A machine learning-based credit risk prediction engine system using a stacked classifier and a filter-based feature selection method

One of the earliest applications of machine learning was for the prediction of credit risk, which uses financial data to predict the risk of customers defaulting a loan, credit card, and other lending services [1]. Credit risk prediction is a challenge for financial institutions, and several research works have attempted to address this problem [2]. The proper utilization of credit risk prediction tools can lead to increased profitability for financial institutions. Credit card and loan applications are two areas where this can be applied. Creditors who have been unable to adequately predict the credit risk of potential clients have had severe losses. Hence, proper risk assessment is crucial for the survival of these financial institutions [3].

Credit risk prediction has been a trending topic for the past few decades; credit card default prediction is among the most crucial tasks facing creditors. This is because the numbers of default transactions considerably outnumber the non-default transactions [4]. Therefore, the datasets used for credit risk prediction can be considered to have a class imbalance problem. Prior studies have shown that class imbalance can lead to poor classification performance of machine learning (ML) models that results in model bias towards a specific class at inference time [5]. In literature, several techniques have been proposed to solve the class imbalance problem, and they can be classified into three groups: ensemble learning, cost-sensitive learning, and re-sampling methods. Among these three methods, ensemble learning has been widely studied [6]. Ensemble learners perform better than a single model since they combine the advantages of several base learners. Furthermore, ensemble models can be divided into two groups: classifier ensemble and hybrid classifier. The former implies an ensemble model that combines an attribute selection technique or hyperparameter tuning prior to the classification whereas the latter combines numerous classifiers that run side by side [7].

Moreover, the datasets that are used to build credit risk prediction systems may possess a large feature space [16]. This can lead to an increased complexity while training machine learning models [37]. It is therefore vital to implement a feature selection (FS) algorithm that can alleviate the growing issue of feature space. FS algorithms are categorized as follows: filter, wrapper and hybrid. The filter-based FS make its decision based on the intrinsic nature of the dataset and therefore, is independent from the estimator that is used. The wrapper-based FS selects an optimal subset of features based on the performance obtained using estimator. Finally, the hybrid-based FS algorithms combines the filer and wrapper-based methodologies [8, 9].

In this research we implement a filter-based FS method that uses Information Gain (IG) [28]. IG is inspired from Information Theory [29]. The filter-based FS technique is selected because it is computationally less expensive in comparison to the wrapper and hybrid-based approaches [10].

Furthermore, we develop a multilevel ensemble-based model using the stacking method. Stacking or Stacked generalization is a technique that stacks the output of individual algorithms and uses a single classifier for the final prediction. This method uses the effectiveness of each individual classifier within a stack and utilizes their results as the input the final estimator [17]. The structure of the stack includes the following algorithms: Gradient Boosting [18], Random Forest [21] and Extreme Gradient Boosting [19].

The major contributions of this research are as follows:

The remaining part of this paper is structured as follows. "Related work" section presents a review of related works. In "Machine learning methods" section, we provide a background of the various machine learning algorithms used in this research. "Datasets" section provides an overview of the datasets. "Research Methodology" section presents the methodology that was followed in this research. "Feature Selection" section provides the details about the experimental settings. "Proposed Credit Risk Prediction Framework" section discusses the results and "Experimental Setup and Performance metrics" section concludes this paper.

Pande et al. [11] conducted a credit risk analysis using machine learning classifiers. In this analysis, the authors considered several methods including Artificial Neural Network (ANN), k-Nearest Neighbour and Naive Bayes (NB). To evaluate the performance of the ML models, the authors used the German credit risk dataset and the accuracy was considered as the main performance metric. The results demonstrated that the ANN, NB and KNN obtained accuracies of 77.45%, 77.20%, and 72.20%, respectively. Although these results represent a step in the right direction; the authors did not evaluate their models using additional metrics such as the F1-Score and Area Under the Curve (AUC) score.

Zhang et al. in [12] presented a credit scoring algorithm using adaptive sup- port vector machine (AdaSVM). This method was assessed on the Australian credit risk dataset and evaluated using the accuracy. The results demonstrated that the AdaSVM obtained an accuracy of 80%. This paper did not expand further in terms of evaluating the quality of classification by using additional metrics such the precision and recall.

Nasser and Maryam [13] developed a customer credit risk assessment system using Artificial Neural Networks (ANNs). In this research, the authors considered learning method such as the Gradient Descent. Moreover, the accuracy was the main performance metric that was utilized to assess the effectiveness of the proposed method. Furthermore, the authors used the Australian, Japanese and German credit risk datasets. The outcome of the experiments demonstrated that the ANN-GD obtained accuracies of 78.11%, 76.87%, and 68.26% for each dataset, respectively.

Hsu et al. [14] implemented an enhanced recurrent neural network (RNN) for combining static and dynamic attributes for credit card default prediction. This method was developed using an enhanced RNN and was evaluated using the Taiwan credit risk dataset. To enhance the RNN, the authors Gated Recurrent Units (GRUs) as the base nodes. The outcome of the numerical experiments showed that the RNN model achieved an AUC of 0.782 and a lift index of 0.659.

In [15], the authors presented a combination strategy of integrating super- vised learning coupled with unsupervised learning for credit risk assessment. In this work, the researchers used datasets such as the German dataset to assess the effectiveness of their proposed algorithms. Additionally, metrics such as the accuracy and the AUC were used to assess the performance of methods. In the instance of cluster-based approach, the KNN achieved an accuracy of 76.80 % and an AUC of 0.788. The RF achieved an accuracy of 72.10 % and an AUC of 0.811. The ANN obtained an accuracy of 78.6% and an AUC of 0.843. Finally, the cluster-based consensus (combined model) obtained an accuracy 80.8%.

Ha et al. [16] implemented an improved credit risk prediction model for online peer-to-peer (P2P) lending systems using a feature selection (FS) method and deep learning (DL). In this study, the first step consisted of preprocessing the data. The second step involved feature selection using Restricted Boltzmann Machines (RBMs). In the third step, the authors implemented the modeling process using machine learning (ML) methods such as Linear Discriminant Analysis (LDA), Artificial Neural Networks (ANN), k-Nearest Neighbors (KNN), and Random Forest (RF). These models were evaluated on various datasets, including the Australian and German credit risk datasets. Accuracy was the primary performance metric considered in the experiments. For the German dataset, the results showed the following accuracies: 76.50%, 75.8%, 67.10%, and 67.72% for LDA, ANN, KNN, and RF, respectively. For the Australian dataset, LDA, ANN, KNN, and RF achieved the following accuracies: 85.80%, 71.45%, 65.94%, and 67.72%, respectively. Although these results demonstrated some improvements compared to existing methods, the authors did not consider additional metrics such as precision, recall, and AUC.

This section provides an overview of the machine learning methods that were considered in this paper.

The RF algorithm computes its predictions by using a group of n Decision Trees (DTs) [20]. DT is a supervised ML technique that is used for classification and regression problems. A DT has the following categories of nodes: leaf node, decision node, and root node. The decision node represents a splitting point in a DT. A leaf node computes the final decision of the DT. The root node represents the initial state in the DT approach. Majority vote is a process that the RF algorithm uses to compute the predictions [21] as follows: let RF = {f (X, d_i)}, where i is the number of DTs and X represents an input vector and d_i is a set of DTs. The majority vote process is computed by d_i. The class with the most votes represents the prediction.

K-Nearest Neighbor (KNN) technique is a supervised ML method used for classification and regression tasks. The KNN approach uses the standard Euclidean (ED) method to compute the distance between data points as follows [22]: let n and m data points in space Q, the distance between n and m, D(n, m), is computed using the expression in (3).where t is total number of data points in space Q. The KNN approach estimates a prediction n₀ in Q by computing the ED between n₀ and its k closest data points within Q. As a result, n₀ is assumed to be like its neighbors [23].

Artificial Neural Network (ANN) is another type of ML algorithm that is used for classification and regression tasks. In this research, we used feed for- ward ANNs. ANNs are built using Artificial Neurons (ANs). An AN processes information from its input and forwards it to its output. Moreover, an AN is designed to solve both linear and non-linear problems. This is achieved by using different types of activation functions such as the Sigmoid, \(\sigma = \frac{1}{1+{e}^{-2}}\); the Rectified Linear Unit (ReLU): f (x) = max(0, x); or an hyperbolic tangent in (2).

Gradient boosting (GB) is a technique used to build regression and classification models to improve the learning process of the final model. In the GB algorithm, a meta-learner is built by using a group of weak estimators such as DTs. Each estimator is gradually added to the base group in a sequential manner. The aim of this process is to optimize the performance of the ensemble model by rectifying the mistakes made by the previous meta-learner [18]. This can be mathematically expressed as follows:where g represents the ensemble, t is the total number of estimators, h_n represents a single learner, and θ_n is a tunable parameter.

In this research, we selected feed forward ANNs because of their simplicity and training efficiency. ANNs are generally simpler in their structure compared to GANs. This simplicity is evident in their operational mechanics, as FFNNs involve a straightforward processing of inputs through hidden layers to outputs, using weights and biases, followed by an activation function. This linear processing makes FFNNs inherently less complex and more efficient in training than GANs, which require training two networks simultaneously (generator and discriminator). This complexity in GANs can lead to longer training times and increased computational cost [38].

Furthermore, we have selected ANNs because of the low computational cost and high scalability. From a computational standpoint, ANNs are generally more cost-effective. They require less computational power due to their simpler architecture, which also makes them more scalable for handling large datasets typical in credit risk analysis. In contrast, the dual-network structure of GANs demands more computational resources, leading to higher costs, especially when scaling up for extensive datasets. [39]. Additionally, we used ANNs because of model stability and predictive accuracy as explained in [40].

Finally, it must be noted that Generative Adversarial Networks (GANs) [36] or a Transformers based architecture could be considered in lieu of ANNs. However, GANs or Transformers are computationally expensive to train and require long training times. Moreover, GANs are better suited for tasks that involve data generation or more complex scenarios where adversarial training is beneficial.

All the datasets used in this work were obtained from the University of California, Irvine (UCI) machine learning repository. The Australian credit approval dataset [25] contains 690 instances and 14 attributes; in this dataset, there are 307 creditworthy clients and 383 defaulting clients. The German credit dataset [26] comprises 1000 cases and 20 features, with 700 creditworthy clients and 300 defaulting clients. Meanwhile, the Taiwan default of credit clients dataset [27] contains 30000 instances and 24 attributes, with 23364 creditworthy clients and 6636 defaulting clients. The German and Taiwan datasets are highly imbalanced, whereas the Australian credit dataset is relatively balanced. A summary of the number of features and instances in these datasets is provided in Table 1. The details about the nature of features in each dataset are provided in Tables 2, 3, 4. Moreover, these datasets are mostly made up of financial records and personal information, which were encoded for confidentiality reasons.

The proposed credit risk prediction framework is depicted in Fig. 1. This architecture includes two main phases, namely, the data processing phase (phase 1) and the modelling phase (phase 2). In the first phase, the full credit card fraud dataset is normalized and processed using the IG-based FS method. Moreover, the full dataset is split into a training data subset and testing data subset. In the modelling phase, the following individual classifiers are considered: RF, KNN, ANN, GB, and XGB. The proposed stacked classifier is built using the GB, XGB and RF estimators. Furthermore, once phase 1 is completed; each of the estimators in phase 2 are trained and tested using the training and testing sets generated from phase 1. The evaluation process is conducted using the accuracy, the f1-score and the Area Under the ROC Curve (AUC) as explained in "Feature Selection" section. The Compare Results block compares the metrics generated by each classifier and forwards the results to the Select Best Classifier for model selection.

The experiments were implemented on Google Colab [31]. The compute specifications are as follows: Intel(R) Xeon(R), 2.30GHz, 2 Cores. The ML framework used in this research is the Scikit-Learn [32].

Performance metrics are important factors to consider when evaluating the performance of classifiers. In this work, the following performance metrics are considered: accuracy, F1-score, and Area Under the ROC Curve [33‐35]. These metrics are computed using the true positive (TP), true negative (TN), false positive (FP), and false negative (FN):

The Accuracy is the ratio of correctly predicted instances; it is, however, not an effective metric in evaluating classifier performance when the data is imbalanced since it is sensitive to the distribution of the data. The F1-score is a more effective performance metric that represents the harmonic mean of the precision and sensitivity (recall) of the classifier. AUC demonstrates the tradeoff between the true positive rate (TPR) and false-positive rate (FPR), and it is an indication of the model’s ability to classify positive samples correctly. The mathematical representations of the performance metrics are shown below:

This section discusses the results that were obtained after conducting the experiments in a simulated environment.

Table 5 shows the number of features that were selected using IG-FS. In the instance of the Australian dataset, 9 features were selected. For the German dataset, 13 features were selected. In the case of the Taiwan dataset, 17 attributes were picked. These selected features are used for the experiments presented in this proposed study.

Table 6 shows the results that were obtained using the Australian dataset and the Stacked model had the structure and hyperparameters shown in Fig. 2.

In this instance the model that achieved the highest accuracy is the RF model with an accuracy of 87.68%. The model that underperformed in comparison to other estimators is the KNN method with an accuracy of 70.28%, a F1-Score of 60.19%, and an AUC of 0.683. In contrast, the Stacked model achieved the best and most optimal results with an accuracy of 86.23%, an F1-Score of 84.58%, and an AUC of 0.934. These results demonstrated that using a Stacked approach substantially improves the F1-Score and the AUC.

Table 7 outlines the results that were achieved using the German dataset and the structure and hyperparameters of the Stacked model in Table 7 are showing in Fig. 3. The model that performed the best is the Stacked algorithm with an accuracy of 82.80%, a F1-Score of 86.35 %, and an AUC of 0.944. Moreover, the Stacked model outperformed all other methodologies in terms of overall performance. In contrast, the model that underperformed is the KNN method with an accuracy of 68.40%, a F1-Score of 48.82%, and an AUC of 0.547. In terms of accuracy, the other models that performed optimally are the RF, GB, XGB, ANN, and DT with the following scores, respectively: 75.20%, 72.40%, 74.80%, and 73.60%. Table 5 shows the results that were obtained using the Taiwan dataset a.

In terms of accuracy, the method that performed optimally is the RF with an accuracy 87%. In terms of overall performance, the Stacked algorithm achieved an accuracy of 86.23%, a F1-Score of 84.58% and a AUC of 0.934 %. The experiments on the Taiwan dataset demonstrated the same pattern that has been observed on the Australian and German datasets. Using the Stacked-based methodology has proven to produce results that are superior to individual estimators.

In comparison to the research that were proposed in [11] using the German dataset, the proposed Stacked model outperformed the ANN, NB, and KNN by the following accuracy margins, respectively: 5.35%, 5.6%, and 10.6%. The research in [12] considered the AdaSVM and achieved an accuracy of 80% on the Australian dataset. In contrast, our proposed Stacked model obtained an accuracy that is 6.23% higher than the AdaSVM. The research in [13] used ANNs-GD on the Australian and German datasets and obtained accuracies of 78.11% and 68.26%. In comparison to the ANNs-GD, the Stacked model obtained the following superior results using the same datasets: 86.23% and 82.80%. Furthermore, the researchers in [14] used RNNs and obtained AUC 0.782 using the Taiwan dataset. In contrast, the Stacked model obtained an AUC of 0.870 on the same dataset. This represents an increase of 0.088. Additionally, the researcher in [16] used the KNN, RF, and ANN using credit risk datasets such as the German dataset and obtained an accuracy of 76.80%, 72.10%, and 78.6%, respectively. In terms of AUC, the KNN, RF, and ANN achieved 0.788, 0.811, and 0.843, respectively. In contrast, the Stacked method obtained much higher performance results as shown in Table 8. The structure and the hyperparameters of the Stacked model are depicted in Fig. 4.

This research presented the development and implementation of a ML-based credit risk prediction model. This method was implemented using a FS method based on IG in conjunction with a stacking algorithm. These processes were implemented on the Australian, German, and Taiwan datasets. The accuracy, the F1-Score, and AUC were the performance metrics the were used to evaluate the performance of the proposed method. To put the experimental process into context, the following additional ML methods were considered: RF, GB, XGB, KNN, ANN, and DT. The outcome of the numerical experiments demonstrated that the proposed Stacked algorithm achieved an accuracy of 86.23%, a F1- Score of 84.58% and AUC of 0.934 in the instance of the Australian dataset. With regards to the German dataset, the Stacked method obtained an accuracy of 82.80%, a F1-Score of 86.35% and AUC of 0.944. Finally, for the Taiwan dataset, the Stacked method achieved an accuracy of 85.80%, a F1-Score of 51.35 % and AUC of 0.870. These results were superior to those obtained by individual estimators and other existing algorithms. In future work, our aim is to delve deeper into the realm of feature selection and augmentation techniques with the objective of improving the performance of the proposed machine learning model. We envisage a comprehensive investigation into the applicability and efficacy of transformer-based architectures, which have recently gained prominence in various domains such as text generation and classification, to address the intricate challenges associated with credit risk prediction.