Internet financial risk control model based on XGBoost

Abstract

The invention provides an Internet financial risk control model based on XGBoost. The Internet financial risk control model based on XGBoost comprises the following steps: S1, extracting an appropriate modeling sample client; s2, obtaining online loan data of a sample client, and extracting a feature variable corresponding to the online loan data; s3, defining 'good ' and 'bad' of the modeling sample according to the repayment behavior of the customer, the quality of the target customer group and the product type; s4, data processing, including dirty data cleaning, missing value processing andabnormal value processing; s5, feature engineering including feature construction and feature screening; s6, data set division: randomly or cross-time dividing a training set and a verification set;s7, performing training by applying an XGBoost algorithm, and performing model parameter adjustment; and S8, evaluating the model: evaluating the quality of the model according to the evaluation indexes. On one hand, third-party data are additionally used, the dimensionality of risk identification is increased, and meanwhile, the efficiency and robustness of a model algorithm are optimized throughan XGBoost algorithm with high prediction capability; on the other hand, the accuracy of the model is continuously optimized through XGBoost algorithm parameter adjustment and model evaluation, and the method is more suitable for the demand of big data risk control.

Description

technical field [0001] The invention relates to the technical field of Internet financial risk control, in particular to an XGBoost-based Internet financial risk control model. Background technique [0002] my country's domestic P2P is developing rapidly, and Internet finance such as cash loans and consumer credits have sprung up like mushrooms after a spring rain. However, Internet financial risk control using big data is a relatively new topic in China. Most Internet financial companies still use the traditional American FICO score Card model risk control modeling. The traditional scorecard model is a linear model, which is characterized by a linear relationship between the dependent variable and the independent variable. It is simple and easy to explain, the model performance is stable, the degree of overfitting is low, it is easy to parallelize, and it can easily process hundreds of millions of data. The model can only explain the linear relationship between variables, a...

AI Technical Summary

Problems solved by technology

But its shortcomings are also obvious. One is that the logical relationship inside the model is as complex as a black box and cannot be explained, and it is prone to overfitting, which means that the predictive ability of the model on the predicted data or in practice will decline rapidly. It needs to be constantly updated, and the stability is not enough; the second is that it consumes a lot of space, and needs to save the characteristic values ​​​​of the data and the results of the characteristic sorting

[0004] In summary, traditional scorecard models and machine learning models are two extremes in terms of model complexity, predictability, and stability

Images