Classical statistical models with robust variable reduction techniques
Champion-Challenger mode of model deployment
Models have a high shelf-life
Well established monitoring & reporting systems
Risk compliance framework well defined
Credit Scoring in the Modern World
Lending to unbanked & under-banked
Limited credit history
Lending apps innovatively collect data on unrelated but indicative aspects
Alternate data is often sparse; is different for each smaller homogenous group
Machine Learning Algorithms perform better in such scenarios
Ensemble of models has shown to improve discrimination power
Limited regulatory oversight
Radix Experience in Credit Scoring
100+ years of collective experience in building credit scoring models
Experience in both, traditional and modern approaches
Clients across multiple domains such as Auto loan providers, SME, Telco, equipment
leasing, insurance providers, e-commerce courier
Experience in building various types of scorecards such as application, behaviour,
collection, fraud, attrition, pre-delinquency scorecards etc.
Case Study
Application Scorecard For SME Bank Loan
Issues & Objectives
To develop application scorecard for SME portfolio of an Indonesian bank
The SME portfolio was new to this bank which had been recently acquired by a
multinational bank headquartered in Australia
The scoring model would be used for SME loan origination decisions
Critical component in the plan for scaling up operations in accordance with Indonesian
government directive
Challenges
The number of data points was very small ≈ 400 making it difficult to obtain reliable
results through predictive modeling
Such sparsity of data is not uncommon in Asia
The SME portfolio was new so the history of defaults had not been well established
Solution & Results
Bootstrapping was used to overcome the limitation of a small sample
Reject rates were taken as a surrogate for default rate
Built high quality scorecard
Process Automation
Ensures Consistency in decision making
Predictive modeling replaces gut feel
Scores re-calibrated with default data after sometime
Case Study
Calibration of Expert Scorecard by ML Methods
Issues & Objectives
For the first time in India, a scorecard was developed for the client to keep vigil on the listed companies to avoid potential financial disaster
Scorecard was based on financial as well nonfinancial events such as changes in auditors, board of directors, litigation, news etc.
The task was to refine expert scorecard with ML methods
Challenges
Listed and unlisted flag was incomplete in the database
Many companies had substantial missing data
Frequent modification of event logic
Running ML models and processing score with new weights took several hours posing a challenge to multiple iteration
Solution & Results
Used modern techniques such as Decision tree, Random forest and Gradient boosting to obtain weights of the events
ML methods were run using h2o
Discriminatory power of the calibrated scorecard was found to be higher than the expert scorecard
Applied a decision overlay which enhanced the predictive power of the scorecard
Better separation between GOOD and BAD companies in modelled score
Expert vs Model Performance
Case Study
Credit Scoring for Leasing Company
Issues & Objectives
A large company in UK finances lease of office equipment, primarily to small and medium companies with ticket size less than £10K
Leased items depreciate rapidly and seizure of collateral does not recover the debt
The company currently cherry picks customers who seldom go bad
They want to expand customer base while controlling risk
For this they want a scorecard to replace rule driven underwriting for better screening
Challenges
Company book identified only 2.5% bad lease – payment history data was fraught with inconsistent figures
After incorporating liquidation/insolvency/dissolution status and rating from credit bureau record, the incidence was boosted to 12%. The process classified non takers of loan into Good and Bad by a logical method and not by reject inference
Many financial fields had substantial missing data
Solution & Results
Scorecards with and without credit bureau ratings were delivered
Discriminatory power of the scorecards were found to be high
4 definitions of defaults were considered
Used various modelling techniques to differentiate between Good, Probable Bad and Definite Bad
Scorecard developed by statistical method replaced purely rule based method
Facilitated work of underwriters by restricting scrutiny to limited proposals
Case Study
Fraud Scoring for Insurance Claims
Issues & Objectives
A major insurance company in Singapore used to manually examine each travel insurance claim to identify potentially fraudulent one
Suspicious claims were subject to a more detailed investigation
This involved considerable manual effort & inconsistent processes
The project objective was to develop a score to identify potentially fraudulent claims which would be subject to greater scrutiny
Challenges
Data included 77,445 claim records of which only 120 had been determined to be potentially fraudulent
So identified potentially fraudulent claims are rare events (0.15%) and therefore hard to detect
It was however expected that there could be a large number of undetected fraudulent claims
Solution & Results
A powerful machine learning algorithm, Gradient Boosting, was used for detecting potentially fraudulent cases
Substantial lift demonstrated. – on the client test data set it sufficed to examine 7.75% of all claims to identify 91.67% of all fraudulent claims
Process automation, ensuring consistency, cost saving and increased accuracy
Scrutiny restricted to high scorers reducing manual work by a factor of 5 -10
