Personal Credit Risk Prediction Based on Minimum Weight Value Error Combination Model

In the era of big data, personal credit data has become increasingly complex. This paper proposes a modeling method based on the minimum weighted value error, which integrates sequence models and non-sequence models. Firstly, the chi-square test and Gradient Boosting Decision Trees (GBDT) are utilized to select features, reducing redundancy in personal credit data. Next, the sequence information within this data is treated as language text data. The Transformer architecture, built on BERT pre-training, is employed to create a sequence model for time series features. Concurrently, non-sequence features are constructed using Deep Neural Networks (DNN) and Multi-Layer Perceptrons (MLP). Finally, the minimum absolute error method is enhanced, leading to the proposal of an improved minimum weighted value error method. This method facilitates dynamic weight allocation between the sequence and non-sequence models, resulting in a more accurate integrated model. Experimental results demonstrate that the classification accuracy of this model on the validation set significantly surpasses that of individual Transformer, DNN, and MLP prediction models.