The International Journal of Social Sciences and Humanities Invention 5(06): 4820-4823, 2018

DOI: 10.18535/ijsshi/v5i6.09 ICV 2015: 45.28

ISSN: 2349-2031
© 2018, THEIJSSHI

Research Article
Research on bank credit default prediction based on data mining algorithm
Li Ying
School of Business Administration, China University of Petroleum-Beijing, Beijing, 102249, China

ABSTRACT: It is of great importance to identify the potential risks to the bank's loan customers. Based on data mining
technology, it is an effective method to classify loan customers by classification algorithm. In this paper, we use Random
Forest method, Logistic Regression method, SVM method and other suitable classification algorithms by python to study
and analyze the bank credit data set, and compared these models on five model effect evaluation statistics of Accuracy,
Recall, precision, F1-score and ROC area. This paper use the data mining classification algorithm to identify the risk
customers from a large number of customers to provide an effective basis for the bank's loan approval.

Key words: Bank credit, Risk prediction, Data mining, Classification algorithm, python
INTRODUCTION Data sample collection and preprocessing
In today's information and digit age, bank credit default is still This paper uses the bank credit data set loan_model_sample in
frequent, how to establish an effective model for the the kaggle as the target data set for the study[6]. There are
prediction whether bank customers will default on the loan for 11017 samples and 199 attribute features. After the data
recognition of the risk in bank from a mass of loan applicants collection is completed, the data is viewed and pre-processed.
is of great significance. At present, many scholars at home and The overall framework of data preprocessing is shown in
abroad have studied the probability prediction model of bank Figure 1.
credit default and put forward some forecasting methods
which almost have different restrictions and defects. Beaver Loan Sample Data
proposes to use a single-factor method of financial ratios to
analyze the technical credit default prediction of Split train and test
enterprises[1]. Pomp at al constructed a default prediction
model using multivariate discriminant analysis (MDA)[2]. Missing value processing
Yang at al used Logistic Regression method establishing a Yes
probability prediction model of listed companies’ credit No
Missing value scale<20%?
default, and identified the most influential corporate financial
indicators[3]. Zhang at al proposed a SVM model, which
constructed the technical credit default prediction model of Delete the Mean value
small and medium-sized enterprises by constructing the feasures interpolation

evaluation index system of different input variables[4]. A

large number of studies have shown that the data mining Class imbalance
Test StandardScaler
processing
classification algorithm can find hidden rules from mass data
and use this rule to classify and predict new unknown data[5]. Feature Engineering
Through data mining classification algorithm, banks can set up
classification models by using the relevant personal Dimension specification
information and consumption data of the loan applicants in the
past, and find out the characteristics of risk customers. Then, Figure 1 Data pre-processing framework
use the classification model making classification prediction
Firstly，In order to prevent the occurrence of data leakage
for new loan applicants, from which identify the risk
problem, the data set is divided into two parts: training and
customers, so as to reduce the risk of default repayment.
testing. Training is used to train the model, and testing is used
Therefore, in this paper we use the Random Forest method,
to test the model classification accuracy. Secondly, view the
Logistic Regression method, SVM method and other suitable
missing values of the data set and process the missing values.
classification algorithms to study and analyze the bank credit
The visualization of the missing values of the original data is
data set, and compared these models on five model effect
shown in Figure 2, in which white lines represent missing
evaluation statistics of Accuracy, Recall, precision, F1-score
data. The threshold is set to 20%, and the features whose
and ROC area to identify the risk customers from a large
missing values are greater than the threshold will be directly
number of customers and provide effective approaches for the
deleted, the feature attributes whose missing values are
bank's loan approval.
4820 The International Journal of Social Sciences and Humanities Invention, vol. 5, Issue 06, Jun, 2018
 Li Ying / Research on bank credit default prediction based on data mining algorithm
smaller than the threshold will be averaged using the A series of measures for measuring the performance of
preprocessing module of sklearn in scikit-learn. The processed learning systems such as Overall Accuracy, Recall, precision
data visualization is shown in Figure 3, it can be seen that the and F1-score, can be defined based on the confusion matrix.
entire data set has been filled completely. The definition is as follows.
Accuracy is the correct proportion of the overall number of
predictions, it is defined by the formula[11]:

Where TP is the number of true positives, TN is the number of

true negatives, FP is the number of false positives, and the
number of false negatives[12,13].
Figure 2 Missing data visualization
Precision is to measure the proportion of truly positive
samples in samples that are predicted as positive samples by
the classification model[14]:

Recall is a measure of the ability of prediction model to select

instances of a certain class from a data set[7], which is also
represent TPR(true positive rate)[15]:
Figure 3 Missing data visualization after processing
Then, use the StandardScaler in scikit-learn to standardize test
set data and observe whether the class exists the imbalanced F1-score is the harmonic average of precision and recall, it is
issue. Perform feature engineering on datasets and select defined by the formula[11]:
feature attributes that are valid for the classification model.
Finally, in order to avoid overfitting problems and reduce the
complexity of the model, using the PCA method, Pearson
correlation coefficient or other automatic screening feature FPR(false positive rate) is to describe the proportion of the
methods to perform dimension reduction on the training set. model negative class in the samples predicted to be
Establishment and Evaluation of classification models positive[14]:

Based on the pre-processed data sets, use Random Forest[7],

Logistic Regression[8] and SVM[9] method through python ROC(receiver operating characteristic curve) is a technique for
programming language to establish classification models visualizing, organizing and selecting classifiers based on their
respectively and adjust the hyper parameters with performance[15]. It is a comprehensive index reflecting the
GridSearchCV method. Thus, the parameters are obtained continuous variables of sensitivity and specificity. The ROC
when the model classification effect is best. Using five model curve is a two-dimensional curve with FPR as the X axis and
effect evaluation statistics: Overall Accuracy, Recall, TPR as the Y axis, which ranges from (0,0) to (1,1). A
precision, F1-score and ROC area to compared the common method to compare classifiers is to calculate the area
classification prediction effect of these classification models. under the ROC curve, abbreviated AUC[16,17]. The larger the
The most straightforward way to evaluate the classification AUC, the better the model classification is.
model performance is based on the confusion matrix analysis. Since the purpose of the bank credit default model is to
Confusion matrix is a concept from machine learning that identify credit default risk customers from a large number of
contains information about actual classifications and predicted loan application customers, the risk of predicting a risk
classifications done by a classification system. A confusion customer (Class = 0) as a normal customer (Class = 1) is much
matrix has two-dimensions, one is indexed by the actual class larger than predicting a normal customer (Class = 1) to
of an object, the other is indexed by the class that the classifier customers (Class=0). Based on the above analysis, this paper
predicts[10]. In the bank's credit data in this paper, the number pays more attention to whether the model can correctly
0 represents the credit default customer category, and 1 classify Class=0 when assessing model classification results,
represents the normal customer category. The confusion so the classification algorithms results in this paper are all
matrix is shown in Table Ⅲ. from Class=0.
Table Ⅲ CONFUSION MATRIX Findings and discussions
Predicted Class
Class=1 Class=0 This paper uses RandomForest method, LogisticRegression
method and SVM method to establish classification models.
Actual Class Class=1 TP FN
Perform research and analysis on pre-processed bank credit
Class=0 FP TN
default data sets, and use the GridSearchCV method to search
4821 The International Journal of Social Sciences and Humanities Invention, vol. 5, Issue 06, Jun, 2018
 Li Ying / Research on bank credit default prediction based on data mining algorithm
for the best parameters. The above three classification model penalty parameter C of the error term and gamma which
construction methods are respectively correspond to represents kernel coefficient for ‘rbf’, ‘poly’ and ‘sigmoid’. C
RandomForestClassifier, LogisticRegression, and SVC is the penalty factor. The larger its value is, the heavier the
algorithms in scikit-learn. penalty for misclassification of training samples and the
In the RandomForestClassifier algorithm，what need to higher the requirement for correct classification are. Gamma is
perform GridSearchCV to find the best parameter value in a parameter of the insensitive loss function. The smaller the
order to enhance the prediction effect of classification model value is, the more support vectors are; and the larger the
is n_estimators，which represents the number of trees in the distance between two large edges is, the easier it is to find the
forest. In addition，with the number of trees in the forest maximum hyper-plane Maximum Marginal Hyperplane[5].
After parameter optimization, we obtain the comparison
increasing, the classifier is becoming more and more
results of RandomForestClassifier, LogisticRegression, and
complicated, which probably brings overfitting problem to the
SVC algorithms on the loan_model_sample, as shown in
model. To decrease the model complexity, we also adjust
values of some other parameters in the Table Ⅳ and algorithm ROC curves are shown in Figure 4.
RandomForestClassifier algorithm, such as max_depth The above comparison results show that
min_samples_split、min_samples_leaf and max_features. RandomForestClassifier algorithm has a good classification
effect for large sample as well as high dimensional attribute
In the LogicsticRegression algorithm, we use GridSearchCV
to search the best Parameter combination, one is the feature datasets，because the value of its Recall, F1-score,
penalty(Penalty item) between l1 and l2, the other is C which Accuracy, Precision and ROC area are all larger than the other
represents the reciprocal of the regularization coefficient λ. Classifiers. And the LogisticRegression algorithm has
The objective function can be summarized as follows[18]: relatively high recall, but with the lowest value of precision.
The SVC algorithm has the relatively high accuracy, precision
∑ and F1-score, but with a lowest recall, which to a certain
Where the first item L is the training error, and the second degree shows SVC algorithm may be not suitable for the bank
item is the penalty item. The first item is to minimize the credit default prediction models, because it may well omit risk
training error and get the best fitting data; the second is to loan applicants and cause huge losses to banks. Thus, we
simplify the model, prevent overfitting, and get better consider RandomForest algorithm as the most suitable
generalization ability. algorithm for the bank credit default prediction problem,
SVC algorithm is the classifier model of SVM. In the SVC especially when the dataset is very large or has high
algorithm, we use GridSearchCV to find the best Parameter dimensions.
combination of kernel {‘linear’,‘poly’, ‘rbf’}、 Conclusions
Table Ⅳ COMPARISON OF CLASSIFICATION This paper establish bank credit default prediction models
ALGORITHMS RESUITS using RandomForest, LogisticRegression and SVM
classification algorithms under the python language
precisio recal F1-
Algorithm Accuracy environment. And compare the classification effect of the
n l score
classifiers through five model effect evaluation statistics:
RandomForestCla
0.8863 0.18 0.67 0.28 Accuracy, Recall, precision, F1-score and ROC area.
ssifier
Comparative analysis Experimental results show that
LogisticRegressio
0.7130 0.09 0.64 0.16 compared to LogisticRegression and SVM classification
n
algorithms, RandomForest algorithm is more suitable for the
SVC 0.8580 0.15 0.55 0.24 bank credit default precision model because its high
classification effect for Class=0, especially when the dataset is
AUC=0.71 AUC=0.68 AUC=0.71 very large or has high dimensions. This paper provides an
effective experimental basis for bank credit approval to
identify risk customers from a large number of loan applicants
using data mining classification algorithms.
Since the number of appropriate public data sets for bank
credit is small, the number of samples in this paper is only
11,017, therefore, the experiment may not be comprehensive.
In the future, it is necessary to collect more datasets of large
number and features for further improvement.
References
RandomForestClassifier LogisticRegression SVC [1] Beaver, W. Financial Ratios as Predictors of
Figure 4 Comparison of algorithm ROC curves Failure.Empirical Research in Accounting: Selected
Studied[J]. Journal of Accounting Research, 1966，(4).

4822 The International Journal of Social Sciences and Humanities Invention, vol. 5, Issue 06, Jun, 2018
 Li Ying / Research on bank credit default prediction based on data mining algorithm
[2] Pompe, P.P.M., Bilderbe, J. The Prediction of Bankruptcy [18] Hilbe J M. Logistic regression models[M]. CRC press,
of Smalland Medium-sized Industrial Firms[J]. Journal of 2009.
Business Venturing, 2005,20.
[3] Yang Pengbo, Zhang Chenghu, Zhang Xiang. Prediction
model of credit default probability of listed companies
based on Logistic regression analysis [J]. economic
latitude and longitude,2009(02):144-148.
[4] Zhang Jie, Zhao Feng. [J]. statistics and decision making
of SME credit default prediction based on support vector
machine,2013(20):66-69.
[5] Mei Mei. Application of data mining classification
algorithm in credit card risk management [J]. modern
computer,2013(19):13-16.
[6] https://www.kaggle.com/datasets
[7] Liaw A, Wiener M. Classification and regression by
randomForest[J]. R news, 2002, 2(3): 18-22.
[8] Hosmer Jr D W, Lemeshow S, Sturdivant R X. Applied
logistic regression[M]. John Wiley & Sons, 2013.
[9] Joachims T. Making large-scale SVM learning
practical[R]. Technical report, SFB 475:
Komplexitätsreduktion in Multivariaten Datenstrukturen,
Universität Dortmund, 1998.
[10] Deng X, Liu Q, Deng Y, et al. An improved method to
construct basic probability assignment based on the
confusion matrix for classification problem[J].
Information Sciences, 2016, 340: 250-261.
[11] Ohsaki M, Wang P, Matsuda K, et al. Confusion-matrix-
based Kernel logistic regression for imbalanced data
classification[J]. IEEE Transactions on Knowledge and
Data Engineering, 2017, 29(9): 1806-1819.
[12] Branco P, Torgo L, Ribeiro R P. A survey of predictive
modeling on imbalanced domains[J]. ACM Computing
Surveys (CSUR), 2016, 49(2): 31.
[13] Fanshawe T R, Power M, Graziadio S, et al. Interactive
visualisation for interpreting diagnostic test accuracy
study results[J]. BMJ Evidence-Based Medicine, 2018,
23(1): 13-16.
[14] Davis J, Goadrich M. The relationship between Precision-
Recall and ROC curves[C]//Proceedings of the 23rd
international conference on Machine learning. ACM,
2006: 233-240.
[15] Fawcett T. An introduction to ROC analysis[J]. Pattern
recognition letters, 2006, 27(8): 861-874.
[16] Bradley A P. The use of the area under the ROC curve in
the evaluation of machine learning algorithms[J]. Pattern
recognition, 1997, 30(7): 1145-1159.
[17] Breheny P. Classification and regression trees[J]. 1984..

4823 The International Journal of Social Sciences and Humanities Invention, vol. 5, Issue 06, Jun, 2018