**3. Methods**

In this section, we present the novel CT-XGBoost prediction model with cost-sensitive and threshold methods. Figure 1 shows how XGBoost is modified into CT-XGBoost in this paper. In XGBoost, the misclassification costs for both classes are the same, and the threshold is simply set as 0.5. Thus, we improved XGBoost, turning it into CT-XGBoost, by solving two challenges: How to assign misclassification costs for the two classes properly. How do you set the threshold rationally? In this paper, two corresponding strategies (cost-sensitive strategy and threshold method) are adopted to overcome the challenges, misclassification cost is determined based on the imbalance ratio of the dataset, and a threshold is set considering the number of different classes samples. Then, XGBoost is modified into CT-XGBoost systematically.

**Figure 1.** The process of modifying XGBoost into CT-XGBoost.

In order to introduce CT-XGBoost logically, we first explain the theory of XGBoost and then illustrate how we modify XGBoost into CT-XGBoost. After that, the commonly used default prediction models are introduced, which are used to compare with our proposed model. Lastly, performance evaluation methods of the credit default prediction are explained.
