Loans with LendingClub: Predicting Loans with Positive ROI
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
Introduction
San Francisco-based LendingClub is a peer-to-peer lending institution founded on Facebook in 2007. It counts as a digital-first investment platform with its algorithms mainly focused on matching investors with private loans. The company receives income through 1% - 6% origination fees to the borrower and 1% service fees to the investor.
- Borrower requests unsecured installment loans of 3 or 5 years between $1,000 and $40,000 by posting a qualifying loan request on the Site.
- Client buys fractional loans as an unsecured lender after reviewing multiple criteria through a search function.
- Multiple clients share individual notes with a $25 minimum buy-in and receive monthly principal and interest payments.
Brief History of LendingClub
2007: Founded on Facebook, LendingClub received $10.26 million in a Series A funding round in August 2007.
2007 - 2014: PRE-IPO GROWTH, In Nov 2012, LendingClub surpassed $1 billion in loans issued since inception and announced cash flow positive.
12/10/2014: IPO, Viewed as a pioneer in fintech industry, the company was the largest technology IPO of 2014 in the US, raising $1 billion.
2016: CEO SCANDAL, Founder Renaud Laplanche was ousted as CEO for issues with data integrity and contract approval monitoring and review processes.
Dec 2018: FRAUD CHARGES, LendingClub was ordered to pay a $4 million penalty by the SEC for misrepresenting the quality of some of its loans.
Lending Club Issues
Though viewed as the gold standard in fintech startups, LendingClub experienced problems in early 2016, with difficulties in attracting investors, a scandal over some of the firm's loans. The scandal has taken a major toll on the company’s NYSE stock listing, with shares trading at around 10% of the IPO price in December 2019.
Part 1: Loan Approval Criterion Reverse Engineering
Rejected Loans
-
Observations: 20M (2007 – 2018Q4)
-
9 Features: Before Loan Issued
-
Basic Applicant Information
Accepted Loans
-
Observations: 2M (2007 – 2018Q4)
-
151 Features: Before / After Loan Issued
-
Basic Applicant Information | Detailed Applicant Information
-
Loan Details | Payments | Loan Status | Settlement | Hardship
Merge & Process
-
Selected 7 Features: Loan Amount | Purpose | Dti | State | Employment Length | Risk Score | Application Month/Year
-
Assumptions: Risk Score in Reject = Average FICO score in Accept | Application Month in Reject = Loan Issued Month in Accept
-
Feature Engineering: Impute missing values (Dti/Employment Length/Risk score) | Dummify: Loan Purpose (classify to 8 categories ) / State
Challenges
-
Data size beyond coverage of supervised / unsupervised algorithms
-
Highly imbalanced classification for supervised techniques
-
Unmatched data standard between rejected and accepted loans
Solutions
-
Start with small samples and push the size up to 200,000
-
Down sample majority class / Try different models / Logistic balanced argument
-
Make simplifying assumptions
Unsupervised Learning - Clustering
Supervised Models
Models with higher accuracy have more type I error than type II error.
Feature Importance
Credit Policy Analysis Based on reverse engineering
-
Risk Score: Minimum FICO score of 660
-
Employment Length: 60% of Accepted Loan Application has at least 5 years of employment length. Only 9% of Accepted Loan Application has less than 1 year of employment length, and it has average FICO score of 700.
-
DTI: More than 98% Accepted loans with Debt-to-income ratio of below 40%
-
Purpose: For accepted loan application, more than 50% is for Debt Consolidation; more than 20% is for Credit Card. Loans with Other Purpose have higher possibility of being rejected
-
Loan Amount: Accepted loans have higher average amount than rejected loans
-
Conclusion:
-
LendingClub’s credit policy is straight-forward.
-
Prediction models can achieve high accuracy with limited features.
-
Return of Investment
-
Modern Portfolio Theory was improvised for assembling portfolio of assets. It is a formalization and extension of diversification of investing, consider each investment contributes to a portfolio’s overall risk and return
-
MPT assumes that investors risk averse, given two portfolios that offers. The same expected return, investors will prefer the less risky one.
-
In general, the expected return is calculated by:
-
Distribution of ROI Curve via Monte Carlo Simulation
- 20% of probability investors will end up with a negative return
Predicting Fully Paid vs Charged-Off/Default Loans
In order to pick out loans with positive returns, the highest factor are picking out loans that first pays off the loan completely therefore machine learning models were used predict which loans would come to full payment versus charged off.
The models that were used are listed below with its accuracy and precision scores.
Random Forest Classifier
- Accuracy: 81%
- Precision: 82.2%
- n_estim = 500
- max_feature= .25
- balanced weight
Kneighbor Classifier
- Accuracy: 77.6%
- Precision: 81.7%
- n = 5
Neural Network
- Accuracy: 65%
- Precision: 80%
- hidden layers(10, 10,5,3)
- balanced weight
Logistic Regression
- Accuracy: 79.9%
- Precision: 80.8%
- C= 100
- L2 Regularization
Model Conclusion
-
Random Forest seemed to have the best overall performance
-
The general trend looks as though if precision increases, accuracy decreases and vice versa
-
There are no determination on which model performs the best as the model used would depend on the goal achieved depending on higher accuracy or precision
The Trade Offs
It is important to understand how expensive False Negatives are compared with False Positives. For lending club, False Negatives are much more expensive as the amount of loss that 1 False Negative could bring far outweighs its potential return. Therefore higher precision models might be more favored in return for lower accuracy.
Some models were able to reach 90% precision but sacrificed accuracy as it would drop to 70%.
In order to build a more sustainable business model, it is better to generate less revenue for future growth and stability by increasing precision of the loans and increasing the potential ROI of the loans
Future Works
Work on NLP for loan description or social media if information is available to increase prediction accuracy
Run the models for each sub-grade to increase accuracy and precision for the loans
About Authors
DongHwi Kim
View all posts by DongHwi Kim >
Lily Kuo
View all posts by Lily Kuo >
Huimin Ou
View all posts by Huimin Ou >