Algorithm investing – Part 1: What is a secondary credit model?

Part 1: What is a secondary credit model?
Part 2: Investing through LendingRobot

For years investors have boosted their overall return by filtering the platforms for higher performing loans. This worked because Lending Club and Prosper’s earliest loan grade assignments were not completely accurate. Investors, informed by the open historical data, could take advantage of these inaccuracies to boost their overall return. The high water mark of filtering was actually in 2011 when Prosper had not considered the increased creditworthiness of their repeat borrowers, and investors were increasing their ROI by over 4% (Lend Academy) simply by focusing on repeat borrower loans.

85% of the filters I used in 2012 are very weak or unusable.

Fast forward to early 2015, both Lending Club and Prosper have significantly improved their pricing, meaning, they have become much better at assigning the correct interest rates to the different borrowers who come to them for a loan. So while a number of filters still seem to work, I would say 85% of those I used in 2012 are very weak or unusable.

This is ok. In fact, I would even say the death of simple filtering is good news overall since it indicates a larger maturation of this industry. Each year both Lending Club and Prosper get better at what they do, which is giving us access to the stable and rewarding investment of consumer credit. Having accurate interest rate assignments is part of what it means for these companies to become more stable and excellent.

Secondary Credit Models: The Future of Filtering

Does this mean filtering the platforms for higher performing loans is gone forever? By no means. Some of the biggest investors in online loans (banks, hedge funds, etc.) do indeed filter the available loans. But they don’t use highly simplistic filters like Inquiries=0. Instead, they approach the available loans with an outside credit model.

What is a credit model?

Let’s pause a moment to define the term ‘credit model’. Lending Club and Prosper deny 90% of the borrowers who come to them for a loan because of poor credit history, low income, or because they carry too much debt. The thing is, a company like Prosper does not look at this myriad of factors with human eyes, but with powerful and complex mathematical algorithms, basically a computer program that can instantly determine how likely each borrower is to repay their loan.

Every credit report contains hundreds of data points.

Remember, every borrower who comes to Lending Club or Prosper for a loan gives them their credit history, and this report contains hundreds and hundreds of data points (IE: all the fields in this Prosper back tester). Did this borrower go bankrupt in 2009? That’s in their credit report. Are they still carrying $49 in credit card debt from buying jeans last month? That’s in there too.

In short, Prosper and Lending Club’s highly accurate credit model is responsible for peer to peer lending’s consistency and low default-rate. Their credit model is what makes this a good investment.

What is a secondary credit model?

Some of the biggest investors at Lending Club and Prosper filter the available loans using a secondary credit model/algorithm. Their model allows them to instantly calculate the likely ROI of each available loan, identifying where there is an opportunity to earn above average returns. Then they simply invest in those loans.

Only a few of these models are available to the public. One of them is LendGuardian by Bryce Mason of P2P-Picks. LendGuardian works on both Lending Club and Prosper, statistically identifying the loans that are better investments.

For example, in the above graphic Bryce has created a scatter plot of loans on Lending Club’s platform from January 2014. The vertical axis is ROI, meaning, the mathematical return that each loan is likely to give an investor, calculated by the LendGuardian statistical algorithm. The horizontal axis is the default rate, with the algorithm again having calculated how likely each loan is to default on its payments.

The dark lines are buy-lines. Line A would be the buy-line for an investor who is trying to minimize their risk (while still earning a great return). Someone in retirement might choose to invest above Line A. In contrast, the Line B buy-line is what I personally would use. Line B solely seeks to maximize returns (ignoring risk altogether). Most investors would probably want to tilt that line a bit for a safer investment.

[Bryce’s model is widely celebrated by the public, and has been used by major investors like Titan Bank.]

How is investing through a credit model different from investing with a simple filter?

This is a great question. The short answer is, investing with a credit model is more holistic. Instead of bluntly approaching the available loans with a single credit variable (Inquiries=0), a credit model filters the platforms using all of the borrower variables at the same time.

A great example of this has been demonstrated by the other credit model that is available to the public: LendingRobot. As seen in this blog post, Emmanuel Marot of LendingRobot has calculated the way each loan attribute adjusts the expected default rate.

A variable with a negative value means that variable decreases the estimated default rate. A positive number increases the default rate. See the two red boxes? Credit card is the less likely loan purpose to default; loans purposed for Small business are more likely to default. Of course, the point is not to isolate a single variable like Purpose: Credit card. That would be simple filtering. The point is to show how Emmanuel’s algorithm/credit-model holistically invests on these platforms by considering all of the borrower credit variables at the same time.

[The Expected Return calculation at LendingRobot is at the center of their great investment tool, recently garnering them $3 million in investor funding]

Credit Model Investing vs. Simple Filtering

Investing through such beautiful mathematics may seem more elegant than simple filtering, but how do they actually stack up?

#1. Credit model investing is more consistent

Credit models are more consistent over time. Filters come and go. Remember how I said 85% of the filters I used in 2012 are weak or dead? If I had ignored these changes, my returns would have steeply decreased over the past two years. Filtering takes work (though I enjoy it). In contrast, if I had been investing with P2P-Picks since its launch, I would be earning an above average return for the past two years without having to adjust anything on my end whatsoever.

#2: Credit model investing’s nuance allows larger investments

In my own experience, filtering works best when the filters are somewhat tight. As a simple filter becomes wider and wider, it seems to lose its power, particularly in today’s environment where Lending Club and Prosper have become far more accurate in their loan grade assignments (IE: Simply filtering for higher incomes at Prosper may no longer work).

In contrast, credit models can be widened farther and farther (IE: moving the buy-line in Bryce’s above graphic), allowing much larger sums of cash to be put to work than simple filtering allows. Independent investors like myself who invest less than $1,000 per week may use simple filtering just fine. But banks like Titan and savvy hedge funds with million dollar accounts need something more robust.

#3: Credit model investing is much safer

The loan grades at Lending Club and Prosper are largely outside of our control. We have to trust them. And if they wanted to, these companies could degrade the quality of their loans without the consent of average investors. Yet secondary credit models would be instantly aware of such a change. If bad loans were on the platform, my simple filters would still invest in them. Those with secondary credit models would not invest at all; their cash would be safely kept off the platforms.

#4: Simple filtering is easier to understand

This article barely touched the raw mechanics of credit modeling. The reality is, the math behind these algorithms is heavy with terminology and not understandable by the average investor. That said, simple filtering is straightfoward and easy to grasp. IE: Inquiries=0 has historically given investors a higher than average return, so we can do that ourselves by moving a slider on Lending Club’s website. Simple.

#5: Simple filtering is more independent

Most of us are not rockstar statisticians. If we choose to invest through LendingRobot or P2P-Picks, we place a lot of trust in these companies and the math they practice. Indeed, these credit models really are black boxes, even for Bryce and Emmanuel.

But simple filtering keeps our loan selection almost purely in our own hands. Yes, many of us depend on Michael Philips to have accurate math behind NSR, but that’s about it, so the more independent investors will probably feel more comfortable with simple filtering.

Why I’m Switching to an Algorithm

Simple filters have worked for years – pretty much since peer lending began. Many feel very comfortable with it. Self-directed investors have increased their ROI (see my returns) by taking advantage of their ability to filter the list of available loans. In contrast, investing with a credit model is still novel to the average investor.

But this changing. As simple filtering becomes obsolete, using a credit model will arise as the best way to boost our return (see Bryce’s quarterly returns). I will likely switch my own filtering over to a credit model by late-2015/early-2016.

Conclusion: Celebrating the Open Data of P2P Lending

In closing, I want to take a step back and celebrate this investment and how it makes credit model investing possible, particularly at such a young stage in its growth. Since Lending Club and Prosper have published their loan data for anybody to download, people like Bryce or Emmanuel are able to offer these models to the general public.

In my opinion, this is just another example of the inherent quality in p2p lending. Not only does the open data prove this investment to be rewarding and stable, but everyday Americans are able to click a few boxes on 3rd-party websites to utilize incredibly powerful algorithms that increase their overall return.

[image credit: ‘Mathematics‘ by Tom Brown CC-BY 2.0]