Best practices in machine learning for small business lending - Sip and Solve

50 %
50 %
Information about Best practices in machine learning for small business lending - Sip and...

Published on October 17, 2020

Author: ExperianBIS

Source: slideshare.net

1. 1 © Experian Sip and Solve Best practices in machine learning for small business lending Featuring: Matt Shubert Director of Data Science Experian Business Information Services

2. 2 © Experian Machine learning and small business lending

3. 3 © Experian Machine learning Subset of artificial intelligence including both supervised and unsupervised learning

4. 4 © Experian Machine learning Subset of artificial intelligence including both supervised and unsupervised learning

5. 5 © Experian Not Interpretable Not Transparent Not Explainable Difficult to Implement Machine learning obstacles Barriers to small business lender adoption

6. 6 © Experian Not Interpretable Not Transparent Not Explainable Difficult to Implement Black Box Machine learning obstacles Barriers to small business lender adoption

7. 7 © Experian • Continuous learning models – Benefits • Limited degradation • Reduced redevelopment and deployment efforts – Issues • Lack of stable outputs • Governance is difficult to maintain • Static learned models – Benefits • Stable outputs for regulatory compliance • Improved governance controls – Issues • Degradation likely to occur • Longer cycle to redevelop and deploy Machine learning Continuous vs. Static models

8. 8 © Experian • Support vector machines • Neural networks • Random forests (tree-based method) • Gradient boosted trees (tree-based method) Machine learning Variety of supervised classification methods

9. 9 © Experian Decision trees • Easy to explain decisions • Simple to understand • Limited regulatory scrutiny • Efficient for scoring processes • Easy to implement • Basis for machine learning tree-based methods * Example is simplified for illustration purposes. Decision Trees typically have more depth/width to account for all the possible outcomes. Tree based methods have decision tree roots Foundation of tree-based methods

10. 10 © Experian Gradient boosted trees • Improves predictive power with corrective series of shallow decision trees • Superior performance compared to traditional credit risk modeling methods • Transparent and explainable at each ending tree split • Interpretable at individual and summarized variable levels • Monotonicity enforcement creates consistent explanations throughout the process • Less regulatory scrutiny due to decision tree based roots 10 Gradient boosted trees Tree-based machine learning method

11. 11 © Experian Machine learning Small business lending applications • Origination • Marketing • Fraud • Portfolio Management • Optimization • Collection Recovery

12. 12 © Experian How to implement machine learning?

13. 13 © Experian Aging Infrastructure Lack of Scalability On Demand Availability Expensive to Deploy Barriers to deployment Difficult to implement

14. 14 © Experian Aging Infrastructure Lack of Scalability On Demand Availability Expensive to Deploy Technology Modernization Required Barriers to deployment Difficult to implement

15. 15 © Experian Mainframe Data warehouse Legacy systems and processes Big data capable distributed processing on demand and batch real time support scalable Modernized Technology Platform Legacy technology impedes machine learning Technology modernization Modernization

16. 16 © Experian To partner or not to partner?

17. 17 © Experian • Time to modernize technology platform = 2-3 years • Time to develop machine learning model = 6-12 months • Time to deploy machine learning model = 3-6 months • Time to gain model risk management approvals = 6-12 months • Total time to start using = 3-6 years Small business lender with $10+ billion in assets Time to realize machine learning value without a partner

18. 18 © Experian • Time to evaluate on your portfolio = 3 months • Time to gain model risk management approvals = 3-6 months • Total time to start using = 6-9 months Small business lender with $10+ billion in assets Time to realize machine learning value with a partner

19. 19 © Experian Summary/Key takeaways • Not all machine learning methods are created equal – Tree based methods can be explainable to achieve regulatory compliance • Modernization – Aging technology infrastructures need to be upgraded • Partnership may be the easiest route to quick adoption and value realization

20. 20 © Experian Thank you! Have questions or want to learn more about machine learning for small business lending? • Send us a message in the “Chat” function or via: bit.ly/sip-solve-chat Upcoming Sip and Solves - • Join the list to stay in the loop: bit.ly/sip-and-solve

Add a comment