Free Updates

Exclusive content to your inbox for FREE!

How I am Investing in Lending Club and Prosper in 2012

by Peter Renton on February 9, 2012

One of the things I love about p2p lending is the transparency. By this I mean that anyone can download the entire loan history of Lending Club and Prosper and analyze the data for themselves. I am trying to bring a level of transparency to my own operations on this blog by giving you an inside look at my investments.

Last month I provided a snapshot of all my p2p lending accounts and today I will continue along on that journey by revealing exactly how I am investing in Lending Club and Prosper today. I first detailed my investment criteria nine months ago in a post that described how I was investing with Lending Club and Prosper back then. I gave you two strategies each for both companies and today I am going to expand on that.

No More Conservative Lending Strategies

The biggest change in my investing in the last nine months is that I have ditched the conservative lending strategy at Lending Club. In my main Lending Club account I had been focusing on B- and C-grade loans for quite some time. But I decided that was simply leaving money on the table so late last year I switched course and decided to focus purely on loans grades of D and below at Lending Club on all my accounts (except for my Lending Club PRIME account).

Now, this created something of challenge. As I detailed in my last post I want to invest in multiple p2p lending accounts without investing in the same note twice. So, after spending way too much time on Lendstats exploring hundreds of combinations of selection criteria I came up with these sets of filters that provide no duplication of notes. I have provided a link below each filter to the Lendstats page that shows the returns one might expect when running these filters. If you don’t know what some of these fields mean you should learn more about credit reports (just google “understand credit report” and you will find plenty of articles).

Lending Club Filter 1 – High Income

Loan Grade: D, E, F, G
Inquiries = 0
DTI% <= 23%
Open credit line >= 8
Public records = 0
Monthly income >= $7,500
Loan purpose: All except other, small business and vacation
States – exclude CA
Link to Lending Club Filter 1 on Lendstats

Lending Club Filter 2 – Medium Income

Loan Grade: D, E, F, G
Inquiries = 0
DTI% <= 25%
Open credit line >= 8
2 Yr Deliquencies = 0
Public records = 0
Monthly income >= $3,000 and < $7,500
Loan purpose: All except other, small business and vacation
States – exclude CA, GA and TX
Link to Lending Club Filter 2 on Lendstats

Lending Club Filter 3 – Inquiries 1+

Loan Grade: E, F, G
Inquiries >= 1
2 Yr Deliquencies = 0
Public records = 0
Monthly income >= $7,000
Loan purpose: All except small business
States – exclude CA
Link to Lending Club Filter 3 on Lendstats

You can see that the main difference between Filter 1 and Filter 2 is the stated monthly income. I use that field to ensure that there is no overlap between loans when I am investing in multiple accounts. You will also notice that both filter 1 and filter 2 use inquiries = 0 as a criteria so this opens up the door to use Inquiries of one or more for Filter 3. Because all three filters don’t invest in loans originated in California I could easily setup a fourth unique filter for loans just issued in that state. I haven’t done this mainly because there are not enough loans that meet my criteria.

One point I should make is that if you use the Lending Club website to invest then you will not be able to use these filters as is. The filtering capabilities on their website are not flexible enough to allow for this kind of precision and some fields such as monthly income are not even available. So what I do is download the spreadsheet of all available loans from the Browse Notes page – there is a small Download All link in the bottom right of the screen. Then I can do the filtering in Excel and invest from there.

Prosper Filter 1 – Previous Borrower 0-1 Inquiries

Loan Grade D, E, HR
Payments on previous loans >= 12
Number of late payments  <= 9%
Allow credit score drop up to 100 points
Inquiries <= 1
Current delinquencies <= 1
Link to Prosper filter 1 on Lendstats

Prosper Filter 2 – Previous Borrower 2-5 Inquiries

Loan Grade D, E, HR
Payments on previous loans >= 10
Number of late payments  <= 10%
Allow credit score drop up to 100 points
Inquiries >= 2 and <= 5
Current delinquencies <= 1
Link to Prosper filter 2 on Lendstats

Prosper Filter 3 – New Borrower 0 Inquiries

Loan Grade D, E, HR
Payments on previous loans = 0
Inquiries = 0
Current delinquencies = 0
Open credit lines >= 10
Debt-to-income ratio <= 75%
Link to Prosper filter 3 on Lendstats

Prosper Filter 4 – New Borrower 1-2 Inquiries

Loan Grade D, E, HR
Payments on previous loans = 0
Inquiries >= 1 and <=2
Current delinquencies = 0
Delinquencies in last 7 years = 0
Bankcard utilization <= 95%
Open credit lines >= 10
Debt-to-income ratio <= 75%
Public records last 10 years = 0
Employment status: exclude Unemployed, Not Available
Link to Prosper filter 4 on Lendstats

The bulk of my new investments on Prosper go towards repeat borrowers. I have found repeat borrowers to be an excellent group of borrowers and you can see by clicking on the Lendstats link with each filter that they provide excellent returns.

Long time readers will know my love of the Number of Inquiries filter so you might be surprised by Filter 2 where I go with number of inquiries between 2 and 5 (I have long maintained that inquiries = 0 is one the best filters you can have). But I let the Lendstats ROI numbers be my guide here. And even though a previous borrower has two or more inquiries on their credit report with the additional filters in place here you can still generate an excellent ROI.

Now, I don’t want to be one dimensional and ignore new borrowers, so filters 3 and 4 provide a way to invest in new borrowers that is also likely to produce a good return. Again I am using number of inquiries as the way to separate the note selections to avoid duplication.

So, there you have it. These are the criteria I am using to invest today. It makes investing with multiple accounts a breeze or you can just as easily use these criteria on one account. Feel free to use these filters yourself if you like. Or you can always critique them and provide your own suggestions in the comments.

Like what you read?
Then please join over 1,000 people who receive exclusive weekly p2p lending tips, and get a FREE COPY of my ebook, Understanding Peer to Peer Lending. Just enter your name and email below:

{ 69 comments… read them below or add one }

Dan B February 9, 2012 at 8:38 am

Again, on behalf of my fellow Californians, I must protest on our seemingly blanket exclusion from consideration. We are a great state with incredibly responsible people. You can always rest easy with the knowledge that your investment is completely safe with us. No Californian would ever consider walking away from a debt except as a second or third resort. :)

Reply

Danny S February 9, 2012 at 8:43 am

Very interesting loan filters for LC, particularly that you are focusing on lower grade loans now. Have you determined that the higher interest rates those notes pay more than offsets the increase in defaults you experience?

I’ve also been a B and C note investor (and a few A and D here and there). My biggest concern about investing in notes that are charging 20%+ interest (E, F, G) is that these borrowers are less likely to be doing it for debt payoffs, vs things like home repairs, medical or wedding expenses, small business startups, etc… where I think they are riskier uses of the funds.

Reply

Dan B February 9, 2012 at 9:17 am

As myself & others have stated here 127 times previously, the truth is that we really have no idea what the borrower ultimately does with the borrowed money. Personally, I assume that regardless of what they “say”, they will in fact spend it, rather than pay off debt. It frankly makes no difference to me as I’m not here on some misguided social do-gooder mission. I’m just here to get a high return.

I’m no longer a Prosper investor (for totally unrelated reasons) but in a sort of perverse way, those who are HEAVILY concentrating on “repeat” Prosper borrowers (like Peter) also recognize that the ultimate best borrower is one who will keep “repeat” borrowing & remain perpetually in debt while making his high interest monthly payments on time. I have no problem with that reality either. Just being honest about the dynamics & reality of the situation.

Reply

Charlie H February 9, 2012 at 12:16 pm

It should be noted that most income is just stated income and not verified income.

What the loan is used for, as Dan B points out, is 100% unverifiable. I don’t use it for filters other then to exclude people.

Reply

Peter Renton February 9, 2012 at 1:09 pm

@Dan, I would like to include some CA borrowers in my Lending Club investments but I simply cannot make the numbers work with a large enough pool of loans. As more history build up that may change but you can take solace in the fact that I do not discriminate against CA on Prosper.

And yes, I do realize that repeat borrowers on Prosper paying 20%+ are rarely going to be paying down debt. I have somewhat struggled with this philosophically but I have decided that these people are still benefiting from being able to access money even at these high rates. If they were not borrowing here they would probably be racking up more credit card debt or even going the pay day loan route. A fully amortized fixed term 24% loan is a better deal for them in most cases.

@Danny, Yes, I have determined that the lower grade loans provide the best chance at a high return. And that is my focus. I don’t mind getting some defaults on 20% loans if, at the end of the day they produce a 12% return. Getting that kind of return by investing in B and C-rated loans is extremely difficult.

@Charlie, I am fully aware that stated income is mostly not verified. I am not relying on verification for my decision. Nor am I concerned about loan purpose. I let the loan history data be my guide. For my Lending Club Filter 1 I know that many of these people may not be earning more than $7,500 but the loan history provides return information on those people who say they earn that much. That is good enough for me.

Reply

Danny S February 9, 2012 at 2:50 pm

Peter I find some of your insights to be quite fascinating. I dont know if I can get my head aligned with G graded notes, but I am going to take a closer look a D, E, F notes.

My big concerns for those notes are monthly income (even if not verified) vs monthly payment… I try to keep to a ratio of no more than 15% of a person’s income going towards their LC loan. That has seemingly worked quite well in keeping defaults low for me.

I also dont like having borrowers which show any delinquencies within the last 4 years, so that also affects the # of loans in the pool available to me. But will see how many of these lower grade loans are available in my comfort zone.

Reply

Larry V February 9, 2012 at 4:12 pm

Peter,

My continued thanks for you providing a great place to learn about and discuss P2P. Thanks for being open about your investment philosophy.

Reply

Bryce M February 9, 2012 at 8:22 pm

A few thoughts:

I see many people create filter strategies that attempt to get a handle on risk. But I have seen few people use risk in combination with reward to choose their loans. This is my approach. I predict a charge-off probability for each loan in the database (based on historical predictors), and then compare it to the interest rate to calculate an expected value for the loan. I rank by expected value, and choose the top loans in which to invest.

The mutually exclusive filters is a poor idea in my opinion because it creates an artificial partition on the data. You could literally be choosing “the cream of the crap!” It makes much more sense to look at all loans simultaneously and choose the best ones.

Those who don’t pay attention to stated use are leaving valuable information on the table. Stated reason is a statistically meaningful predictor of charge-off. It doesn’t matter whether people use it for the intended purpose; on average it is a valid signal.

Reply

Peter Renton February 9, 2012 at 11:30 pm

@Danny S, It does take a bit of a leap of faith to take a look at the higher risk loans. I was investing for over a year before I even considered a higher risk loan. Now, they are my complete focus because, if done well, they provide the best chance for a high return.

@Larry V, You are welcome.

@Bryce, What you describe is an interesting concept and reminds me a bit about what SmartPeerLending.com is doing. The way I look at it is that Lendstats in essence provides the same thing. Because they take into account defaults and late notes you are getting a window into a charge-off probability.

Looking at all loans simultaneously is the best solution for most people who have just one account. But if you have multiple accounts you need an investing strategy that partitions the loans unless you want to duplicate your portfolio across different accounts.

Reply

Frankie C February 10, 2012 at 1:12 am

Peter,

Thanks a lot for sharing all this valuable info. I’d like to touch on a few topics:

What’s the reasoning behind DTI% = 8 ? Did you just trial and error on lendtats until you found the magic numbers?

And what about avoiding CA? Did you try excluding each state at a time until you found the worst offender? Any idea why?

I’m very surprised to see you allow 2+ inquiries applicants. I understand your desire to come up with different criteria because of your multiple accounts. I’m not sure giving up on what has been a great predictor makes sense for those of us with a single account.

Finally, I’m surprised you don’t filter more selectively on loan purpose. It seems to be an pretty good predictor also. Do you disagree?

Thanks,
FC

Reply

Bryce M. February 10, 2012 at 2:25 am

Partitioning the loans will never yield a profit maximizing solution if you insist on nonzero investments in all portfolios. I think I could prove it, but it seems intuitive to me anyway. Maybe at best it will match the profit maximizing allocation, but seems unlikely.

If one filter is superior to another, then you would always be better off by investing more in that space with the money from the weaker filters portfolio. My 2cp.

Reply

Moe February 10, 2012 at 2:11 pm

How about the P2P companies start offering balance transfers, having the the borrowed money go directly to the credit card companies owed. This would be noted on the note listing, and I’m sure these notes would get funded much quicker. Now let’s see who will jump on this first…

Reply

Chris February 10, 2012 at 2:37 pm

My two cents for what it’s worth:
1) Risk as a measure to whether one should choose A-G rated notes should be compared against one’s entire portfolio, which hopefully does not include just lending club or prosper products. A healthy balance of other assets (bonds, stocks, CDs, ETFs, Real Estate, whatever) will by its very nature permit you to entertain higher risk (E-G) notes if you are wanting to dabble with that asset class. I happen to agree with Peter and the other comments above that it is a class worth taking seriously. Consequently, if you find yourself curious or interested in the E-F notes, simply offset it with something else (outside of member dependent notes altogether) and you should be fine to at least experiment.
2) Speaking of diversification, I have often wondered if a P2P company will one day provide us with a platform where we can invest in credit card debt directly (funding the actual credit card’s “credit”) instead of investing indirectly in credit card debt as we do now (by providing the credit for prior credit card debt consolidation). Considering how large a market credit card use has become, one could only wonder the investment potential awaiting the first P2P platform that could service those instruments directly.

Food for thought…

Reply

Chris February 10, 2012 at 2:38 pm

By the way Moe, you have an EXCELLENT idea.

Reply

James February 12, 2012 at 1:20 pm

Peter,

Thanks for your ideas. I still don’t understand the rationale for excluding certain states?? And how did you choose which to eliminate?

Also the criteria you list don’t fit the filters on Lending Club. For instance you use 23% DTI in your first example and Lending Club allows selection only in 5% increments. Also you want a minimum stated income and there is no such criteria on LC? Or do you use lendstats to find the loans then link back to lending club to purchase them?

Thanks again!

Reply

Bryce M February 12, 2012 at 7:37 pm

The rationale for giving a black mark for states is simple. Certain ones default at higher rates than others. Go download the data and prove it to yourself.

Reply

Peter Renton February 13, 2012 at 3:12 am

@Frankie, I have spent dozens of hours running several hundred queries on Lendstats and I have let the results from this process be my guide. Keep in mind that you need a decent pool of loans to have any query be statistically significant – I always like to see each query generate at least 500 loans before I consider it a good predictor of future results. Then I will re-run the query over several months to see if it continues to provide a consistent ROI. You are right that loan purpose can be a decent predictor of return and I know many people who use that as an important basis for investing decisions. There are plenty of other ways to generate good returns – these are just the ones that I have chosen.

@Bryce, If it was absolute that a certain criteria would produce the best results then I would agree with you that it would always be better to just invest in the best loans. But the fact is we are working with a moving target. The loans that you have isolated as the best loans may end up proving to not be the best when all loans have reached maturity. We need a loan database many times larger than what we have and underwriting standards that don’t change before we can be completely confident in our choices. This is another reason to choose different criteria – some of them will continue to perform well while others will likely underperform expectations.

@Moe, That is an interesting idea and I can see something like this becoming a reality in the not too distant future.

@Chris, Very good point about considering p2p lending’s role in an overall portfolio. That is really a key point that all investors should take into account. For most investors I think it pays to be more aggressive with their p2p lending holdings, because most people hold a mix stocks, bonds and cash. For the reasons you point out this should make investors feel more comfortable with higher risk notes. In reality, though, most people treat p2p lending in isolation and focus on the higher grade notes.

I think one day, probably several years from now, Lending Club and Prosper will get a banking license and issue credit cards directly. Then finally we can get the traditional banks out of the equation. But they would need to be several orders of magnitude bigger before considering such a move.

@James, You have pointed out the main reason I do not invest with the Lending Club platform. I download the in-funding notes to an Excel file and do the filtering there before going back and investing in the loans on Lending Club. On Prosper this isn’t necessary because they have much more flexible investing criteria.

As for the states, I only consider the largest states for exclusion. Even though states like South Dakota has performed poorly at Lending Club I don’t bother excluding them because so few loans are issued there.

Reply

Bryce M. February 13, 2012 at 10:08 am

I don’t think your argument that you want multiple criteria is solid, Peter. For example, if what you say is true then you wouldn’t be neglecting CA loans because things might have changed since the loans used to make that negative relationship finished. You would be using CA loans because they might be stellar performers in the future. But, you do not, presumably because you also hold the belief that about all you can do is use the past to predict the future.

. The best we can do is to choose criteria that have a string theoretical basis and empirical evidence and hope they continue to perform in the future.

Reply

Peter Renton February 13, 2012 at 9:42 pm

@Bryce, I agree that is the best we can do and if I had just one account then I would be doing exactly what you suggest. There are literally millions of combinations of loan filters that we could use and I am choosing just a handful of the top one that have performed well historically.

As for your point on CA loans I would be open to a filter based on those loans because the vast majority of them have performed well. If we can isolate those loans then we would have a successful strategy.

Anyway, we might need to just agree to disagree on this one. I take your points but I will continue on with my strategy.

Reply

TJ May 31, 2012 at 12:25 pm

Peter,

When you manage multiple accounts, how do you avoid duplicate notes? Or do you not worry about that, given the vast diversification you already have?

Reply

Peter Renton June 3, 2012 at 11:26 pm

@TJ, I avoid duplicate notes by having three mutually exclusive filters setup for each account. This way, I can continue to invest with the maximum diversification possible.

Reply

Chris F February 14, 2012 at 12:36 pm

Peter,

I’m curious if you and others are doing any manual filtering based off answers and written descriptions. I use similar filters as above and then base my final decision off seeing some written answers. If someone wants a loan for $35k and they can’t be bothered to write anything anything beyond, “creidt consold” then I skip it. However, this can really limit the number of loans available for me to invest in since a fair amount of borrowers don’t write anything. I am curious if I’m biased against these loans for no good reason and what others take on it is.

I remember reading articles a while back that showed a positive correlation between word count and ROI, but that was a while back and I really haven’t seen too much discussion about it since.

Reply

Larry V February 14, 2012 at 1:21 pm

I’m the same way. I have a variety of filters, but I read every loan description I’m considering. I definitely run away if they are asked specifics and answer with generalities. I’ve also noticed that many of my defaults literally contained some “desperate” language that I used to ignore. On a side note, in case LC is watching, it would be great if I could look at a loan, then hide it or mark it somehow so i don’t keep reading the description.

Reply

Chris February 14, 2012 at 3:00 pm

@ Chris F:
Above you write:

“I remember reading articles a while back that showed a positive correlation between word count and ROI, but that was a while back and I really haven’t seen too much discussion about it since.”

Would you happen to remember where you read this material/article? I would love to look at that data as I have often wondered the same thing with regard to possible correlation between description and default rates.

Thanks for sharing,
Chris S

Reply

Peter Renton February 14, 2012 at 5:54 pm

@Chris F/Chris, Here are two articles from bloggers last year that looked at description length:
Lending Tuber (they did a whole series of posts on loan description) – http://lendingtuber.blogspot.com.au/2011/09/loan-description-length-lending-club.html
Lending Club Modelling (they never did another post after this one) – http://lendingclubmodeling.wordpress.com/2011/04/25/why-loan-descriptions-and-qa-matter/

@Larry, That is a great idea. I would love to see the ability to mark loans as either ignore or invested so you never see them again.

Reply

Roy S February 14, 2012 at 6:06 pm

@Chris, I had to go back through many, many posts to find this…http://lendingclubmodeling.wordpress.com/2011/04/25/why-loan-descriptions-and-qa-matter/

There is also another post somewhere that also has a list with the 10 worst and 10 best words in the description section. I think the ten worst mainly included familial relationships, like child and children. I think another of the worst was “help.” That article might take me a little more time to find.

Reply

Roy S February 14, 2012 at 6:07 pm

Oh look! Peter beat me to posting that article! Oh well…

Reply

Frankie C February 14, 2012 at 6:09 pm

On the topic of descriptions, I wonder if there is a correlation between the quality of the spelling (@Chris’ “creidt consold”) and the default rate. It would be interesting to massively feed descriptions into a spell checker, come up with a metric (errors/100 words?) and then correlate that to default rate over time. Or am I taking this too far? ;-)

Reply

Peter Renton February 14, 2012 at 7:54 pm

@Frankie, Hmmm. I am thinking that would be a very difficult one to implement because people might use prosper nouns or abbreviations that would be picked up as misspellings. It bothers me, too, when people can’t spell but automating that as a filter would be very challenging I would think.

Reply

Bryce M. February 15, 2012 at 9:24 am

Whoops! I posted a couple nice graphs about loan description size but in the wrong thread. They are in the How I Run Multiple accounts thread.

Reply

Peter Renton February 15, 2012 at 5:19 pm

@Bryce. No worries – I have copied your comment from there. Here is what Bryce is talking about:

People were interested in loan description length vs. charge-off. I figured since loan descriptions aren’t going to come back, I’m happy to share a couple graphics. Although it looks like there’s a bowed relationship, it’s statistically borderline (~p=0.09) because the vast majority of the data is between 3 and 7, where it’s essentially the flatish part of the charge-off curve. There’s not enough data to suggest the tips of the curve are really that bowed (as noted by the 95% confidence interval).

http://imageshack.us/photo/my-images/341/histb.png/
http://imageshack.us/photo/my-images/17/qfit.png/

Peter, feel free to use the charts as you like as a contribution to the blog.

Reply

Bryce M February 15, 2012 at 8:44 pm

Just a followup, people should look at both graphics. The first is a histogram of the distribution of the lengths. The second is the money graphic, showing the relationship to charge-off. I note this because the second graphic only had 1 view where the first had 10. People may have thought that the link was just one continuous thing.

Reply

Bryce M February 16, 2012 at 1:03 am

Had some fun with words:

Loans with the word “bills” in the description were 6% more likely to charge off. “Bills” was in the top 50 words used.

Loans with “bible,” “God,” or “pray” defaulted at a 35% rate compared to a population 22% charge off rate, but the result was only borderline significant because there were just 25 such loans among the first 3200 in LC’s history.

Reply

Peter Renton February 16, 2012 at 3:44 pm

@Bryce, This reminds me of a study that Lending Tuber did last year. He did an entire analysis based on the word “need” being used in titles or description. You can check out his entire series here: http://lendingtuber.blogspot.com.au/2011/07/need-series-introduction.html

Reply

Chris F February 16, 2012 at 3:54 pm

@Bryce Very interesting, thanks for passing those graphs along. Just curious, on your loan description axis it goes from 0-8 characters, but have you looked at beyond 8 characters?

The lendingclubmodeling link and lendingtuber where the two articles I remember reading. However, I thought lendingtuber’s findings (on Prosper, not Lending Club) that the shorter the description the greater chance of getting paid was surprising and looking for other research to back that up.

http://lendingtuber.blogspot.com/2011/09/loan-description-length-by-credit-grade.html

Reply

Bryce M. February 16, 2012 at 4:00 pm

Chris,

Per the histogram, there are hardly any loans with descriptions beyond 8 characters. The x scale is in natural logged characters. So, 8 is really e^8 ~ 3000 characters.

Bryce

Reply

Dan B February 16, 2012 at 6:56 pm

Whatever did happen to the “Wordman”? :

Reply

Peter Renton February 17, 2012 at 3:18 am

@Bryce, Thanks for the clarification. I haven’t dealt with natural logs since high school so it is good to get clear on that.

@Dan, He seems to have disappeared from view. While I was never a big fan of his analysis I know plenty of people liked it.

Reply

Bryce M February 18, 2012 at 12:30 am

Inspired by the keyword analysis, I did the same using Lending Club loan titles and found 8 keywords that I will be keeping in my model going forward. There is definitely value to be gained there.

Reply

Peter Renton February 19, 2012 at 7:29 pm

@Bryce, I am a huge fan of studying historical loan data but I am still unconvinced about analyzing keywords. Before I draw any conclusions here I would like to see if the trends remain the same with time and I am not sure we have a big enough loan pool yet to see that.

Reply

Bryce M. February 19, 2012 at 7:44 pm

I wouldn’t waste everyone’s time reporting statistically insignificant results. These are the patterns over all of LCs completed loan history. Some words may not have enough appearances to say much, and that’s why I looked at the top 50 commonly used words only (and many of those were boring and not bother to check like articles the and a).

I suppose one way to validate it would be to predict delinquency on the currently in repayment loans population. I have been thinking of making more use of that data to see if my modeling holds on more current data anyway.

Reply

Peter Renton February 19, 2012 at 9:33 pm

@Bryce, That is precisely the kind of thing that I am interested in. If we can predict delinquency on current loans and if these predictions remain true going forward then we will really have a useful model.

Reply

Bryce M. February 19, 2012 at 9:58 pm

The problem is that defaulting is not the same outcome as charging off. I have done all my modeling with chargeoff or paid in full as the outcomes. Obviously many loans that default wind up curing themselves, so things that predict default may or may not be the same or of the same strength as for chargeoff.

I do two things to test my model. First When I built my model I put a portion of the data aside and then tested the model’s performance on it. Once since I’ve completed it, I tested the model on all the loans that completed since the last loan used on the original model. That is, as time passes, more loans complete to study. But the problem is that they are always three years behind the loans to fund today.

The hope is to build a set of leading indicators to keep an eye out and ensure that the predictors are continuing to perform as expected.

Best wishes!

Reply

Peter Renton February 19, 2012 at 10:12 pm

@Bryce, Thanks for the fuller explanation – I can certainly see you have done your homework. But you do bring up another relevant point: we are working with a moving target. Lending Club has changed their credit policy dramatically from 3 years ago and some loans issued back them would not make it on to the platform. I presume you are excluding those loans from your analysis.

We can’t get a perfect model because Lending Club is always tweaking their risk and underwriting models, but we can still get something that is close.

I would like to share more details about your work with the readers here if and when you are willing.

Reply

Bryce M. February 19, 2012 at 10:35 pm

I definitely knew that LC tweaked their eligibility criteria, and when I built the model I used a technique with an indicator random variable on that set of loans to see if, after all of the other variable I was using, there was still any basic difference for those loans. After all of the work that I did, I no longer had to treat them any differently. It was clear that LC selected better borrowers on criteria that they were disclosing to lenders. So, in the end, we can use those loans to model with also.

I’d be happy to have a conversation with you any time, Peter. You have my email.

Reply

Dan B February 19, 2012 at 11:21 pm

Bryce…………I’m sorry, but defaults & charge offs are in fact virtually the same thing. Almost no loans that default “cure themselves:” as you put it. I have no idea how you could come to that conclusion.

Reply

Bryce M February 19, 2012 at 11:31 pm

Perhaps it is a definition difference, but many people do miss a payment deadline and subsequently fix their situation. The outcomes for loans in repayment are essentially four: (1) paid in full, (2) late of varying degrees, (3) charged off, and (4) current. I am equating “late” to default.

I do not like the word “default” because I have not seen LC define it precisely. It is a meaningless marketing term to me that I cannot associate to any of their public data. I’m sorry if I’m just ignorant of a commonly accepted definition.

Reply

Bryce M February 19, 2012 at 11:34 pm

The main point is that it is hard to analyze loans in repayment in a way that is meaningful and to get a set of leading indicators because we simply don’t know how the outcome of “late” is going to shake out. LC doesn’t give us the whole payment history so we can model that. That’s why I just simply skipped the middle steps and modeled the end state of charge off vs. full payment. That’s what matters in the end anyway.

Reply

Dan B February 19, 2012 at 11:44 pm

Ok I see. So when YOU say “default”, it could actually be someone who is late by a few days. If that is the way you’re defining it then we have no argument here.

Reply

Bryce M February 19, 2012 at 11:48 pm

As further proof that “default” and “charge-off” cannot be the same, LC consistently states its default rate is something like 3%. The facts are that the charge-off rate for all loans that have completed is around 22%. These are so different as to be comical to assert they mean the same thing.

I tried to reproduce this number many ways, but about all I could come up with was that it is the instantaneous proportion of late loans. That is, at any given time, there are about 3% of loans that are delinquent. These risks accrue over time, some cure, some charge off, and we wind up with about a 22% charge-off rate.

Reply

Chris S February 20, 2012 at 12:00 am

Bryce, you bring up an interesting point. I had always assumed that:

Default = Payer has missed too many payments; so now LC/Prosper is going to send off to a third party collections agency with hopes of getting some portion of the balance recouped.

Written-Off = Hopeless possibility of getting any money due to various reasons (borrower has died, filed bankruptcy, etc)

This leads me to wonder – does anyone reading this thread have a more formal/official explanation from LC/Prosper as to how they define these two classifications?

Reply

Bryce M February 20, 2012 at 12:06 am

As a mathematician, it is extremely frustrating (but not at all surprising) to see people throw around words without precise definitions. One of the first things I did with the public data file was to try and reproduce LC’s numbers, such as the 3% “default” rate. I assumed that they meant only 3% of their loans went bad. Imagine my shock when I looked at the loans that had run their 3-year course and found almost 23% had charged off.

A 3% instantaneous delinquency rate sounds much more appealing than 1 in 4 of our loans charges off. I can see why they went that route. But of course all investors do plenty of due diligence to know this, right? =)

Reply

Leave a Comment

Notify me of followup comments via e-mail. You can also subscribe without commenting.

Previous post:

Next post: