior? E3 Journal of Business Management and
Economics, 3(2):90–95.
West, D. (2000). Neural network credit scoring models. Computers & Operations Research,
27(11-12):1131–1152.
Yobas, M. B., Crook, J. N., and Ross, P. (2000). Credit scoring using neural and evolutionary
techniques. IMA Journal of Mathematics Applied in Business and Industry, 11:111–125.
Zhang, H. H. and Lu, W. (2007). Adaptive lasso for Cox’s proportional hazards model.
Biometrika, 94(3):691–703.
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American
Statistical Association, 101(476):1418–1429.
Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net.
Journal of the Royal Statistical Society, 67(2):301–320.
42

SSRN Id3553781

Uploaded by

Copyright:

Available Formats

SSRN Id3553781

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3553781

Uploaded by

Copyright:

Available Formats

Machine Learning or Econometrics for Credit Scoring:

Let’s Get the Best of Both Worlds∗

JEL Classification: G10 C25, C53

Second, our approach can be viewed as a systematisation of common practices in the

2 Threshold effects in logistic regression

where β = (β0 , β1 , ..., βp ) ∈ Rp+1 is an unknown vector of parameters. The estimator βb is

with estimates obtained by replacing β with β.

Proportion of Correct Classification (PCC)

Figure 1: Comparison of performances under univariate and bivariate threshold effects:

2.2 Machine learning for non-linear effects

Dm,l,1 (θm,l ) = (xi , yi ) | xi,j < tm,l,j ,

Dm,l,2 (θm,l ) = (xi , yi ) | xi,j ≥ tm,l,j ,

where the parameter estimates θbm,l satisfy

Figure 2: Example of a decision tree for credit scoring

Figure 3: Comparison of performances under univariate and bivariate threshold effects:

3 Penalised logistic tree regression

Figure 4: Illustration of the two-stage splitting process

3.2 PLTR under threshold effects: Monte Carlo evidence

Proportion of Correct Classification (PCC)

Figure 5: Comparison of performances under univariate and bivariate threshold effects:

4 Model performance with a benchmark dataset

4.2 Statistical measures of performance and interpretability

Table 1: Statistical performance indicators: Kaggle dataset

Methods AUC PGI PCC KS BS

Table 2: Measures of interpretability: Kaggle dataset

Lastly, to highlight the interpretability advantage of our method, we report in Table 3

5 Robustness across datasets

Table 4: Statistical performance indicators: robustness check

Methods AUC PGI PCC KS BS

Methods Size of the decision set Maximal number of predicates

Table 6: Economic performance indicator: Kaggle dataset

Methods Expected Maximum Profit (in %)

Methods Expected Maximum Profit (in %)

Figure A.1: Economic evaluation for the Kaggle dataset

Figure A.2: Economic evaluation for the Taiwan dataset

-20 Non Linear Logistic regression

Figure A.3: Economic evaluation for the Australian dataset

30 Non Linear Logistic regression

Figure A.4: Economic evaluation for the Housing dataset

Inference procedures LLR vs RF LLR vs PLTR RF vs PLTR

Akkoc, S. (2012). An empirical comparison of conventional techniques, neural networks and

Athey, S. (2018). The impact of machine learning on economics. In The economics of

Variable Type Description

Variable Type Description

Table A.5: Description of the variables in the Taiwan dataset

Variable Type Description

Breiman, L. (1996). Bagging predictors. Machine Learning, 26:123–140.

Breiman, L. (2001). Random forest. Machine Learning, 45:5–32.

Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly

Dietterich, T. G. (1998). Approximate statistical tests for comparing supervised classifica-

Doshi-Velez, F. and Kim, B. (2017). Towards a rigorous science of interpretable machine

Durand, D. (1941). Risk elements in consumer installment financing. National Bureau of

EC (2020). White paper on artificial intelligence: A european approach to excellence and

Henley, W. and Hand, D. (1996). A k-nearest-neighbour classifier for assessing consumer

Makowski, P. (1985). Credit scoring branches out. Credit World, 75(1):30–37.

Matignon, R. (2007). Data Mining Using SAS Enterprise Miner.

Mullainathan, S. and Spiess, J. (2017). Machine learning: an applied econometric approach.

Shewchuk, J. R. et al. (1994). An introduction to the conjugate gradient method without

Srinivasan, V. and Kim, Y. H. (1987). Credit granting: A comparative analysis of classifi-