Predictive models for P2P lending

Predictive models for loans approval/rejections, loans default and risk calculation, and borrowers rating classification using data from www.prosper.com

Transcript

1. Predictive Models For Prosper Peer-to-Peer lending platform Jeyaram Ashokraj
2. P2P Lending •Crowd sourced funding •No intermediaries •Investment Options •Lending Club •Prosper
3. Prosper Platform •Prosper has a rich dataset when compared to Lending Club. •Historical data from its inception(2005). •Prosper provides observations across 7 objects •Listings, Loans, Groups, Categories, Marketplaces, Members
4. How it works? Prosper Platform Borrower Investor $$$.. Creates Listing Places BID Please read the offer documentation carefully  Monthly EMI Investment
5. Objectives •Will my loan get approved? •Will my investment default? •What rating a borrower gets?
6. Prosper Data •3.2 GB XML •2.3 million records and 70+ variables •Subset 2008 (95k) and 2013(50k) data •~70 variables
7. Variables Predictors Response Quantitative AmountRequested, BidCount, EstimatedLoss, LenderYield, ProsperScore, DebtToIncomeRatio, OnTimeProsperPayments, ProsperPaymentsLessThanOneMonthLate, ProsperPaymentsOneMonthPlusLate, AmountFunded , AmountRemaining, BidMaximumRate, BorrowerMaximumRate, BorrowerRate, Category, Duration,ActiveProsperLoans, TotalProsperLoans, ProsperPrincipalBorrowed, ProsperPrincipalOutstanding, TotalProsperPaymentsBilled, CreditScoreRangeLower, CreditScoreRangeUpper, MonthlyLoanPayment, BankDraftFeeAnnualRate, GroupLeaderRewardRate, PercentFunded, Term Categorical HasVerifiedBankAccount, IsBorrowerHomeowner, FundingOption,City,State,GroupName,GroupRating ProsperRating, ListingStatus, LoanStatus
8. Key Variables •BidCount •ProsperScore •CreditScore •Debt to Income Ratio •OnTimePayments
9. GBM Relative Influence
10. Loan Default Prediction • 2013 data didn’t work out. • Loan term was 3 and 5 year • Binary Classification • Response: LoanStatus • Defaulted, Complete • Random forests • 86 % prediction accuracy
11. Confusion Matrix
12. Loan Approval Prediction •Approved/Rejected ? •Binary Classification •Response Variable: ListingStatus • Completed, Cancelled •Random forests didn’t work (more than 32 categories) •Naïve Bayes Classifier •87% prediction accuracy
13. Borrower Ratings •7 Ratings •Key for Investment •AA least risky •HR high risky •Prosper Algorithm
14. Borrower Ratings • Multiclass Classification • Response: ProsperRating • Completed - Naïve Bayes • 90% accuracy • In-Progress - Gradient Boosted Trees, SVM
15. Confusion Matrix
16. Tools •Grep, Sed, AWK •Libxml – python •MySQL •R – Caret, e1071, gbm, libsvm •Rattle •Weka
17. Future Scope •Improving prediction accuracy – Analyzing individual credit profile data •Money flow – Analyzing bids placed •Impacts of Social Networking (Friends, References) •Influence of groups and categories
18. Thanks for your patience