Knock-out Rule Optimization

Historically banks have relied on expert scorecards for scoring their credit applications. This meant setting up of knock-out(KO) rules (e.g. age of the applicant > X years, applicant should be employed etc.). However for efficiency reasons there is a requirement to reduce the no.of KO rules applied, while keeping the default rate low. In this talk we will see, how a Multi-objective Optimisation method (NSGA-II) can be used to reduce the no.of KO rules while keeping the default rate as low as possible.

Transcript

1. Knock Out Rule Optimization NSGA-II Online, 28th Apr 2020 Data Science Meetup
2. • Lending Process and Knockout rules. • Optimization Objectives. • NSGA-II. • Results. • Questions. Agenda 2
3. Customers/Potential customers apply for a loan either via online/offline channels. Scrutiny by the bank: • Knock out rules. • Decision models. • Operational/ Risk managers Disbursement to client. 3 Typical Lending Process | Flowchart Application Scrutiny by Bank Loan disbursed
4. • Expert-based decision rules for eliminating high risk customers. E.g. : Loan amount <= 20x monthly net income. Monthly Income >= 1000 Eur. Credit bureau score >= Cutoff score etc. • In addition to the credit decision model, knock-out (KO) rules are applied during the lending decision process in order to decide whether a loan application is accepted or rejected. • These KO rules are either applied before or in conjunction with the decision models. • ~ 50-60 rules for primary and secondary applications. • Some KO rules are automatically checked in the IT systems during each scoring step others may require manual processing. Knock Out Rules | Introduction 4
5. Knock Out Rules | Good & Bad 5 Pros • KO rules were created by expert knowledge, so they are effective both for expert knowledge based scoring models & machine learning models. • Some KO rules are mandatory from a regulatory perspective .(e.g. rules related to nationality, minimum age of the applicant etc.) Cons • KO rules disqualify a considerable no. of applications, which reduces genuine lending opportunities. • Many KO rules introduce processing overhead leading to longer time to disbursement.
6. We focus on the following objectives while optimizing KO rules: • Decrease the number of KO rules. • Increase the number of loans disbursed. • Decrease/Maintain the default rate of the overall applications. These objectives are competing against each other e.g. reducing the no.of KO rules will lead to an increase of default rate of the applicants . KO Optimization | Competing Objectives 6
7. • Multi-objective Optimization algorithm based on Genetic algorithm. • Proposed by Deb K and others in 2002. • Provides a set of optimal solutions. Pareto fronts ( Non-Dominated Fronts ) : Set of solutions, that are better than or equal to other optimal solutions. Crowding Distance: When solutions lie in the same NDF, solutions from the less crowded region are selected. NSGA-II | Non-Dominated Sorting with Genetic Algorithm 7
8. NSGA-II | Flow Chart 8 Start Create a random initial population Evaluate fitness for each population Optimal Solution found ? Select desired population through Fast NDS and Crowding Distance Mutate Organisms Stop Yes No
9. 9 KO Optimization | IRL 1. KO Rules + Objectives 2. Decision Model 3. Data Multi Objective Optimization Optimized set of KO rules
10. Data : Vehicle Loan Default All KO Rules : Optimized KO Rules : KOR Optimization | Demo 10 # Rules No.of Application passed Default Rate 10 26239 1.15 % # Rules Rules No. of Application passed Default Rate 1 perform_cns_score >= 200 31054 1.18 % 2 avg_acct_age_m >= 6 perform_cns_score >= 200 27083 1.13 %
11. 11 RPAA Team | We are hiring
13. Code : https://github.com/anilkumarpanda/kooptimize NSGA : https://www.youtube.com/watch?v=Hm2LK4vJzRw Flaticon Icons : https://www.flaticon.com/authors/eucalyp Python Packages for Evolutionary Multiple Objective Optimisation: https://pymoo.org/ https://deap.readthedocs.io/en/master/ References & Attributions 13