The Tale of LendingClub
LendingClub, a financial services company headquartered in San Francisco, was the first Peer-to-Peer (P2P) lender that registered its offerings as securities with the SEC and offered loan trading on a secondary market. It was once the world’s largest P2P lending platform. However in 2020, it acquired Radius Bank to transform into a neobank, and it even announced the shutdown of its renowned P2P platform.
What happened to LendingClub? Why would they make the pivot?
Taking on the role of an external analyst, I aim to find an answer by analyzing LendingClub’s historical loans through the years. My exploration and analysis focus on two areas:
(1) Whether LendingClub is offering attractive returns for investors, and whether advanced techniques could help investors increase returns
(2) With additional data related to macroeconomic conditions, how the default probability of borrowers would be affected during times of uncertainty.
The dataset used for analysis contains all accepted loans by LendingClub issued from 2007 up to the end of 2018. There were over 2 million loans with 151 attributes, which covers all the loan features, like term, interest rate, amount and purpose, as well as a wide range of borrower information like income and credit history.
I also introduced macroeconomic indicators to examine whether market conditions have any impact on loan performance. CPI growth rate and real GDP growth rate data were collected from OECD and statewide unemployment rate data from the Department of Labor.
Exploratory Data Analysis
Loan Volume and Loan Count
Loan volume and count both had rapid growth every year. In 2014, the year of its IPO, it had over $3 billion of loans issued. Also, note that 3-yr loans had a higher volume than 5-year loans every year.
Visualizing loan volumes geographically, the state with the most loans issued was California, followed by New York, Texas and Florida. As the population of New York is only ranked 4th among these states, LendingClub was more popular in New York than in Texas and Florida.
According to LendingClub, the loan grades take into account not only credit score but also a combination of several indicators of credit risk from the credit report and loan application. The grades range from A1 to G5.
Grade B and grade C are the most prominent in terms of both size and counts, where the two grades account for nearly 60% of the entire portfolio.
Interestingly, lower-graded loans in general have larger loan sizes, as shown in the violin plot.
LendingClub’s interest rates take into account credit risk and market conditions, it is calculated by adding Adjustment for Risk & Volatility to the Lending Club Base Rate. Interest rates correspond to the loan grades. Interest rates had many fluctuations but the majority of them were in overall upward trends from 2007 to 2018.
The exception here is Grade A loans, where their interest rate trends were decreasing, more or less in line with the trend for US interest rates in the last decade.
Debt consolidation and credit card are the most common purposes in LendingClub loans, accounting for about 83% of the entire portfolio by volume. As for the average loan size by purpose, the largest is for starting small businesses and the smallest need comes from vacations.
For loan applicants, LendingClub typically makes approval decisions based on their FICO score, debt-to-income and credit history. For those who were approved, their DTI ratios are almost evenly distributed with a slight right skew, with the vast majority of those accepted having DTI of less than 40%. As for their FICO scores, the middle 50% (1st-3rd quartiles) of approved applicants range from 672 to 717.
Loan return is the most important measure to investors. Because there are loans with various conditions and different actual payout schedules, it is more useful to use Return-on-Investment (“ROI”) for a loan’s entire time horizon to make the comparison.
Looking at ROI by grade and term, the analysis shows that grade B loans have the highest ROI, which is 5.57% for 3-yr loans, and 3.91% for 5-yr loans.
Secondly, returns for 3-yr loans in 6 of the 7 grades all outperform the 5-yr loans, with the exception of grade G. This is counterintuitive from the overall return perspective, but it could be the consequence of having a higher possibility of default on 5-yr loans.
In addition, for grade G loans, although they offer higher interest rates on paper, historical data indicate that the high default rate led to larger losses for investors. In fact, LendingClub has discontinued the issuance of grade F and G notes since the end of 2017.
For modeling purposes, the following major data processing procedures were applied to the loan dataset,
- Subset complete loans of which loan status is fully paid or default
- Combine income and credit history features for joint applicants by taking averages, taking minimum/maximum, taking summation of variables, etc. Specific treatment of variables depends on the nature of each feature
- Conduct Missing value imputations
- Label-encode ordinal categorical variables to numerical values
The feature selection process utilized the following criteria and tests,
- Preliminary selection based on the combination of my industry experience and explanatory analysis;
- VIF test
- Correlation matrix analysis.
Finally, 21 features were selected and used for modeling. They were in three categories,
- Loan feature: Grade, Term, interest rate, loan amount (funded), purpose
- Borrower info: Income, DTI, employment status, credit history, historical records
- Macroeconomic indicators: CPI growth rate, real GDP growth rate and unemployment rate
Machine Learning Modeling and Analysis
For interpretation purposes and model performance, the logistic regression model was fitted as the benchmark model for further analysis. The target is the loan status of individual loans.
As the original data is highly imbalanced with 80.02% to 19.98% for fully paid and default respectively, a balanced dataset was sampled for modeling.
The model was tuned using grid search method and the best model has an accuracy of 0.6465 on training data and 0.6374 on testing data. The AUC score is 0.6463 and the confusion matrix is as follows,
My model serves two business goals,
- To help investors find good borrowers and Notes. This means to predict as accurately as possible on the fully paid loans (True). I hence checked the Negative predictive Value, which is 0.6325.
- To examine factors leading to loan default and to find out what makes the P2P platform less attractive to investors
Analysis 1: How ML techniques can help investors
For investors on a P2P lending platform, a common investment strategy is to simply invest a fixed dollar amount for each available security. This is especially true for LendingClub where you could put in as little as $25 for each Note, encouraging investors to diversify. I mimicked this strategy by randomly picking a sample of 8,000 loans that have the same loan status distribution as the entire LendingClub portfolio.
If an investor purchased all 8,000 of them, the overall return on investment would be 3.39%. But by utilizing my model and only investing in loans that my model predicted to be fully paid, the return would be 6.10%.
The same test was also run on several different test datasets for selective issuance years. Using LendingClub with my ML model would provide a good alternative to stock investors in bearish years. For example, if an investor had invested in S&P 500 in 2007 before the financial crisis, the return would be -23% over 3 years, but investing in LendingClub notes would get 2.44% and 11.49% with the help of ML techniques over the same period.
Analysis 2: What major factors influence loan default
For my model result on features, notable positive coefficients below lead to a higher odds ratio and thus higher default probability,
- Grade: the largest with a coefficient of 0.4. A larger grade number (e.g. Grade C has a higher number than B) leads to a higher default probability
- Loan term: For default probability, a 5-yr loan is higher than a 3-yr loan
- Number of credit inquiries for the last 6 months: Having many credit inquiries in the recent 6 months can be interpreted as an attempt to substantially expand available credit, and is inherently associated with a higher default possibility
As for the features with negative coefficients,
- GDP growth rate: Most notable negative A lower GDP growth rate leads to a higher probability of default
- Unemployment rate: This is a counterintuitive result. According to the model, a lower unemployment rate in a certain state would equate to a higher probability of default in that state
- Employment length provided: A significant portion of applicants have provided no “employment length”, and I found out that a loan borrower who failed to provide an employment length would be more likely to default
- Purpose - credit card: Different loan purposes have inherently different risk levels. A credit card refinancing/consolidation loan has a much lower likelihood of default compared with starting a small business
To return to my original question, why is LendingClub pivoting? The first reason, according to the analysis, is that LendingClub lost its appeal to retail investors. For a retail investor who buys Notes at random or buys all Notes available, with no Machine Learning or advanced quantitative techniques, it would be difficult to achieve a return comparable to the S&P 500, especially when in the last decade the stock market was having high returns every single year.
In fact, in later years, LendingClub was attracting mostly institutional investors like hedge funds, commercial banks and investment banks. They have advanced models and expertise to value and choose loans, and purchasing entire LendingClub loan portfolios is an alternative investment for them.
While losing appeals to retail investors, LendingClub also wanted to lower its funding cost, and the way to achieve this is to become a bank. In 2020, they acquired Radius bank, an online bank. They then started offering online loan products targeting a similar audience, but with lower funding costs.
Another reason was the COVID-19 outbreak. The pandemic brought panic and uncertainty with economic outlook when it wrecked the US economy and stock market in early 2020. With a gloomy expectation during the initial phase of COVID, the demand for personal loans went down. Also, based on the analysis of the machine learning model, borrowers are more likely to default with the economic downturn expectations. Moreover, with people being hit with difficulties, more borrowers with bad grades would flock to the platform, and this would not be LendingClub’s ideal situation. All of these factors may have accelerated the closure of LendingClub’s P2P lending platform in 2020.
The logistic model is a good model choice for business analysis and interpretation. To increase the prediction accuracy of default probability, tree-based models can be introduced to generate more accurate selections on loans.