Lending Club: Optimizing Lending Club Portfolios
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
LinkedIn | GitHub | Email | Data | Web App
The advent of electronic peer to peer (P2P) lending, also known as marketplace lending, was invented in 2005 with the launch of the London-based, Zopa. In 2006, Lending Club was launched in San Francisco, becoming one of the first peer-to-peer 'marketplaces' in the United States. The overall market for these types of loans was still relatively small at the time, but for Lending Club and the industry as a whole that would soon change.
From a borrower's perspective, Lending Club allows individuals to lend to other individuals without the red tape, overhead and restrictions banks impose. From an investor's perspective, you are excluded from earning loan-level interest rates altogether. For example, a working professional who happened to get his or her self into credit card debt, paying near 30% annual interest rate can go to a P2P platform like Lending Club to receive a loan at a more favorable interest rate from an investor or peer. The investor, in theory gets a much better rate of return than he or she would at their local bank. Currently the average Savings account returns 0.9% in the United States, while the average annual interest rate on LendingClub.com is 12.3%.
This business model & additional flexibility to traditional lending practices proved to be a novel concept, as we see the Lending Club experienced exponential growth between 2007 & 2016. Their largest period of growth occurred between 2012 & 2014, when they received investments from Google, John Mack (former Morgan Stanley CEO), surpassed $1 Billion in total loans & filed for their IPO.
How it Works
Lending Club aims to use big data & machine learning algorithms to make credit decisions that were traditionally made by loan officers & their mangers at Banks. Without the overhead of loan officers & physical locations, lower costs and in turn, better outcomes could be offered to both borrowers and lenders. This ignited a new category within the FinTech Sector, with much promise for its role in the future of lending. But one key question remained to be answered: Could Machine Learning truly replace loan officers?
It appeared that FinTech companies like Lending Club could one day surpass traditional loans, should they continue their parabolic ascent. However, the first signs of turbulence came in 2016 after an increase in interest rates made other investment vehicles more attractive, but more importantly, the rising rates of borrower default began to cut into returns. As a result, Lending Club, which was muddling through a scandal and a change in leadership, entered a three-year slide in share prices. The P2P lending market continues to grow today, but there is great importance in accuracy of the machine learning models & the technology behind Lending club. In the end, they are still accountable to investors needs: good rates of return and minimizing risk.
We found the loans average around $8,231 and are typically used to refinance expensive credit card debt. The average interest rate over all loans is 12.3%, but after factoring in early repayment, loan defaults, and platform fees, investors average a 5% annual return. For lenders owning more than one hundred notes, returns can be fairly consistent (low volatility), but for those investing in just a few loans, the returns can vary wildly, ranging from above 10% to subzero (negative) returns. Given the wild variance in the outcomes from the lender's perspective, we began to dig deeper.
The most desirable outcome for a lender is that the loans are consistently paid on time rather than late or early for the entire lifetime of the loan, and terminating in repayment at the due date. Approximately 30% of the 36 month loans, and 15% of the 60 month loans follow this pattern, with most of the other loans being repaid early to one extent or another, and 5% or fewer incurring costly late fees. 10-15% of the 36 month loans end up defaulting at some point, and 25% of the 60 month loans. Most of the 60 month loan defaults occur prior to the three-year mark.
The loans which default are well-distributed across the loan sizes, but show the expected dependence on the loan grade, which is Lending Club's measure of the borrower reliability. The grade is determined in part by the FICO score, as shown below.
The interest rates are set based on Lending Club's evaluation of the risk as expressed by the grade. The interest rates for high-risk loans have gradually increased over time as Lending Club increasingly leveraged them for a better return.
The data for the accepted loan applications contained over 2 million entries spanning from June 2007 to December 2018, and had 110 features. Since our objective was to develop a predictive model for determining bad loans from the loan application, we had to eliminate all of the features containing data from the loans themselves, except for the end status, which was our dependent value. After also eliminating high-missingness features (any feature missing 70% or more of the values), this left us with 40 usable features.
We tested seven different classifiers, and dropped those with poor performance. The final model was a logistic regression stacking of XGB, random forest, and gradient boost classifiers. Accuracy was only 67% Feature importance showed a significantly different reliance on the features by XGB, which gave considerable weight to many features. The Others had a strong dependence on only three features, including loan grade, duration, and either interest rate or installment size.
When the threshold is near zero, the returns are the average returns computed by the data set: 5.79% for the 36 month loans, and 5.88% for the 60 month loans. As the threshold is increased to apply greater selectivity using the model, the most obviously bad loans are left out of the portfolio and the returns increase rapidly. Both classes of loan peak near 60-70%, after which point the increasing selectivity begins to eliminate too many profitable loans, and the returns suffer. At peak, the 36 and 60 month loans have their returns boosted by about 3 and 4%, respectively.
We showed that machine learning can be successfully applied to reduce risk, and to increase returns by treating the bad loans as a classification problem. However, there are more optimization measures that could be applied which require regression. Early repayment of loans is a phenomenon with a continuous dependent variable, where the highest profit requires the latest possible payment within the deadline. Defaulted loans too, are variable in their impact, with some causing nearly 100% loss of the invested capital, and with others still bringing in a nominal profit. In order to build the perfect portfolio of Lending Club loans, we need to apply regression modeling to find the most lucrative possible portfolio.