Modeling consumer loan default prediction using ensemble neural networks

Abstract

Abstract-In this paper, a loan default prediction model is constricted using three different training algorithms, to train a supervised two-layer feed-forward network to produce the prediction model. But first, two attribute filtering functions were used, resulting in two data sets with reduced attributes and the original data-set. Back propagation based learning algorithms was used for training the network. The neural networks are trained using real world credit application cases from a German bank datasets which has 1000 cases; each case with 24 numerical attributes; upon, which the decision is based. The aim of this paper was to compare between the resulting models produced from using different training algorithms, scaled conjugate gradient backpropagation, Levenberg-Marquardt algorithm, One-step secant backpropagation (SCG, LM and OSS) and an ensemble of SCG, LM and OSS. Empirical results indicate that training algorithms improve the design of a loan default prediction model and ensemble model works better than the individual models.Index Terms -credit risk, loan default, neural network, scaled conjugate gradient backpropagation, Levenberg-Marquardt algorithm and One-step secant backpropagation.

Effect of Hidden Neuron Size on Different Training Algorithm in Neural Network

“…They also showed that OSS is bad choice for three phase power quality monitoring. Hassan and Abraham [17] showed that OSS was slowest algorithm.…”

A Boosted Decision Tree Model for Predicting Loan Default in P2P Lending Communities

“…Peer to peer lending [3], [5], [7], [11], [12], commercial banking [2], [4], [13]- [15], insurance [6], agriculture [16], mortgage [17], and small and medium enterprises (SMEs) [8], [12] are different application areas of loan default prediction studies. However, because of certain specific problems and different available dataset, the studies employed different machine learning models.…”

“…This, as shown, suggests the type of applicable machine learning model, especially where the size of the available dataset determines the performance. The data features used in building the models reviewed range from eight (8) [2], seventeen (17) [4] to twenty-four (24) [13]. Studies with considerable large datasets are Bagherpour [18] of about 20 million loan observations between 2001 -2006, and Rivet [6] which used 1 million loans data.…”

“…Studies with considerable large datasets are Bagherpour [18] of about 20 million loan observations between 2001 -2006, and Rivet [6] which used 1 million loans data. Yang et al [3], which employed a dataset of 10,000 anonymous users of a peer lending service, Sivasree and Sunny [4] of 4520 data instances for loan credibility prediction system, and Hassan and Abraham [13] with 1000 cases are others with small data sizes. In these, K-Nearest Neighbors [18], Decision Tree [4], [6], [18], Random Forest [18], logistic regression [5], [6], Support Vector Machine [18], Artificial Neural Network [3], [13], [19], and Naïve Bayes [2], [11] are the recorded machine learning models.…”

“…Yang et al [3], which employed a dataset of 10,000 anonymous users of a peer lending service, Sivasree and Sunny [4] of 4520 data instances for loan credibility prediction system, and Hassan and Abraham [13] with 1000 cases are others with small data sizes. In these, K-Nearest Neighbors [18], Decision Tree [4], [6], [18], Random Forest [18], logistic regression [5], [6], Support Vector Machine [18], Artificial Neural Network [3], [13], [19], and Naïve Bayes [2], [11] are the recorded machine learning models. Asides from studies that did comparative study for loan default prediction performances [2], [5], [18], others employed single [4], [17] and hybrid/ensemble models [3], [6], [8], [9].…”