Slide 1

Slide 1 text

2019 DevDay LINE Score: How To Build Alternative Credit Scoring Model > Sangwoo Kim > LINE Plus

Slide 2

Slide 2 text

> Sangwoo Kim > Manager of Credit Scoring & Risk Management team > Responsible for developing LINE Score This session will benefit those who are interested in data analytics, especially credit score modeling Hello Developers and Data Scientists!

Slide 3

Slide 3 text

> Developing a scoring model > Credit scoring system Credit scoring model Building a scoring model

Slide 4

Slide 4 text

Agenda > Key Concept > LINE Score as an Alternative Credit Score > Overview of Credit Score > Techniques for Credit Score Modeling > Measurement methodology for performance > Achievements of LINE Score

Slide 5

Slide 5 text

Key Concept

Slide 6

Slide 6 text

Predictive Model vs. Credit Scoring Model Predictive Model The general predictive model usually used in non-financial companies to achieve a specific purpose. > Mostly focusing on efficiency or effectiveness > Indirectly affects businesses or companies Credit Scoring Model The credit scoring model usually used in financial institutions or credit bureaus to evaluate a customer’s creditworthiness. > Mostly focusing on default > Influenced by various regulations > Directly affects the profit or losses of financial institutions

Slide 7

Slide 7 text

Credit Score vs. Alternative Credit Score Credit Scoring Model The credit scoring model is a traditional evaluation method that usually uses financial data. > Uses financial data for scoring > Mainly used in countries with advanced financial industries > Only for those who have a history of financial transactions in the past Alternative Credit Scoring Model The alternative credit scoring model is an evaluation method that is in the limelight recently. It usually uses non-financial data. > Uses non-financial data for scoring > Available in all countries > Possible for those who don't have a history of financial transactions in the past

Slide 8

Slide 8 text

LINE Score

Slide 9

Slide 9 text

“Enrich daily life” LINE Score

Slide 10

Slide 10 text

Evaluate credit score by analyzing behavior data on the LINE platform LINE Score will only be calculated upon the user's consent

Slide 11

Slide 11 text

Various services are running on the LINE platform Big Data

Slide 12

Slide 12 text

An ecosystem in which data continuously flows into LINE scores can be utilized in various businesses at the same time Creating the LINE Score Platform Data source Financial Institutions Sharing Economy C2C Platforms Reservation Services Rental Businesses Score platform

Slide 13

Slide 13 text

Overview of Credit Scoring

Slide 14

Slide 14 text

> A credit scoring model for evaluating customers who apply for loans Application Scoring System Behavior Scoring System > A credit scoring model for evaluating the behavior of existing loan customers Credit Scoring System > A statistical model that utilizes a customer's past history to predict creditworthiness in the future Key Terminology

Slide 15

Slide 15 text

> Descriptive information used to predict defaults Feature or Variable (Input) ‘Build’ or ‘Develop’ credit scoring model > A process of developing a statistical model that predicts 'defaults' using 'features' Default (Target) > Mainly uses a loan delinquency history of more than 60 days or 90 days as target information Key Terminology

Slide 16

Slide 16 text

Basic Concept of Credit Scoring Model How to Build a Credit Scoring Model Past default record (y) Past transaction history (x) Fitting the best algorithm to predict default y = f (x) How to Operate a Credit Scoring Model Funtion Loan applicant's future possibility of default Loan applicant's past transaction history Selected features (x) Probability of default (y)

Slide 17

Slide 17 text

Techniques for Credit Score Modeling

Slide 18

Slide 18 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition

Slide 19

Slide 19 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Establishing the purpose of the model and selecting groups or subjects to be used for model development Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition

Slide 20

Slide 20 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition Selecting the target value that needs to be predicted

Slide 21

Slide 21 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition Creating and selecting features that are highly predictable and highly correlated with the target

Slide 22

Slide 22 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition Exploring the highly predictable and optimal algorithms suitable for the purpose of model development

Slide 23

Slide 23 text

Process Overview The credit scoring model building process can be divided into 5 steps as follows: Feature creation & selection Model fitting Scaling & calibration Purpose of model Target definition Converting to a score or grade form that is easy for customers to understand

Slide 24

Slide 24 text

Several types of credit scoring models for different purposes > Behavior Scoring Model > Collection Scoring Model > Application Scoring Model Purpose of Models > Fraud Detection Model > Early Warning Model

Slide 25

Slide 25 text

Set up a target that fits the model’s purpose Target Definition > Behavior Scoring Model > Collection Scoring Model > Application Scoring Model > Fraud Detection Model > Early Warning Model → Event of default → Event of default → Possibility of collection → Event of fraud → Event of issue related to credit risk

Slide 26

Slide 26 text

Feature Creation and Selection Examples of candidate feature creation > Type of mobile phone > Number of received or sent calls in a day/week/month/year > The ratio between sending and receiving calls/messages > Main time zone

Slide 27

Slide 27 text

Factors to consider when selecting features > Information value > Stability > Diversity > Correlation Feature Creation and Selection

Slide 28

Slide 28 text

Traditionally, logistic regression has been widely used Recently, machine learning has also been used to increase predictability Model Fitting Logistic Regression A model usually can be derived in the form of a highly-explainable scorecard > Higher Interpretability > Simple > Relatively small amounts of data > Relatively low performance Machine Learning Algorithm A model usually can be derived in the form of a highly-complex and hard-to- interpret black box > Lower interpretability > Complicated > Requires a large amount of data > Relatively high performance

Slide 29

Slide 29 text

Process for improving user utilization and understanding Scaling and Calibration Scaling Calibration Probability of Default Score 0.000% ~ 0.124% 1,000 ~ 901 0.125% ~ 0.249% 900 ~ 876 0.250% ~ 0.374% 875 ~ 851 0.375% ~ 0.499% 850 ~ 826 0.500% ~ 0.624% 825 ~ 801 0.625% ~ 0.749% 800 ~ 776 0.750% ~ 0.874% 775 ~ 751 … … Score Range Grade 820 ~ 1000 1 800 ~ 819 2 780 ~ 799 3 750 ~ 779 4 730 ~ 749 5 700 ~ 729 6 650 ~ 699 7 … …

Slide 30

Slide 30 text

Measurement Methodology for Performance

Slide 31

Slide 31 text

Measuring the performance of a model in terms of stability and discriminability How to Measure Model Performance Stability Indicator The difference in the distribution between the development data set and validation data set, which is used to verify the stability of the features, score, and grade > PSI Discriminative Power A numerical measure of how well rank ordering was done on predicting the target, which is used to show the discriminant power of the features, score, and grade > AUROC (AR) > K-S

Slide 32

Slide 32 text

Discriminant Power Indicator: AUROC False Positive Rate True Positive Rate Perfect Model Developed Model Random Model Area Under the ROC = A+B A B

Slide 33

Slide 33 text

Assume that there are 5 customers in default out of 100 customers, and 5 customers in default are of the 5th, 10th, 20th, 25th, and 30th lowest scores Example of AUROC Calculation False Positive Rate True Positive Rate Random Model A

Slide 34

Slide 34 text

Assume that there are 5 customers in default out of 100 customers, and 5 customers in default are of the 5th, 10th, 20th, 25th, and 30th lowest scores Example of AUROC Calculation False Positive Rate True Positive Rate Developed Model Random Model A Perfect Model 5 10 20 25 30

Slide 35

Slide 35 text

Assume that there are 5 customers in default out of 100 customers, and 5 customers in default are of the 5th, 10th, 20th, 25th, and 30th lowest scores Example of AUROC Calculation False Positive Rate True Positive Rate Developed Model Area Under the ROC = A+B Random Model A Perfect Model 5 10 20 25 30 B

Slide 36

Slide 36 text

Achievements of LINE Score

Slide 37

Slide 37 text

Achievements of LINE Score Offer Save Increase many benefits based on LINE Score costs by providing more reasonable financial services accessibility to financial services

Slide 38

Slide 38 text

Thank you