Automatic model optimization algorithm based on stepwise optimal feature selection

Abstract

The invention discloses an automatic model optimization algorithm based on stepwise optimal feature selection. The algorithm comprises seven steps of establishing a primary scoring model, setting a limiting condition, gradually optimizing a feature selection algorithm, screening to obtain an optimal model KS value, judging whether a model KS difference variable combination reaches the standard ornot, re-screening the non-standard KS difference variable combinations, and establishing a final scoring card model. According to the algorithm, an optimal scoring card model can be automatically generated, the prediction capability of the feature variables on the target variables is mined to the maximum extent, the interference of the subjective consciousness of modeling personnel is eliminated,the waste of the time cost is reduced, and the finally generated scoring card model gives consideration to the effectiveness of a model prediction effect and the stability of a model application effect. The algorithm is embedded into the establishment of a traditional scoring card model, so that the quality of a traditional credit model in the financial industry can be obviously improved, the approval accuracy is improved, and more fraud overdue applications are rejected.

Description

technical field [0001] The invention belongs to the technical field of Internet financial risk control, and in particular relates to an automatic model optimization algorithm based on step-by-step optimal feature selection. Background technique [0002] The credit scoring model is based on various historical credit data of banks or Internet financial customers to obtain credit scores of different grades. According to the customer's credit score, the credit institution can determine whether to grant credit and credit by analyzing the possibility of customer repayment after the loan amounts and interest rates. [0003] Traditionally, banks or financial institutions adopt manual approval methods, and make subjective approval judgments based on the personal experience of the approvers, making the approval decision easily influenced by subjective factors, resulting in inconsistent approval results, unable to quantify the risk level, and unable to achieve risk control. Hierarchic...

AI Technical Summary

Problems solved by technology

[0007] 1. The traditional model tuning method does not achieve the optimal effect of the model, and cannot maintain the stability of the model when the effect is better. It is easy to ignore some meaningful characteristic variables, and it is impossible to select the variable with the best combination of all variables. Enter the model, resulting in the loss of information;

[0008] 2. Traditional model tuning methods are relatively subjective and do not have uniform applicability, and the effects of different modeling engineers are not stable;

[0009] 3. The traditional model tuning method takes a long time and must rely on the interactive operation of the modeling engineer, which cannot realize the automatic establishment of the scorecard model

Images