Two-Step Logistic Regression for High-Variable Credit Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing credit evaluation models are limited by linear independence constraints, preventing the use of more than 10 explanatory variables and underestimating statistical significance, while machine learning methods lack explanatory power in financial transactions.
Innovation Solution
A two-step logistic regression model that selects important variables through frequency calculations, time windows, and weighted values to generate and apply variables in a step-wise manner, enhancing explanatory power and model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional logistic regression is used to maintain linear independence of explanatory variables, then statistical significance is preserved, but the number of variables is limited to 10 or fewer
Solution Approach 1:
The patent segments the variable selection process into multiple stages: first selecting variables that meet linear independence criteria, then progressively adding additional variables while managing multicollinearity. This segmentation allows the model to incorporate more than 10 variables while maintaining statistical rigor through staged evaluation.
Solution Approach 2:
The patent transforms the traditional single-dimension approach (selecting only linearly independent variables) into a multi-dimensional framework that considers both linear independence and controlled multicollinearity. By adding a second dimension of variable selection that tolerates certain levels of multicollinearity, the model can incorporate additional predictive variables beyond the traditional 10-variable limit.
2Adaptability or versatility
If machine learning methods are used to utilize all available information, then the number of variables can be increased, but explanatory power is lost
Solution Approach 1:
The patent applies local quality by treating different variable groups differently: some variables are selected based on strict linear independence criteria to maintain explanatory power, while other variables are incorporated with controlled multicollinearity to expand information utilization. This localized approach allows different parts of the variable set to serve different functions.
Solution Approach 2:
The patent creates a composite modeling approach that combines traditional logistic regression principles (for explanatory power) with modern variable selection techniques (for information utilization). By compositeing multiple variable selection strategies, the model achieves both interpretability and comprehensive information usage.
3Adaptability or versatility
If more than 10 explanatory variables are used in traditional logistic regression, then variable coverage increases, but linear independence is violated and statistical significance is underestimated
Solution Approach 1:
The patent performs preliminary action by pre-screening variables for linear independence before incorporating them into the final model. This preliminary filtering ensures that only variables meeting statistical criteria are selected, preventing underestimation of statistical significance while still allowing comprehensive variable coverage through multiple selection rounds.
Solution Approach 2:
The patent implements feedback mechanisms where the impact of each added variable on statistical significance is continuously evaluated. If adding a variable causes excessive multicollinearity or significance underestimation, the system provides feedback to adjust the selection criteria or remove problematic variables, thereby maintaining measurement precision while maximizing variable coverage.
Data Source
AI summary
A credit evaluation model operating method performed by a credit evaluation server linked to a financial server, the credit evaluation model operating method comprising, a step of receiving log data of a user and selecting basic variable items included in the log data, a step of generating candidate variables by calculating a frequency of the basic variable items in the log data, a step of generating a plurality of first derived variables by applying different time windows or different calculation methods to the candidate variables, a step of selecting important variables by comparing values related to the plurality of first derived variables with a predetermined standard value, a step of deriving a first-step model by using the important variables as input variables and using information on the user's credit as a dependent variable, a step of selecting a first final variable to be applied to the first-step model among the important variables and calculating a first weighted value for the first final variable, a step of generating a second derived variable by using the first final variable and the first weighted value, a step of deriving a second-step model by using the second derived variable as an input variable and using information on the user's credit as a dependent variable, and a step of selecting a second final variable to be applied to the second-step model from among the first derived variables and calculating a second weighted value for the second final variable.