Online Loan Default Risk Identification for Small Businesses Based on Samples Weighting

Reliable models have been urgently needed to measure the loan default risk of small businesses. Using part of the 2015-2017 PPDF online lending data as the sample, this study develops an online loan default risk identification model for small businesses with a machine learning approach. The dataset employed is from Model Whale. On the sample dataset, a maximum information value method is employed to bin features. And then the one-hot encoding technique is used to split the original features into feature variables by their numerical size, based on which we train a modified stacking heterogeneous ensemble model. Each sample point in the training set is given a weight based on its contribution to the forecasting accuracy in the 5 base classifiers in level-0 model. Specifically, k-fold cross-validation is employed, where 5 base classifiers are used to classify each sample in the train set individually. The weight assigned to each sample point depends on the times of correct identifications for this very point. The aim is to adjust adaptively the influence of these points in fitting the aforementioned model. Our work shows that, the proposed stacking model with weights achieves significantly higher accuracy in default risk identification than the conventional stacking model or other commonly used ones.