A comparative study of forecasting Corporate Credit Ratings

using Neural Networks, Support Vector Machines, and Decision

Trees

Parisa Golbayani1 , Ionuţ Florescu2 , and Rupak Chatterjee3

1
Financial Engineering, School of Business, Stevens Institute of Technology
2
Hanlon Financial Systems Center, Financial Engineering, School of Business,
Stevens Institute of Technology
3
Department of Physics, Center for Quantum Science and Engineering, Schaefer
School of Engineering and Science, Stevens Institute of Technology

July 15, 2020

Abstract

Credit ratings are one of the primary keys that reflect the level of riskiness and reliability of corpo-

rations to meet their financial obligations. Rating agencies tend to take extended periods of time to

provide new ratings and update older ones. Therefore, credit scoring assessments using artificial intel-

ligence has gained a lot of interest in recent years. Successful machine learning methods can provide

rapid analysis of credit scores while updating older ones on a daily time scale. Related studies have

shown that neural networks and support vector machines outperform other techniques by providing

better prediction accuracy. The purpose of this paper is two fold. First, we provide a survey and a

comparative analysis of results from literature applying machine learning techniques to predict credit

rating. Second, we apply ourselves four machine learning techniques deemed useful from previous

studies (Bagged Decision Trees, Random Forest, Support Vector Machine and Multilayer Perceptron)

1
to the same datasets. We evaluate the results using a 10-fold cross validation technique. The results of

the experiment for the datasets chosen show superior performance for decision tree based models. In

addition to the conventional accuracy measure of classifiers, we introduce a measure of accuracy based

on notches called ”Notch Distance” to analyze the performance of the above classifiers in the specific

context of credit rating. This measure tells us how far the predictions are from the true ratings.

We further compare the performance of three major rating agencies, Standard & Poors, Moody’s and

Fitch where we show that the difference in their ratings is comparable with the decision tree prediction

versus the actual rating on the test dataset.

1 Introduction

Machine learning models have been widely applied on a range of applications in financial fields due

to their capability of detecting embedded trends in financial data. For example, asset prices follow a

random, non linear dynamic pattern due to many factors such as political events, economic conditions

and traders behavior (Huang et al., 2005). Despite this non linear nature of price movements, machine

learning techniques provide an accurate estimation of asset prices (Kim, 2003; Huang et al., 2005;

Kanchymalay et al., 2017; Ganguli and Dunnmon, 2017; Culkin and Das, 2017; Feng et al., 2018; Gu

et al., 2019; Sambasivan and Das, 2017). In addition to asset pricing, machine learning techniques

gained a lot of interest in credit rating prediction problem. Credit ratings have been used as an

informative indicators of the level of riskiness of companies and bonds (Huang et al., 2004; Lee, 2007).

They reflect the likelihood that a counterparty will default on his/her financial obligation (Martens

et al., 2010). Rating agencies such as Standard and Poor’s and Moodys analyze various aspects

of companies financial status to come up with these credit ratings (Huang et al., 2004). Such an

assessment is an expensive complicated process often taking months since it involves many experts to

analyze different variables that reflect the underlying corporations’ reliability (Hajek and Michalak,

2013). One solution to reduce this financial and time cost is to come up with a predictive quantitative

model based on the historical financial information of a company.

In this study we provide a comparative analysis of the literature of Corporate credit rating pre-

diction problem. We focus on modern statistical and Artificial Intelligence techniques. We review

and summarize the performance of techniques that were demonstrated to be the most accurate in the

literature. We implement these most accurate techniques and compare their performance when ana-

2
lyzing three different sectors of the US economy. To our knowledge, this is one of the first papers that

implements corporate rating models on the same data and thus compares algorithms’ performance on

an equal footing.

Corporate credit rating is a different problem than the typical image classification techniques.

Thus, we introduce a new evaluation measure: the Notch Distance in section 4.2 on page 23. This

measure allows us to better quantify and distinguish between performance of various machine learning

methods.

This paper is organized as follows. In Section 2, we provide a review of prior literature on the

credit rating problem. Section 3 contains an explanation of each machine learning method studied in

order to predict corporate credit ratings. Section 4 explains the nature of the input and output data

sets in this study as well as the evaluation methodology used. Experimental results and analysis are

provided in Section 5. Final conclusions are drawn in section 6, together with future applications of

the methodology.

2 Prior Methods on the Credit Rating Problem using Ma-

chine Learning Techniques

Bond Rating The majority of literature on the credit rating problem is focused on predicting bond’s

probability of default and thus bond rating (Saha and Waheed, 2017). Garavaglia (Garavaglia, 1991)

simulated S&P corporate bond rating using the unidirectional version of the counter-propagation

neural network. This neural network model introduced by Robert Hetch-Nielsen. Chaveesuk et al.

(Chaveesuk et al., 1999) studied three supervised neural networks concepts, backpropagation, radial

basis function and learning vector quantization. However, they concluded, neither a neural network

nor a regression model performs well on bond rating assignments. Shin and Han (Shin and Han, 2001)

combined inductive learning approach based on decision trees with case indexing process in order to

build a knowledge-based system that can classify Korean corporate bond ratings. Kim & Han (Kim

and Han, 2001) clustered financial data among bond rating groups. They used competitive artificial

neural network to generate the centroid values of the clusters. They perform a case-indexing study

based on these cluster information and combine it with the classification technique in order to present

a hybrid model for forecasting corporate bond ratings. They apply their method on 167 financial

variables including five categorical variable and 162 financial ratios. After performing some statistical

3
tests such as ANOVA, factor analysis and step-wise method, they select 13 financial ratios as the final

input data. Abdou (Abdou, 2009) studied feed forward neural network to evaluate consumer loans

within the Egyptian private banks. Lessman et al. Lessmann et al. (2015) provided a deep analysis

of machine learning techniques used in forecasting retail credit scoring. They studied 41 classification

algorithms used to predict probability of default for retail loans. The authors used multiple datasets

and criteria for evaluation and concluded that their hybrid method is superior to other methods.

Gangolf et al. Gangolf et al. (2016) provided a comparison of statistical and artificial intelligence

techniques used for forecasting probability of default for bonds. In case of predicting probability

of default, Angelini et al (Angelini et al., 2008) studied two different neural networks to evaluate

credit risk of Italian small businesses. They showed that neural network architectures can be useful

for estimating the probability of default of borrowers. Tsai et al. (Tsai and Wu, 2008) studied the

application of multilayer perceptron network on bankruptcy prediction and credit risk evaluation.

Their analysis proved that single neural network provides accurate results compared to multi-layers

classifiers. Addo et al. (Addo et al., 2018) studied seven binary classifiers (elastic net, random forest,

gradient boosting model and four different neural network architectures) in predicting loan default

probability. They implemented these models on 117019 data points in the year 2016 and 2017 provided

from European banks. They used 60% of data as training set, 20% as validation and 20% as the test

set. They compared the performance of these models on a selected subset of important features based

on variable importance of each model. They concluded three important points. First, that decision

tree based models outperform multilayer artificial neural networks. Second, that standard performance

criteria such as AIC or R2 are not sufficient enough to compare different models. Third, that the

selected subset of important features based on variable importance of models, does not necessarily

provide stable results.

Corporate credit rating In this work we are concerned with corporate credit rating. Although,

the two problems are related, the bond ratings are not necessarily the same as the issuing company’s

corporate rating due to the priority of claim of various bonds. There are many more bonds than there

are corporations and thus the data availability is much more limited. When rating publicly traded

corporations the input is based generally on the 10-Q and 10-K statements. However, bonds are traded

and thus bond values can be marked to market far more frequently than quarterly. Furthermore,

while the basic bond structure is similar for all issuing companies regardless of sector, the quarterly

4
statement issued by a bank is fundamentally different from that issued by a company from say the

Consumer Staples sector. Thus papers dedicated to corporate credit rating are generally constructing

models for companies from a specific sector. The lack of large data sets is thus typical of this domain

and this makes the credit rating problem challenging.

Methodologically, statistical models and artificial intelligence approaches are used in the field

of credit rating prediction. In this section, we provide a review of prior and relevant literature of

models used. We start from standard models and later present more recent and complex models. We

summarize these studies in Tables 1, 2 and 3. We are basing our subsequent analysis on the accuracy

numbers in these tables to determine which models will be implemented on the financial statements

data.

Statistical Methods Standard statistical techniques such as logistic regression analysis (LRA) and

discriminant analysis (DA) have been used to construct credit scoring models (Horrigan, 1966; West,

1970; Kaplan and Urwitz, 1979; Kim et al., 1993; Kamstra et al., 2001; Oelerich and Poddig, 2006;

Abdou, 2009; Akkoç, 2012; Abdou et al., 2016). Modified version of these techniques have been de-

veloped to improve credit scoring models. (Hwang et al., 2009, 2010) modify the usual ordered probit

model by replacing the linear regression function with a semi-parametric function to increase the

choices of regression function. They apply this prediction method on four market driven variables,

nineteen accounting variables and industry indicator variables. Their conclusion shows that an or-

dered semiparametric probit model (OSPM) provides a better prediction than ordered linear probit

model (OLPM). However, different assumptions behind most of these statistical approach, such as lin-

earity, normality and distribution-dependency limit their predictive capability comparing to Artificial

Intelligence techniques (Huang et al., 2004; Pai et al., 2015).

Nonetheless, recent papers are still using traditional statistics methods and new developments.

Gogas et al. (2014) used ordered probit regression model to forecast long term bank credit rating.

The benchmark was Fitch credit rating in 2012. Irmatova (2016) introduced a new rating model

based on principal component analysis and k-means clustering algorithm to predict 30 country’s

rating. Karminsky and Khromova (2016) implemented an ordered probit regression model to predict

bank ratings using the Bankscope database. Hajek et al. (2016) developed a methodology based on

latent semantic analysis to extract topical content from company-related documents. The authors

use naive bayesian network to forecast corporate credit ratings. Recently, (Petropoulos et al., 2016)

5
developed a methodology based on Student’s-t Hidden Markov Models (SHMMs) to investigate the

temporal dimension and heavy tailed distribution of firm-related data.

We next present three prevalent AI techniques used for corporate credit rating: Artificial Neural

Networks (ANN), Support Vector Machines (SVM) and Classification and Regression Trees (CART).

Artificial Neural Networks (ANN) Artificial Neural Networks draws attention from many re-

searchers with its robustness, flexibility and higher accuracy (Dutta and Shekhar, 1988; Surkan and

Singleton, 1990; Kim et al., 1993; Maher and Sen, 1997; Kwon et al., 1997; Brennan and Brabazon,

2004; Abdou, 2009; Abdou et al., 2016). West (West, 2000) implemented five neural network architec-

tures: multi-layer perceptron, mixture-of-experts, radial basis function, learning vector quantization

and fuzzy adaptive resonance. He compared the results of these neural networks with linear discrim-

inant analysis, logistic regression, k nearest neighbor, kernel density estimation, and decision trees.

They studied German and Australian credit scoring data and conclude that multi-layer perceptron

is not the most accurate neural network architecture and logistic regression outperforms all neural

network architectures. Kim (Kim, 2005) applied adaptive learning networks (ALN) which is a non-

parametric model, on both financial and non financial variables to predict S&P credit ratings. Yu et

al. (Yu et al., 2008) proposed a six stage neural network ensemble learning model to assess a credit risk

measurement on Japanese consumer credit card applications approval and UK corporations. Kashman

(Khashman, 2010) investigated three neural networks based on back propagation learning algorithm

on German Credit Approval dataset. The architecture of these neural networks are different accord-

ing to various parameters used in the model such as hidden units, learning coefficients, momentum

rate and random initial weight range. Pacelli and Azzollini (Pacelli and Azzollini, 2011) provided an

overview of different types of neural networks used in credit rating literature.

Support Vector Machines Among all artificial intelligence techniques, Support Vector Machine

(SVM) has indicated a powerful classification ability (Cortes and Vapnik, 1995; Kim and Sohn, 2010;

Vapnik, 2013; Xiao et al., 2016). However the standard SVM, addresses the two-class classification

problems. For multi-class classification problems, SVM can be represented as several two class tasks

such as one-against-one (OAO-SVM) (Schölkopf et al., 1999), one-against-all (OAA-SVM) (Bottou

et al., 1994) and direct acyclic graph (DAG-SVM) (Platt et al., 2000; Wang et al., 2009). Cao et al.

(Cao et al., 2006) studied all these three approaches on US companies from the manufacturing sector.

Hajek & Olej (Hájek and Olej, 2011) studied Support Vector Machines with supervised learning as well

6
as kernel-based approach with semi-supervised learning in order to predict corporate and municipal

credit ratings. Kim & Ahn (Kim and Ahn, 2012) studied a new type classifier based on Support

Vector Machine, called (OMSVM) which applies ordinal partwise partitioning (OPP) to extend the

binary SVM.

Decision Trees According to the literature, some researchers focus on the higher level of inter-

pretability and feature extraction flexibility of decision tree-based approach in multi-class classification

problems (Lee et al., 2006; Kao et al., 2012; Abdou et al., 2016). The fundamental of these techniques

are introduced first by (Breiman et al., 1984). Paleologo et al. (Paleologo et al., 2010) proposed a

missing data imputation method and ensemble classification technique, subagging, particularly for

unbalanced credit rating data sets. The proposed subagging technique is based on different models

such as decision trees and support vector machines. Khandani et al. (Khandani et al., 2010) studied

general classification and regression tree technique (CART) on a combination of traditional credit

factors and consumer banking transactions to predict consumer credit risk. Pedro Veronezi (Veronezi,

2016) applied Random Forest(RF) and Multi Layer Perceptron (MLP) to predict corporate credit

ratings using their financial data. First, he used RF to perform feature selection and then the result

was fed to two different RF models. One model aims to predict the changes and another model

predicts the direction. The result of these two models have been added to the input dataset. After

creating a new dataset, two models were trained: a MLP model and a RF model. He concluded that

RF outperforms MLP by providing a more accurate and stable results in a shorter period of time.

Hybrid Models In addition to all these frequently used techniques, some researchers study other

approaches to provide a credit scoring model. Peng et al. (Peng et al., 2011) introduced three

multiple criteria decision making (MCDM) methods to evaluate classification algorithms for financial

risk prediction. Chen (Chen, 2012) investigated rough set theory (RST) approach to classify Asian

banks’ ratings. Some researchers take one step further and integrate multiple techniques to achieve a

higher accuracy. Yeh et al. (Yeh et al., 2012) combined random forest feature selection with different

approach such as rough set theory (RST) and SVM. They tested these techniques on 17 financial

variables in addition to the market based information which has been obtained by Moody’s KMV

tool for each corporation. Wu & Hsu (Wu and Hsu, 2012) introduced an enhanced decision support

model which is a combination of relevance vector machine and decision tree. Hajek (Hájek, 2012)

used fuzzy rule-based system adapted by a feed-forward neural network to classify credit ratings of

7
US municipalities and companies. Pai et al. (Pai et al., 2015) used Decision Tree Support Vector

Machine (DTSVM) integrated to multiple feature selection strategies and rough set theory (RST) to

predict credit ratings. Description for other hybrid techniques are available in (Chandra et al., 2009;

Akkoç, 2012; Chen and Cheng, 2013).

Evaluation and Comparison The Comparison of the discussed techniques with different param-

eters and architectures on several datasets, have been widely studied in credit scoring field. Huang et

al. (Huang et al., 2004) studied support vector machines (SVM) and back-propagation neural network

(BNN) to obtain more accurate credit rating predictions for Taiwan financial institutes and United

States commercial banks. Lee (Lee, 2007) used 5-fold cross-validation to find the optimal parameters

of SVM, multiple discriminant analysis (MDA), case-based reasoning (CBR), and three-layer fully

connected back-propagation neural networks (BPNs) on Korean Datasets. Ye et al. (Ye et al., 2008)

compared the performance of four different approaches: Linear Regression, Bagged decision tree with

Laplace smoothing, SVM and Proximal SVM (PSVM) over four industries manufacturing, wholesale,

retail and services. Ryser & Denzler (Ryser and Denzler, 2009) used k-fold cross validation proce-

dure to compare the performance of different credit rating models. They concluded that although

non-parametric approaches such as random forest, neural network and generalized additive models

still outperform the parametric models, they may overfit during the training process. Hajek & Micha-

lak (Hajek and Michalak, 2013) studied the performance of supervised machine learning techniques

such as Multilayer Perceptron (MLP), Support Vector Machines (SVM), radial Basis Function Neural

Network (RBF), Naive Bayes (NB), Random Forest (RF), Linear Discriminant Classifier (LDC) and

Nearest Mean Classifier (NMC) on both US and European non-financial companies. They focused

on 81 financial and non-financial variables of 852 US companies and 43 financial and non-financial

variables of 244 European companies. They reduced the dimension of data by applying feature se-

lection techniques such as wrapper algorithms. They concluded that US credit ratings are more

affected by the size of companies and market value ratios while the European ratings depend more

on profitability and leverage ratios. Zhong et al. (Zhong et al., 2014) provided a comparison study

over reliability, overfitness and error distribution of different learning algorithms such as SVM, Back

propagation neural network, Extreme Learning Machine (ELM) and Incremental Extreme Learning

(IELM). They considered 6 financial variables on both S&P and Moody’s rating categories to classify

the US corporate credit ratings.

8
 Bahrammirzaee (Bahrammirzaee, 2010) provided a comparative analysis of the application of three

categories of techniques: artificial neural networks, expert systems and hybrid intelligent systems on

credit evaluation. Their analysis categorized articles where artificial neural network outperform other

methods as well as articles where neural network performs as same as or even worse than other tra-

ditional methods such as decision trees and linear regression. Dima et al. (Dima and Vasilache,

2016) compared the performance of neural networks with Bayesian approach on a large sample of

companies (3000 companies) applying for credit to the same bank in Romania. Khemkhem et al.

(Khemakhem and Boujelbene, 2015) implemented neural networks and linear discriminant analysis to

forecast Tunisian corporate credit ratings. They concluded neural networks outperform discriminant

analysis. Hajek et al. (Hájek and Olej, 2014) compared the performance of artificial immune classi-

fication algorithms (AIRS1, AIRS2, AIRS2-p and CSCA) and several well-known machine learning

techniques (MLP, Decision tree, SVM and RBF) on investment and non-investment grade firms. Both

groups of input variables were collected for 520 U.S. companies in the year 2010, 195 classified as in-

vestment grade and 325 as non-investment grade. They used accuracy and misclassification cost as

performance measures of the studied classifiers. They also provided confusion matrix for each classifier

(TP, TN, FP, FN). First, they conclude that AIRS2 outperforms AIRS1, AIRS2-p and CSCA on the

test set based on the studied measures. They also show that MLP and Decision tree are the best

models by providing the highest accuracy and the lowest misclassification cost, respectively.

Wallis et al. (Wallis et al., 2019) provided a comparative analysis of several most popular machine

learning techniques (multinomial logistic regression, linear discriminant analysis, regularized discrim-

inant analysis,artificial neural network, support vector machine, gaussian process classifier, random

forest, gradient boosting machine) to predict Moody’s lone term credit rating. They studied these

models on 308 of the S&P 500 companies from January 2016 to November 2017. They concluded

the top three performing models were random forest, artificial neural network and support vector

machines where random forest outperforms the others with 64.6% accuracy.

Moscatelli et al. (Moscatelli et al., 2019) applied random forest, gradient boosted trees, logistic

regression, penalized logistic regression and linear discriminant analysis to predict default probability

of Italian non-financial firms from 2011-2017. They studied economic and financial ratios as potential

indicators of corporate defaults. They also added a set of credit behavioral indicators on the firm-

bank relationship to the studied financial ratios. They concluded random forest outperform all other

models in predicting default/non-default credit rating.

9
 A summary of discussed approach including number of financial variables, input dataset, output

data set provider and the corresponding accuracy is provided in tables below. Table 1 refers to ANN-

based approach. Table 2 addresses SVM-based models and Table 3 represents Decision tree-based

methods.

10
 Reference Year Num. of Input Var. Input Data Credit Rating Provider Accuracy

(Garavaglia, 1991) 1991 87 797 Companies S&P 85%

(Moody and Utans, 1994) 1994 10 196 Indudtrial Firms S&P 36%

(Huang et al., 2004) 2004 5 36 Commercial Banks S&P 80.00%

(Huang et al., 2004) 2004 6 & 16 Taiwan Dataset Taiwan Ratings Corporation 75.68%

(Kim, 2005) 2005 1080 US companies S&P 84%

(Brabazon and O’Neill, 2006) 2006 8 791 Non-financial US Firms S&P 85%

(Cao et al., 2006) 2006 17 US Manufactoring Companies S&P 80.28%

(Lee, 2007) 2007 10 3017 Korean Companies Korea Information Service, Inc 59.93%

(Yu et al., 2008) 2008 13 Jap. Consumer Credit Card Application Approvals 88.08%

(Yu et al., 2008) 2008 12 60 UK Corporations UK Ratings 85.87%

11
(Khashman, 2010) 2010 24 German Dataset German Ratings 85.9%

(Kim and Sohn, 2010) 2010 10 Korean SMEs Korean Default Rates 64.23%

(Hájek, 2012) 2012 52 852 US Mining Companies S&P 80%

(Hájek, 2012) 2012 14 154 US Municipalities Moody’s 71.62%

(West, 2000) 2000 24 German and Australian Data German and Australian Rating error:0.2243
Table 1: Comparison results for ANN approach

(Khemakhem and Boujelbene, 2015) 2015 15 86 Tunisian companies Tunisian credit rating 82.55%

(Zhao et al., 2015) 2015 20 German Credit Dataset German Credit Ratings 87%

(Hájek and Olej, 2014) 2014 20 Investment and non-investment grade S&P 88.44%

(Addo et al., 2018) 2018 181 and 10 117019 companies asking for a loan European Bank RMSE: 0.044

(Daniel et al., 2019) 2019 17 Random US companies S&P and Moody’s 83%

(Wallis et al., 2019) 2019 27 308 companies from S&P 500 Moody’s 63.6%
 Reference Year Num. of Input Var. Input Data Credit Rating Provider Accuracy

(Huang et al., 2004) 2004 14 36 Commercial Banks S&P 80.00%

(Huang et al., 2004) 2004 6 Taiwan Dataset Taiwan Ratings Corporation 79.73%

(Cao et al., 2006) 2006 17 US Manufactoring Companies S&P 84.61%

(Lee, 2007) 2007 10 3017 Korean Companies Korea Information Service, Inc 67.22%

(Ye et al., 2008) 2008 33 Four US industries Moody’s 64%

(Yu et al., 2008) 2008 13 Jap. Consumer Credit Card Application Approvals 79.91%

(Yu et al., 2008) 2008 12 60 UK Corporations UK Ratings 77.84%

12
(Kim and Sohn, 2010) 2010 10 Korean SMEs Korean Default Rates 66.16%

(Hájek and Olej, 2011) 2011 5 852 US firms S&P 85.96%

(Hájek and Olej, 2011) 2011 12 452 Czech municipals local experts 89.76%

(Kim and Ahn, 2012) 2012 14 Korea Dataset Korea Ratings Corporation 67.98%

(Yeh et al., 2012) 2012 18 2470 Taiwanese companies Taiwanese Rating 74.4%
Table 2: Comparison results for SVM approach

(Hájek and Olej, 2014) 2014 20 Investment and non-investment grade S&P 87.28%

(Pai et al., 2015) 2015 18 Taiwan Dataset Taiwan Ratings Corporation 86%

(Wallis et al., 2019) 2019 27 308 companies from S&P 500 Moody’s 60.1%
 Reference Year Num. of Input Var. Input Data Credit Rating Provider Accuracy

(Shin and Han, 2001) 2001 12 Korean Companies Korean Ratings 69.3%

(Koh et al., 2004) 2004 74.2%

(Ye et al., 2008) 2008 33 Manufacturing/Wholesale/Retail/Services Moody’s 47.1%

(Hájek and Olej, 2014) 2014 20 Investment and non-investment firms S&P 86.13%

13
(Veronezi, 2016) 2016 33 Healthcare, Financial, Technology Sectors in USA S&P and Moody’s > 70%

(Addo et al., 2018) 2018 181 and 10 117019 companies asking for a loan from a bank European Bank RMSE: 0.044

(Wallis et al., 2019) 2019 27 308 companies from S&P 500 Moody’s 64.6%

(Moscatelli et al., 2019) 2019 26 Italian non-financial firms Italian Credit Register > 84%
Table 3: Comparison results for Tree-based approach
3 Analytical Methods

Supervised learning methods are based on learning from observations. Let X ⊂ Rn and Y be the

input and output space respectively. The output space identifies the classification type . When

Y = {−1, +1}, the problem is a binary classification. If Y = {1, 2, .., m}, the problem is a multiple

class classification. Y ⊂ Rn leads to a regression type problems. During the learning process, the

classifier assumes there exists a hidden function Y = f (x) in a training set L = {(xi , yi )}N
i=1 . The

classifier attempts to find a prediction for this function such that the error is minimized.

There are multiple ways to determine the exact nature of function f . In computer science and

statistics these are generically named machine learning algorithms. We are focusing an a few specific

machine learning algorithms described in this section. The previous section and the summary Tables

1, 2, and 3 detail these methods as the most popular in credit rating problem. This is the reason we

chose them for implementation.

3.1 Bagged Decision Trees (BDT)

Let L be the learning/training set consists of N instances. Bagging or bootstrap aggregation algorithm

creates B bootstrap samples where each sample consists of n < N randomly chosen instances but with

replacement from L. Each of these B bootstrap samples train the same classifier which in this case is

a decision tree. In order to predict an outcome for each instance of the testing set, the new case must

be run down each of the B decision trees. The prediction is obtained by a maximum number of votes

among the B classifiers.

This ensemble structure combines predictions of multiple classifiers to provide a better performance

than unique learner. The combination of learners in bagging structure reduces the risk of classification

error (variance) for those unstable high variance classifiers such as decision trees.

In Bagged Decision Trees (BDT), the possibility of over-fitting of individual tree is less concerned.

For this reason, the individual decision trees are grown deep and the trees are not pruned. Bagged

Decision tree has only two hyper-parameters, the number of trees and the number of samples. However,

they may have many similarities which leads to a high correlation in predictions. The structure of

Random Forest will address this problem. A more detailed description of these methods can be found

in (Breiman et al., 1984).

14
3.2 Random Forest (RF)

Random Forest (RF) is another algorithm based on ensemble of classification trees which is developed

by Breiman (Breiman et al., 1984; Breiman, 2001). The only difference between RF and BDT is that

RF takes one extra step. In addition to taking the random subset of data, it also chooses randomly

a subset of X at each node and calculate the best split at that node only within the given subset of

X. This structure provides uncorrelated or weakly correlated predictions. Also, there is no pruning

step, which means all the trees of the forest are grown deep. RF has only two hyper-parameters,

the number of variables in the random subset at each node and the number of trees in the forest.

Moreover, RF ranks variables by the importance of a variable based on the classification accuracy,

while considering the interaction between variables.

3.3 ANN: Multi Layer Perceptron (MLP)

A perceptron is a linear classifier that separates two classes with a straight line. Let X be the input

vector and y be a perceptron that produces a single output:

Xn
y = Φ( wi xi + b) = Φ(wT x + b)
i=1

where w is a vector of input weights, b is a bias and Φ denotes a non-linear activation function.

A multilayer perceptron is a deep, artificial neural network composed of more than one perceptron.

In this NN, the input layer receives the input vector, the output layer makes a decision and provide

a prediction, and in between these two, there exists a number of hidden layers. The input signal

moves forward from the input layer to the output layer, through the hidden layers. Let l represents

the number of input units, m represents the number of hidden units and n represents the number of

output units. Each hidden layer unit is defined as:

(1) (1)
X
hj = f ( wij xi + w0j ) = f (w(1) x)
i→j

where f is an activation function on hidden layers, i = {1, 2, · · · , l} and j = {1, 2, · · · , m}. The
ez
common choice for f is a logistic function f (z) = 1+ez . this hoice is important and depends on the

nature of the data.

15
 Similarly, the output units are denoted as:

(2) (2)
X
yk = g( wjk hj + w0k )
j→k

where g is another activation function and k = {1, 2, · · · , n}. Again, the common choices are logistic

functions.

Considering the layers:

m l
(2) (2) (1) (1)
X X
yk (x, (w(1) , w(2) )) = g(w0k + wjk f ( wij xi + w0j ))
j=1 i=1

The number of parameters required to build this neural network is :

l.m + m + m.n + n = (l + 1)m + (m + 1)n

So depending on the weights and the choice of activation functions, different values of outputs are

obtainable.

When the decision of the output layer is computed against the real outputs, the error will be

calculated. The parameters of the model, the weights and the bias terms, will be adjusted during the

training in order to minimize the error.

∆w = η ∗ ek ∗ xjk

where ek represents the error on the output unit k and xjk is the input that caused the error and η

is the learning rate which defines how much to change the weight to correct the error.

Partial derivative of the error w.r.t. the weights and biases are back propagated through the MLP.

Therefore the change of weights in the hidden layer follows the previous formula except the error term

which will be computed as:

ej = wj ∗ ek ∗ d(yk )

where d is the transfer derivative. The network keeps doing these forward and backward passes until

the error can go no lower. This state is called convergence.

16
3.4 Support Vector Machines (SVM)

SVM is an algorithm that implements non linear boundaries between classes by transforming the input

data into a high dimensional space. This mapping into a new space is a task of kernel functions which

make the input data set linearly separable. In the new space, SVM constructs a maximal margin

hyperplane which provides a maximum separation between output classes. The training observations

that are closest to the maximal margin hyperplane are called support vectors. It introduced first by

Vapnik in a form of quadratic optimization problem with bound constrains and one linear equality

constraints. The standard SVM classifier is only applicable on binary type classification problems.


wT Φ(xi ) + b ≥ +1

 if yi = +1

wT Φ(xi ) + b ≤ −1

 if yi = −1

such that w and b represent the weight vector and bias respectively. Φ() : Rn → Rnk represents a

mapping function and wT Φ(xi ) + b = 0 the optimal separating hyperplane.

The formula above is equivalent to

yi [wT Φ(xi ) + b] ≥ +1

for ∀i = {1, 2, · · · , N }

This formula constructs two hyperplanes on the opposite sides of the optimal hyperplane with the
2
total margin size of ||w||2 .

The classifier takes a decision then based on a decision function

sgn(wT Φ(xi ) + b)

In order to solve the linearly non-separable problems, it is general to use a slack variable ξ which allows

a miss-classification error. The optimization problem containing the weighted miss-classification error

will be formed as:

17
 N
1 X
min wwT + C ξi
w,b,ξ 2 i=1

yi [wT Φ(xi ) + b] ≥ 1 − ξi

 i = 1, 2, · · · , N
subject to

ξi ≥ 0
 i = 1, 2, · · · , N

where C ∈ R+ is a tuning hyperparameter that weights the importance of miss-classification errors.

This optimization problem is solvable using Lagrangian where the Lagrangian multiplier α exists for

each training observation. The non-zero α’s represent the support vectors in the training set.

Therefore, the formula above leads to a dual problem:

1 T
max α Qα − eT α
α 2

0 ≤ αi ≤ C
 i = 1, 2, · · · , N
subject to
y T α = 0



where e is the vector of ones, Qij = yi yj K(xi , xj ) is a N ∗ N positive semi-definite matrix and

K(xi , xj ) = Φ(xi )T Φ(xj ) is a kernel function. The choice of kernel functions depends on the type of

the problem in hand. The linear kernel function is K(xi , xj ) = xTi xj , the radial basis kernel function

is K(xi , xj ) = exp{−γ||xi − xj ||2 }, the polynomial kernel function with degree d is K(xi , xj ) =

(γxTi xj + r)d and the sigmoid kernel is K(xi , xj ) = tanh{γxTi xj + r}

The final SVM classifier will be in a form of:

N
X 
sgn αi yi K(x, xi ) + b
i

3.5 Multi Class Support Vector Machines

The conventional SVM explained above is a binary classifier. For multi-class classification problems,

we need to modify its structure. There are different approaches that extends binary SVM structure to

solve the multi-class classification problems, two of which have been explained and used in this study.

3.5.1 One versus All Approach

Consider an M-class classification problem. The problem can be decomposed into M binary subprob-

lems. One-against-all approach constructs M binary SVM classifiers, each of which separates one class

18
from the other M-1 classes. Let N be the number of training samples (x1 , y1 ), (x2 , y2 ), . . . , (xN , yN ).

The i−th SVM classifier is trained with all the training examples of the i−th class with positive labels,

and all the others with negative labels (Liu and Zheng, 2005; Kim and Ahn, 2012).

3.5.2 One versus One Approach

M (M −1)
This approach constructs binary SVM classifiers for all pairs of classes; in total, there are 2

pairs. For each given pair, binary SVM classifier follows the optimization problem above to maximize

the margin between classes (Kim and Ahn, 2012).

4 Research Methodology

In order to have a fair assessment of the methodologies developed to predict credit rating, we need

to have an equal evaluation footing for the methods. To do so we need to use the same dataset with

the same structure of training, validation and testing. Second we need to have a fair measure for

performance evaluation. We introduce a so called Notch measure in section 4.2 that we believe it is

appropriate for assessing the performance of credit rating algorithms. In our applications we focus on

methods described in section 3. These are the methods used by the majority of credit rating literature

(see Tables 1, 2, and 3).

4.1 Data sets

Previous Studies show that machine learning techniques are able to predict corporate credit ratings

using their historical financial variables (Huang et al., 2004; Hájek, 2012; Pai et al., 2015). In this

study, we consider 52 stocks of financial sector, 28 stocks of energy sector and 44 stocks of healthcare

sector. The input data set covers these corporate historical financial variables from 1990-2018 for

financial sector and from 2009 to 2018 for energy and healthcare sectors. These variables are taken

from both Bloomberg and Compustat and have been merged together to reduce the number of missing

values. Among all available variables, we selected 16 financial variables for financial sector and 20

variables for both energy and healthcare sectors in order to predict the credit ratings. The primary

factors in the selection of these variables are the availability of data and the influence of a variable

on the credit rating based on previous studies (Huang et al., 2004). The output data which are the

corresponded credit ratings are taken from S&P and it contains 19 rating classes starting from AAA

19
to CC. These variables are listed in Tables 4,5. Table 6 and Figure 1 present the distribution of each

sector with respect to rating classes.

Table 4: Financial variables used in financial sector

Variables

V1 Asset/Equity

V2 Total Common Equity

V3 Total Asset

V4 Total Invested Capital

V5 Total Debt/Total Equity

V6 Total Debt/Total Asset

V7 Total Liabilities

V8 Sgort and Long term Debt

V9 Long Term Borrow

V10 Return on Asset

V11 Debt/Market Cap

V12 Operating Margin

V13 IS-OPER-INC

V14 Net Income

V15 Profit Margin

V16 EPS-FOR-RATIOS

20
Table 5: Financial ratios used in energy and healthcare Sectors

Ratio

R1 Debt/EBITDA

R2 FFO/Total Debt

R3 EBITDA/Interest

R4 FFO/Interest

R5 CFO/Debt

R6 FFO/Net Profit

R7 NWC/Revenue

R8 Current Asset/Current Liabilities

R9 (FFO+Cash)/Current Liabilities

R10 EBITDA/Revenues

R11 Cash/Total Debt

R12 Total Debt/Tangible Net worth

R13 Total Debt/Revenue

R14 Debt/Capital

R15 Cash/Asset

R16 Total Fixed Capital/Total Fixed Assets

R17 Equity/Asset

R18 NWC/Total Assets

R19 Retained Earnings/Total Assets

R20 EBITDA/Total Assets

21
Table 6: Number of Companies in each class of rating on December 2018

Rating Classes Financial Sector Energy Sector Healthcare Sector

AAA 1 1 6

AA+ 3 1 1

AA 12 1 5

AA- 24 4 5

A+ 30 4 10

A 34 7 16

A- 36 11 18

BBB+ 33 15 19

BBB 18 18 20

BBB- 12 16 17

BB+ 6 10 14

BB 4 9 13

BB- 2 5 6

B+ 1 3 3

B 1 1 2

B- 1

CCC+ 1

CCC 1

CCC-

CC 1

Figure 1: Distribution of credit ratings for each sector

22
4.2 Evaluation Methodology

The typical way to evaluate a machine learning model is the standard accuracy as the proportion of

observations correctly classified in the test data. For example, the column “Accuracy” in Tables 1,

2, and 3 present a summary of these percentages in prior work. A quick inspection of these numbers

appear to say that accurate predictions based on publicly available financial variables using machine

learning techniques is achievable. However, predicting credit rating is a problem that goes beyond

most other classification problems.

First, the credit rating data is temporal. That is, a prediction has to be done at a point in time

and data from future quarters may not be used in prediction. Most techniques divide data into 80%

training and 20% test data sampled at random which means invariably that future observations will

be used in the model development. We will investigate the models performance when the test data is

made of future observations.

Second issue is that corporate credit ratings do not change very often. Due to this, in about 90%

of observations the credit rating does not change with respect to the previous quarter and in only

10% of observations the credit rating changes. Therefore, creating a simple model that just predicts

the same credit rating as the previous quarter will have a 90% accuracy! This beats the majority of

the results in Tables 1, 2, and 3. Thus, we believe it is crucial to investigate the performance of the

classifier on those observations where the rating is changing from the previous quarter.

Third issue is that the proportion of observations correctly classified tells us nothing about how far

from the true rating were the inaccurate predictions. We quantify the distance between predictions

and real ratings using what is known in literature as Notches distance. To create a numerical value of

the distance we sort the ratings from best (AAA) to worst (CC). We give a numerical value to each

rating (e.g., AAA is 1)

Definition 1 (Notch Distance). To formulate this distance, we let y denote the true rating of an

observation. Let ŷ denote the prediction given by a particular model. Then define a random variable:

Notch Distance as Y = ŷ − y. To asses the distribution of this variable for a particular model that

makes predictions ŷ1 , . . . , ŷN , we calculate F the frequency of the notch i value:

X
F(i) = I(ŷk −yk =i) /N
k∈N

where N is the total number of observations in the test set, i ∈ {· · · , −1, 0, 1, · · · }, and I is the

23
indicator function taking value 1 every time the condition is satisfied and 0 otherwise.

In the definition above F (0) is the proportion of observations predicted correctly. To create a

measure of accuracy we can use the random variable created. Specifically, we can calculate the

expected value:
X
DC = E[Y ] = i ∗ F(i)
i

and we may regard it as a Dissimilarity Coefficient. The numerical value is the expected notch distance

corresponding to a particular method. In a similar way we may define an Absolute Dissimilarity

Coefficient
X
ADC = E[|Y |] = |i| ∗ F(i)
i

We can of course calculate the standard deviation of the variable Y and that would be a measure of

how far the typical prediction is from the actual rating.

Theorem 1 (Jensen’s Inequality). Let X be a random variable with finite expectation, and let g :

R → R be a convex function whose domain includes the range of X. Then,

E[g(X)] ≥ g(E[X])

If the function f is strictly convex, then the inequality holds with equality if and only if X is degenerate.

As a consequence of Jensen’s inequality, let X = |Z| and g : x → x2 . Then, E[g(X)] = E(Z 2 ) and

g(E[X]) = E[|Z|]2 hence Jensen’s inequality reads

E[|Zt|]2 ≤ E[Z 2 ] = σ 2

Both these measures will be influenced by the percentage of the correct predictions (value 0). To

have a more accurate measure in terms of average notch distance when the prediction is wrong we

can calculate a conditional expectation:

X F(i)
E[Y | Y 6= 0] = i∗ P
i6=0 j6=0 F(j)

This conditional expectation eliminates the correct prediction and its expectation should be a better

24
measure of how many notches we expect the algorithm to be off when the prediction fails.

5 Experiment Results and Analysis

In this section we study the performance of four Machine learning techniques: Bagged Decision Tree

(BDT), Random Forest (RF), Multiple Layer Perceptron (MLP) and Support Vector Machine (SVM)

techniques to predict corporate credit ratings. Figures 2,3,4 illustrate some examples of the type of

results obtained. Figures depict the training set (blue circle), test set (red circle) and predictions

(green x) for various stocks for three methods: Bagged Decision Tree, Random Forest, and MLP.

Figure 2: Train set, Test set and Predictions of AIG using Bagged Decision Tree

25
Figure 3: Train set, Test set and Predictions of FITB using Random Forest

26
 Figure 4: Train set, Test set and Predictions of BK using MLP

5.1 Analysis of Prediction Accuracy for different Machine Learning models

To address overfitting and obtain a more accurate results, we use a, so called, 10-fold cross validation

procedure. Specifically, each sector data is split into three subsets, a training set, a validation set and

a test set. For example, for financial sector dataset, a training set of 70%, which is 4220 data points, a

validation set of 10% which is 603 data points and a test set of 20% which is 1206 data points of input

data have been considered. Table 7 presents prediction accuracy of each of these techniques based

on 10-fold cross validation procedure. The SVM technique is a binary type classification technique.

However, in our case we have several categories that need to be classified not just two. Thus we use

two different SVM-based methods to classify the rating categories (Kim and Ahn, 2012).

27
 Financial Sector Energy Sector Healthcare Sector

Bagged Decision Tree 84.21% 82.11% 83.90%

Random Forest 82.83% 84.45% 82.97%

Multi Layer Perceptron 73.95% 78.19% 76.63%

One Vs. One-SVM 42.12% 75.31% 71.29%

One Vs. All-SVM 40.14% 59.17% 61.89%

Table 7: Accuracy of different techniques

Table 7 contains the type of results that are reported by the majority of articles applying Machine

Learning algorithms to the credit rating problem. According to the results we obtain it is clear that

Bagged Decision Tree and Random Forest outperformed MLP and SVM-based approaches. This type

of performance was observed for all datasets and periods considered. Based on the Tables 1 and 2, most

of the related studies on S&P credit scores, prove that neural network and support vector machine

obtain more than 80% prediction accuracy on US firms. Only one study obtained 36% accuracy of

neural network on industrial firms Moody and Utans (1994).

For a financial analysis it is important to have a measure of how far the prediction is when the

algorithm fails. To this end recall we introduced the Notch Distance in Definition 1 on page 23.

A summary of the distribution of the notch distance Y is presented in Table 8. As expected values

for F (0) are exactly the prediction accuracy percentages in Table 7.

28
 Table 8: Computed Notches for all stocks in three sectors

Sector Methods Zero One Absolute Greater than One Absolute

Financial Bagged Decision Tree 84.21% 12.01% 3.78%

Random Forest 82.83% 12.08% 5.09%

Multi Layer Perceptron 73.95% 15.92% 10.13%

One Vs. One-SVM 42.12% 32.13% 25.75%

One Vs. All-SVM 40.14% 32.28% 27.58%

Energy Bagged Decision Tree 82.11% 13.98% 3.91%

Random Forest 84.45% 11.85% 3.70%

Multi Layer Perceptron 78.19% 15.48% 6.33%

One Vs. One-SVM 75.31% 16.98% 7.71%

One Vs. All-SVM 59.17% 20.64% 20.19%

Healthcare Bagged Decision Tree 83.90% 13.51% 2.59%

Random Forest 82.97% 11.74% 5.29%

Multi Layer Perceptron 76.63% 16.08% 7.29%

One Vs. One-SVM 71.29% 16.03% 12.68%

One Vs. All-SVM 61.89% 24.50% 13.61%

29
 Table 9: Expected Values and Standard Deviation of all notches in three sectors

Sector Methods E[X] σ(X) E[|X|]

Financial Bagged Decision Tree 0.0501 0.7614 0.2719

Random Forest 0.0068 0.7028 0.2731

Multi Layer Perceptron 0.1381 1.0152 0.5114

One Vs. One-SVM 0.2661 1.6917 1.4584

One Vs. All-SVM 0.6263 1.4562 1.4699

Energy Bagged Decision Tree 0.0980 0.5817 0.2851

Random Forest -0.0594 0.5892 0.3212

Multi Layer Perceptron -0.0800 0.8421 0.3858

One Vs. One-SVM -0.0685 0.7224 0.3481

One Vs. All-SVM -0.0066 0.9117 0.6491

Healthcare Bagged Decision Tree 0.1975 0.3512 0.2959

Random Forest -0.2154 0.6317 0.3223

Multi Layer Perceptron -0.0531 0.6854 0.4486

One Vs. One-SVM 0.1191 0.7515 0.6531

One Vs. All-SVM 0.2461 0.9683 0.8372

Table 9 provides expected value and standard deviation of the notch distance for every model. The

columns represent the three sectors we are analyzing: financial sector, energy sector and healthcare

sector. First we look at the expected value as an indication of symmetry of prediction. The standard

deviation and expected absolute value are measure of how far the prediction is from the actual rating.

The results in the table generally indicate that Bagged decision tree and Random Forest methods

have a better performance compared to the ANN and SVM. The largest proportion of predictions

is accurate thus the notches distance is zero. We next calculate the conditional expectation of the

notches measurement value when we are wrong. Table 10 shows this conditional expected values and

the conditional standard deviation.

30
 Table 10: Conditional Expected Values and Standard Deviation of all notches in three sectors

Sector Methods E[X|X 6= 0] σ(X|X 6= 0) E[|X||X 6= 0]

Financial Bagged Decision Tree -0.0018 0.8619 1.2795

Random Forest -0.0489 0.9037 1.1451

Multi Layer Perceptron 0.2391 1.3294 1.3912

One Vs. One-SVM 0.0911 1.8011 2.0594

One Vs. All-SVM 0.7216 1.4991 1.9171

Energy Bagged Decision Tree 0.0770 0.3735 0.8264

Random Forest -0.0885 0.3010 0.7587

Multi Layer Perceptron -0.1317 0.5615 0.9991

One Vs. One-SVM -0.2415 0.3512 0.9387

One Vs. All-SVM -0.5131 0.4411 1.2853

Healthcare Bagged Decision Tree 0.1973 0.2218 0.5313

Random Forest -0.4516 0.3054 0.6214

Multi Layer Perceptron -0.2712 0.3614 0.7778

One Vs. One-SVM -0.0710 0.5784 1.0496

One Vs. All-SVM 0.4052 0.6413 1.2384

Looking at the standard deviation and expected absolute for the unconditional Notches distance

in Table 9 it would appear that the performance for all sectors is roughly similar. However, looking at

the conditional numbers in Table 10 we see that healthcare and perhaps energy sectors are predicted

better than the financial sector. The numerical values in Table 10 also provide information about how

far the prediction is when it is wrong. For example the Bagged decision tree is on the average 0.81

notches away from the true value.

5.2 Comparing the performance of the various Machine Learning Credit

In the previous section we compared the accuracy of the prediction of Machine learning models with

the corporate ratings provided by Standard and Poor’s, Moody’s and Fitch. These three credit

agencies do not always agree on the ratings of companies. Therefore, the variation of their ratings as

31
 Table 11: Notch distances for the three rating agencies S&P, Moody’s, and Fitch quarter

E[X] σ(X) E[|X|] E[X|X 6= 0] σ(X|X 6= 0) E[|X||X 6= 0]

measured by Notch distance is useful in order to assess a measure of an acceptable standard deviation

of a specific credit rating.

In Table 11 we present the same statistics as in the previous table using quarterly ratings from

each of the rating agencies. We can see that the notch distance between the rating agencies shows less

agreement between themselves than most of the Machine learning methods considered in this paper.

This should give some confidence in the machine learning methods in this paper.

5.3 Capturing rating changes

Generally, credit ratings do not change very much from quarter to quarter. Thus for example if a

dataset contains 90% of quarters in which ratings did not change from the previous quarter, a vary

naive model that would predict the same rating as the one from the previous quarter would have

90% accuracy! Thus, the most important issue when predicting credit rating based on information

available is about predicting a rating change when the actual rating will change. Table 12 shows

the percentage of credit rating changes that have been captured by each model. These numbers are

actually much lower than the precision of the models. They are larger than we expected, however, the

performance of the Bagged decision tree and random forest is not that far from the MLP and SVM

in capturing rating changes. These ideas will be investigated in future work as predicting changes is

the most important part of the credit analysis.

Table 12: Captured Changes

Financial Sector Energy Sector Healthcare Sector

Bagged Decision Tree 67.64% 88.88% 77.27%

Random Forest 71.56% 77.77% 83.33%

Multi Layer Perceptron 53.92% 77.77% 66.66%

One Vs. One-SVM 29.41% 77.77% 58.33%

One Vs. All-SVM 25.49% 33.33% 41.66%

32
6 Conclusion and Future Directions

In this work we first provide a survey of current literature using machine learning techniques to predict

corporate credit ratings. The survey pointed toward support vector machine (SVM) and artificial

neural network (ANN) as methods that produce the most successful results. We implemented these

methods and two other methods based on decision trees which in addition have the ability to select

best features for the classification problems. We applied all methods for the same datasets spanning

2009 to 2018 for companies from two different sectors: energy, and healthcare and another dataset

spanning 1990 to 2018 for companies from financial sector. All our results indicate that decision tree

methods are in fact outperforming SVM and MLP. To check the results we introduced a Notches

Distance concept. This measure helps us quantify how far the predictions are from the true values

when the algorithm fails to produce the correct prediction. Using this new measure the results were

unchanged. Since this is a new measure we need to have a benchmark for the results we obtained. To

this end we calculated the Notch Distance using ratings provided by S&P, Moody’s and Fitch on the

same companies. We found that the best algorithms produce a Notch Distance from the true ratings

which is comparable with the distance between ratings produced by different rating agencies on the

same company.

References

Abdou, H. A. (2009). An evaluation of alternative scoring models in private banking. The Journal of

Risk Finance, 10(1):38–53.

Abdou, H. A., Tsafack, M. D. D., Ntim, C. G., and Baker, R. D. (2016). Predicting creditworthiness

in retail banking with limited scoring data. Knowledge-Based Systems, 103:89–103.

Addo, P. M., Guegan, D., and Hassani, B. (2018). Credit risk analysis using machine and deep learning

models. Risks, 6(2):38.

Akkoç, 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.

Angelini, E., di Tollo, G., and Roli, A. (2008). A neural network approach for credit risk evaluation.

The quarterly review of economics and finance, 48(4):733–755.

33
Bahrammirzaee, A. (2010). A comparative survey of artificial intelligence applications in finance:

artificial neural networks, expert system and hybrid intelligent systems. Neural Computing and

Applications, 19(8):1165–1195.

Bottou, L., Cortes, C., Denker, J. S., Drucker, H., Guyon, I., Jackel, L. D., LeCun, Y., Muller,

U. A., Sackinger, E., Simard, P., et al. (1994). Comparison of classifier methods: a case study in

handwritten digit recognition. In Pattern Recognition, 1994. Vol. 2-Conference B: Computer Vision

& Image Processing., Proceedings of the 12th IAPR International. Conference on, volume 2, pages

77–82. IEEE.

Brabazon, A. and O’Neill, M. (2006). Credit classification using grammatical evolution. Informatica,

30(3).

Breiman, L. (2001). Random forests. Machine learning, 45(1):5–32.

Breiman, L., Friedman, J., Olshen, R., and Stone, C. (1984). Classification and regression t rees

(monterey, california: Wadsworth).

Brennan, D. and Brabazon, A. (2004). Corporate bond rating using neural networks. In IC-AI.

Cao, L., Guan, L. K., and Jingqing, Z. (2006). Bond rating using support vector machine. Intelligent

Data Analysis, 10(3):285–296.

Chandra, D. K., Ravi, V., and Bose, I. (2009). Failure prediction of dotcom companies using hybrid

intelligent techniques. Expert Systems with Applications, 36(3):4830–4837.

Chaveesuk, R., Srivaree-Ratana, C., and Smith, A. E. (1999). Alternative neural network approaches

to corporate bond rating. Journal of Engineering Valuation and Cost Analysis, 2(2):117–131.

Chen, Y.-S. (2012). Classifying credit ratings for asian banks using integrating feature selection and

the cpda-based rough sets approach. Knowledge-Based Systems, 26:259–270.

Chen, Y.-S. and Cheng, C.-H. (2013). Hybrid models based on rough set classifiers for setting credit

rating decision rules in the global banking industry. Knowledge-Based Systems, 39:224–239.

Cortes, C. and Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3):273–297.

Culkin, R. and Das, S. R. (2017). Machine learning in finance: The case of deep learning for option

pricing. Journal of Investment Management, 15(4).

34
Daniel, C., Hančlová, J., el Woujoud Bousselmi, H., et al. (2019). Corporate rating forecasting

using artificial intelligence statistical techniques. Investment Management & Financial Innovations,

16(2):295.

Dima, A. M. and Vasilache, S. (2016). Companies default prediction using neural networks. Romanian

Journal of Economic Forecasting, 19(3):127.

Dutta, S. and Shekhar, S. (1988). Bond rating: a non-conservative application of neural networks.

In Proceedings of the IEEE international conference on neural networks, volume 2, pages 443–450.

Los Alamitos: IEEE Press.

Feng, G., He, J., and Polson, N. G. (2018). Deep learning for predicting asset returns. arXiv preprint

arXiv:1804.09314.

Gangolf, C., Dochow, R., Schmidt, G., and Tamisier, T. (2016). Automated credit rating prediction

in a competitive framework. RAIRO-Operations Research, 50(4-5):749–765.

Ganguli, S. and Dunnmon, J. (2017). Machine learning for better models for predicting bond prices.

arXiv preprint arXiv:1705.01142.

Garavaglia, S. (1991). An application of a counter-propagation neural network: simulating the stan-

dard and poor’s corporate bond rating system. In Artificial Intelligence Applications on Wall Street,

1991. Proceedings., First International Conference on, pages 278–287. IEEE.

Gogas, P., Papadimitriou, T., and Agrapetidou, A. (2014). Forecasting bank credit ratings. The

Journal of Risk Finance, 15(2):195–209.

Gu, S., Kelly, B. T., and Xiu, D. (2019). Empirical asset pricing via machine learning. 31st Australasian

Finance and Banking Conference 2018.

Hájek, P. (2012). Credit rating analysis using adaptive fuzzy rule-based systems: an industry-specific

approach. Central European Journal of Operations Research, 20(3):421–434.

Hajek, P. and Michalak, K. (2013). Feature selection in corporate credit rating prediction. Knowledge-

Based Systems, 51:72–84.

Hájek, P. and Olej, V. (2011). Credit rating modelling by kernel-based approaches with supervised

and semi-supervised learning. Neural Computing and Applications, 20(6):761–773.

35
Hájek, P. and Olej, V. (2014). Predicting firms’ credit ratings using ensembles of artificial immune

systems and machine learning–an over-sampling approach. In IFIP International Conference on

Artificial Intelligence Applications and Innovations, pages 29–38. Springer.

Hajek, P., Olej, V., and Prochazka, O. (2016). Predicting corporate credit ratings using content

analysis of annual reports–a naı̈ve bayesian network approach. In FinanceCom 2016, pages 47–61.

Springer.

Horrigan, J. O. (1966). The determination of long-term credit standing with financial ratios. Journal

of Accounting Research, pages 44–62.

Huang, W., Nakamori, Y., and Wang, S.-Y. (2005). Forecasting stock market movement direction

with support vector machine. Computers & Operations Research, 32(10):2513–2522.

Huang, Z., Chen, H., Hsu, C.-J., Chen, W.-H., and Wu, S. (2004). Credit rating analysis with support

vector machines and neural networks: a market comparative study. Decision support systems,

37(4):543–558.

Hwang, R.-C., Cheng, K., and Lee, C.-F. (2009). On multiple-class prediction of issuer credit ratings.

Applied Stochastic Models in Business and Industry, 25(5):535–550.

Hwang, R.-C., Chung, H., and Chu, C. (2010). Predicting issuer credit ratings using a semiparametric

method. Journal of Empirical Finance, 17(1):120–137.

Irmatova, E. (2016). Relarm: A rating model based on relative pca attributes and k-means clustering.

arXiv preprint arXiv:1608.06416.

Kamstra, M., Kennedy, P., and Suan, T.-K. (2001). Combining bond rating forecasts using logit.

Financial Review, 36(2):75–96.

Kanchymalay, K., Salim, N., Sukprasert, A., Krishnan, R., and Hashim, U. R. (2017). Multivariate

time series forecasting of crude palm oil price using machine learning techniques. In IOP Conference

Series: Materials Science and Engineering, volume 226, page 012117. IOP Publishing.

Kao, L.-J., Chiu, C.-C., and Chiu, F.-Y. (2012). A bayesian latent variable model with classification

and regression tree approach for behavior and credit scoring. Knowledge-Based Systems, 36:245–252.

Kaplan, R. S. and Urwitz, G. (1979). Statistical models of bond ratings: A methodological inquiry.

Journal of business, pages 231–261.

36
Karminsky, A. M. and Khromova, E. (2016). Extended modeling of banks’ credit ratings. Procedia

Computer Science, 91:201–210.

Khandani, A. E., Kim, A. J., and Lo, A. W. (2010). Consumer credit-risk models via machine-learning

algorithms. Journal of Banking & Finance, 34(11):2767–2787.

Khashman, A. (2010). Neural networks for credit risk evaluation: Investigation of different neural

models and learning schemes. Expert Systems with Applications, 37(9):6233–6239.

Khemakhem, S. and Boujelbene, Y. (2015). Credit risk prediction: A comparative study between

discriminant analysis and the neural network approach. Accounting and Management Information

Systems, 14(1):60.

Kim, H. S. and Sohn, S. Y. (2010). Support vector machines for default prediction of smes based on

technology credit. European Journal of Operational Research, 201(3):838–846.

Kim, J. W., Weistroffer, H. R., and Redmond, R. T. (1993). Expert systems for bond rating: a compar-

ative analysis of statistical, rule-based and neural network systems. Expert systems, 10(3):167–172.

Kim, K.-j. (2003). Financial time series forecasting using support vector machines. Neurocomputing,

55(1-2):307–319.

Kim, K.-j. and Ahn, H. (2012). A corporate credit rating model using multi-class support vec-

tor machines with an ordinal pairwise partitioning approach. Computers & Operations Research,

39(8):1800–1811.

Kim, K. S. (2005). Predicting bond ratings using publicly available information. Expert Systems with

Applications, 29(1):75–81.

Kim, K.-S. and Han, I. (2001). The cluster-indexing method for case-based reasoning using self-

organizing maps and learning vector quantization for bond rating cases. Expert systems with appli-

cations, 21(3):147–156.

Koh, H. C., Tan, W. C., and Peng, G. C. (2004). Credit scoring using data mining techniques.

Singapore Management Review, 26(2):25.

Kwon, Y. S., Han, I., and Lee, K. C. (1997). Ordinal pairwise partitioning (opp) approach to neural

networks training in bond rating. International journal of intelligent systems in accounting finance

and management, 6:23–40.

37
Lee, T.-S., Chiu, C.-C., Chou, Y.-C., and Lu, C.-J. (2006). Mining the customer credit using classifi-

cation and regression tree and multivariate adaptive regression splines. Computational Statistics &

Data Analysis, 50(4):1113–1130.

Lee, Y.-C. (2007). Application of support vector machines to corporate credit rating prediction. Expert

Systems with Applications, 33(1):67–74.

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(1):124–136.

Liu, Y. and Zheng, Y. F. (2005). One-against-all multi-class svm classification using reliability mea-

sures. Neural Networks, 2:849–854.

Maher, J. J. and Sen, T. K. (1997). Predicting bond ratings using neural networks: a comparison

with logistic regression. Intelligent Systems in Accounting, Finance & Management, 6(1):59–72.

Martens, D., Van Gestel, T., De Backer, M., Haesen, R., Vanthienen, J., and Baesens, B. (2010).

Credit rating prediction using ant colony optimization. Journal of the Operational Research Society,

61(4):561–573.

Moody, J. and Utans, J. (1994). Architecture selection strategies for neural networks: Application

to corporate bond rating prediction. In Neural networks in the capital markets, pages 277–300.

Citeseer.

Moscatelli, M., Narizzano, S., Parlapiano, F., Viggiano, G., et al. (2019). Corporate default forecasting

with machine learning. Technical report, Bank of Italy, Economic Research and International

Relations Area.

Oelerich, A. and Poddig, T. (2006). Evaluation of rating systems. Expert Systems with Applications,

30(3):437–447.

Pacelli, V. and Azzollini, M. (2011). An artificial neural network approach for credit risk management.

Journal of Intelligent Learning Systems and Applications, 3(02):103.

Pai, P.-F., Tan, Y.-S., and Hsu, M.-F. (2015). Credit rating analysis by the decision-tree support

vector machine with ensemble strategies. International Journal of Fuzzy Systems, 17(4):521–530.

38
Paleologo, G., Elisseeff, A., and Antonini, G. (2010). Subagging for credit scoring models. European

journal of operational research, 201(2):490–499.

Peng, Y., Wang, G., Kou, G., and Shi, Y. (2011). An empirical study of classification algorithm

evaluation for financial risk prediction. Applied Soft Computing, 11(2):2906–2915.

Petropoulos, A., Chatzis, S. P., and Xanthopoulos, S. (2016). A novel corporate credit rating system

based on student’st hidden markov models. Expert Systems with Applications, 53:87–105.

Platt, J. C., Cristianini, N., and Shawe-Taylor, J. (2000). Large margin dags for multiclass classifica-

tion. In Advances in neural information processing systems, pages 547–553.

Ryser, M. and Denzler, S. (2009). Selecting credit rating models: a cross-validation-based comparison

of discriminatory power. Financial Markets and Portfolio Management, 23(2):187–203.

Saha, S. and Waheed, S. (2017). Credit risk of bank customers can be predicted from customer’s

attribute using neural network. International Journal of Computer Applications, 975:8887.

Sambasivan, R. and Das, S. (2017). A statistical machine learning approach to yield curve forecasting.

In Computational Intelligence in Data Science (ICCIDS), 2017 International Conference on, pages

1–6. IEEE.

Schölkopf, B., Burges, C. J., and Smola, A. J. (1999). Advances in kernel methods: support vector

learning. MIT press.

Shin, K.-s. and Han, I. (2001). A case-based approach using inductive indexing for corporate bond

rating. Decision Support Systems, 32(1):41–52.

Surkan, A. J. and Singleton, J. C. (1990). Neural networks for bond rating improved by multiple

hidden layers. In Neural Networks, 1990., 1990 IJCNN International Joint Conference on, pages

157–162. IEEE.

Tsai, C.-F. and Wu, J.-W. (2008). Using neural network ensembles for bankruptcy prediction and

credit scoring. Expert systems with applications, 34(4):2639–2649.

Vapnik, V. (2013). The nature of statistical learning theory. Springer science & business media.

Veronezi, P. H. (2016). Corporate credit rating prediction using machine learning techniques. Master’s

thesis, Stevens Institute of Technology.

39
Wallis, M., Kumar, K., and Gepp, A. (2019). Credit rating forecasting using machine learning tech-

niques. In Managerial Perspectives on Intelligent Big Data Analytics, pages 180–198. IGI Global.

Wang, A., Yuan, W., Liu, J., Yu, Z., and Li, H. (2009). A novel pattern recognition algorithm:

Combining art network with svm to reconstruct a multi-class classifier. Computers & Mathematics

with Applications, 57(11-12):1908–1914.

West, D. (2000). Neural network credit scoring models. Computers & Operations Research, 27(11-

12):1131–1152.

West, R. R. (1970). An alternative approach to predicting corporate bond ratings. Journal of Ac-

counting Research, pages 118–125.

Wu, T.-C. and Hsu, M.-F. (2012). Credit risk assessment and decision making by a fusion approach.

Knowledge-Based Systems, 35:102–110.

Xiao, H., Xiao, Z., and Wang, Y. (2016). Ensemble classification based on supervised clustering for

credit scoring. Applied Soft Computing, 43:73–86.

Ye, Y., Liu, S., and Li, J. (2008). A multiclass machine learning approach to credit rating prediction.

In Information Processing (ISIP), 2008 International Symposiums on, pages 57–61. IEEE.

Yeh, C.-C., Lin, F., and Hsu, C.-Y. (2012). A hybrid kmv model, random forests and rough set theory

approach for credit rating. Knowledge-Based Systems, 33:166–172.

Yu, L., Wang, S., and Lai, K. K. (2008). Credit risk assessment with a multistage neural network

ensemble learning approach. Expert systems with applications, 34(2):1434–1444.

Zhao, Z., Xu, S., Kang, B. H., Kabir, M. M. J., Liu, Y., and Wasinger, R. (2015). Investigation

and improvement of multi-layer perceptron neural networks for credit scoring. Expert Systems with

Applications, 42(7):3508–3516.

Zhong, H., Miao, C., Shen, Z., and Feng, Y. (2014). Comparing the learning effectiveness of bp, elm,

i-elm, and svm for corporate credit ratings. Neurocomputing, 128:285–295.

40