Financial default probability prediction model based on LightGBM

Abstract

The invention provides a financial default probability prediction model based on LightGBM. Modeling of the method comprises the steps of sample data acquisition, data preprocessing, feature engineering, data set division, model training and parameter adjustment, and model deployment and monitoring, and automatic, comprehensive and process credit risk assessment and prediction of borrowers are realized, so that credit overdue fraud risks are reduced, and improvement of the Internet financial risk control capability and healthy development are promoted. The core of the prediction model of the technical scheme is as follows. According to the method, a LightGBM model based on a histogram algorithm and a level-wise splitting strategy is utilized; the method is advantaged in that internet financial mass data with many abnormal values, high missing values and wide dimensions can be optimized and rapidly processed, a big data processing capability is realized, because of excellence of the algorithm, reliability, flexibility and expandability of the model are all improved, and the method is more suitable for present big data risk control demands.

Description

technical field [0001] The invention relates to the technical field of Internet financial risk control, in particular to a financial default probability prediction model based on LightGBM. Background technique [0002] In recent years, Internet finance represented by P2P lending and consumer finance has developed rapidly. Most Internet finance companies based on logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting decision tree (GBDT), Algorithms such as extreme gradient boosting tree (XGBoost) are used for risk control modeling. Generally, the risk control model mainly adopts WOE conversion, and then uses logistic regression model for fitting to construct a credit score card. This method is more effective than traditional financial industries in Internet finance. The effect has declined. With the development of big data, Internet financial risk control and machine learning complement each other. It is very meaningful to use more adv...

AI Technical Summary

Problems solved by technology

[0004] Big data is worthy of its name. The data dimension shows explosive growth, with many dimensions and high sparsity. In risk control modeling, structured data cleaning and processing are heavy, data transformation has sparse matrix and too much loss information, feature extraction is difficult, tens of thousands of Dimensionality exceeds the range that traditional scorecard models can handle, and the requirements for machine learning algorithms are getting higher and higher. Algorithms such as LR, SVM, RF, GBDT, XGBoost, and LightGBM have emerged as the times require. XGBoost is currently the mainstream algorithm, and XGBoost On the basis of traditional Boosting, integrate the advantages of RF and GBDT, use the multi-thread parallelization of CPU, introduce regularization items, support column sampling and parallel approximate histogram algorithm, etc., but in practical applications, it also presents a large amount of calculation Huge, slow training speed, high memory usage, easy to produce overfitting and other shortcomings, which gave birth to a more advanced algorithm such as LightGBM

Images