### Author Topic: Develop Loan-Picking Algorithm  (Read 1230 times)

#### Jogi

• Newbie
• Posts: 18
##### Develop Loan-Picking Algorithm
« on: September 29, 2017, 03:58:00 AM »
Dear Community,

I'm currently developing a loan-picking algorithm for Lending Club loans for my master thesis. Since there is much valuable knowledge in this forum, I would like to incorporate your views on two areas. In return, I would share my thesis and post the final PDF.

Area 1:
My thesis also asses es the investment-attractivenes of Lending Club. Therefore, I'm highly interested on your assessment of the investment- attractivenes.
Some areas to cover would be:
- LC steadily downgrades it's "solid" return which can be earned - indicating a downturn in potential profitability for investors?
- Is it a problem that institutional investors pick "cherry loans" within seconds - leaving only shitty loans?
- Do you see LC as an attractive place to invest and why or why not?
- What is a topic that you would like to get examined?
- Do you think P2P lending can continue to offer lower rates for borrowers (compared to bank loans) and higer rates for investors or will regulation drive will change the game?

Area 2:
My approach so far is to calculate the probability of Default and Prepayment over the whole life-time of a loan (over the whole 36 month). To accomplish that, I applied a logistic regression (lasso, ridge + step foward), a multinomial logit model and a random forest. Afterward I compute the NPV of a loan.
Do you have a hint for the model development?
Especially:
- How would you set up your model development?
- Are there some variables that are your smoking gun predictors?
- Since it was stated that the default rate is different over vintages, a predictive model should not be developed on all data, to predict the future - instead, one should develop the model on data spanning over eg. the last year to predict the defaults over the next (ignoring what happened before)?
- Is it advisable to perform predictions separate for each loan class?

These are just some example questions, that came up to my mind. If you have other concerns or hints - I'm grateful for any suggestions!
« Last Edit: September 29, 2017, 04:23:21 AM by Jogi »

#### PhilGD

• Full Member
• Posts: 150
##### Re: Develop Loan-Picking Algorithm
« Reply #1 on: September 29, 2017, 12:59:34 PM »

- Are there some variables that are your smoking gun predictors?

Finding these variables is the most fun part of the model development process! And this information also tends to be among the most closely guarded and is rarely shared outside of the commonly known variables. Any answers you get to this question will likely be limited to the low-hanging fruit i.e. FICO, DTI, number of inquiries.

#### Jogi

• Newbie
• Posts: 18
##### Re: Develop Loan-Picking Algorithm
« Reply #2 on: October 03, 2017, 08:12:24 AM »
@PhilGD: Thank you very much for the reply! Indeed, this was very interesting! What surprised me is that the updated FICO in the payment history file seems to predict the default quite well. Sadely, I don't have data on Folio - such that I could include a selling option into my model.
Even though, I hoped to get a few more details...

Even if the model development area is technical, Area 1 isn't - what about your assesment of the claimed "solid" returns between 4-6% (which has been 6-8% before) ?

#### Fred93

• Hero Member
• Posts: 1966
##### Re: Develop Loan-Picking Algorithm
« Reply #3 on: October 03, 2017, 07:51:53 PM »
What surprised me is that the updated FICO in the payment history file seems to predict the default quite well.

Yes, but of course an UPDATED variable like this has an easier time predicting default than variables sampled only at the beginning of the loan.

Consider this example: A borrower loses his job, or suffers some other serious hardship.  A few months later, say Jan 1st, he misses a payment at LC.  Another month later, that missed payment is 30 days old, and gets reported to the credit reporting agency.  LC comes along and samples the borrower's FICO score some time during that month, and you see a drop.  At this point the LC payment is 1 months late.  Awhile later it is 2 months late.  Another month and its 3 months late.  Another month and it is now 4 months late, at which point the LC loan defaults.

The important date is the stop pay date, ie the date of the first missed payment.  That's the "event" that we want to predict.  It causes the loan default.  Note that in this example, the drop in FICO came after this event, so the FICO drop cannot predict it.

There are lots of different scenarios besides this one of course.  The critical question is not whether the FICO drop comes before the DEFAULT, but whether it comes before the first payment fails.  After the first payment fails, the price that secondary market buyers are willing to pay for a note suddenly goes WAY DOWN.  Buyers know the probability of recovery is low.

To benefit from observing FICO drops, they have to be FICO drops which occur early in the process.  Early enough so that you can take advantage of them.  You shouldn't just correlate FICO drop to default.  You have to look at timing.

#### rawraw

• Hero Member
• Posts: 2719
##### Re: Develop Loan-Picking Algorithm
« Reply #4 on: October 04, 2017, 03:59:45 AM »
Interesting point Fred

#### bkcarp00

• Newbie
• Posts: 4
##### Re: Develop Loan-Picking Algorithm
« Reply #5 on: October 04, 2017, 12:34:07 PM »
Hi - Some responses to your different Areas. Interested in reading more.

Area 1:
Some areas to cover would be:
- LC steadily downgrades it's "solid" return which can be earned - indicating a downturn in potential profitability for investors?
- Yes I am concerned with the lowering of the expected return each month. I only focus on my own returns, and based on changes lendingclub is implementing I hope that the return rate will stabilize soon and lead to more higher quality loans.
- Is it a problem that institutional investors pick "cherry loans" within seconds - leaving only shitty loans?
- No. I don't concern myself with that institutional investors are investing. I don't consider any loans as "shitty". Every loan has a purpose and place depending on an investors level of risk.
- Do you see LC as an attractive place to invest and why or why not?
- Yes. My return is still far above what I could receive in a savings account. I use lendingclub as a holding place for cash that I am going to at eventually deploy into a different investment. I don't expect super high returns from placing money in lendingclub.
- What is a topic that you would like to get examined?
- What are the biggest factors of defaulted loans that could be used to predict future defaults on current loans. Is there any single set of factors that are unique across charged-off/defaulted loans. How do you predict when a loan is nearing default based on collection/notes from customer service that is placed on the loan.
- Do you think P2P lending can continue to offer lower rates for borrowers (compared to bank loans) and higer rates for investors or will regulation drive will change the game?
- Yes as long as there is capital willing to invest they will be able to maintain lower rates. I'm not concerned about new regulations at least in the short term. The current administration has made it clear they are removing regulations so I don't worry about any new ones being added at this point.

Area 2:

Especially:
- How would you set up your model development?
- Are there some variables that are your smoking gun predictors?
-I look at Credit Score, Past Defaults, debt to income ratio, total income, and loans less than \$25,000 total.

- Since it was stated that the default rate is different over vintages, a predictive model should not be developed on all data, to predict the future - instead, one should develop the model on data spanning over eg. the last year to predict the defaults over the next (ignoring what happened before)?
- Could be a good idea. Maybe look at when lendingclub changes their credit rates/model and use that to identify trends. Go back perhaps 2-3 years. Is data from 10+ years ago when lendingclub was starting still applicable to current day loans?

- Is it advisable to perform predictions separate for each loan class? - Maybe. Every loan class seems to work in different unique ways.

#### Jogi

• Newbie
• Posts: 18
##### Re: Develop Loan-Picking Algorithm
« Reply #6 on: October 12, 2017, 06:51:54 AM »
@ Fred93: Thank you very much for your response!
I'm completely in lign with your argumentation! The value of an updated variable matters only over time - as your example showed. It is therefore valuabe for investors to track the updated fico in order to decide, when to sell a loan on FOLIO.

Sadely, I don't have acces to FOLIO prices, such that I don't use this update, but others can do..

#### Jogi

• Newbie
• Posts: 18
##### Re: Develop Loan-Picking Algorithm
« Reply #7 on: October 12, 2017, 07:08:29 AM »
@ bkcarp00
Many many thanks!

Some areas to cover would be:
- LC steadily downgrades it's "solid" return which can be earned - indicating a downturn in potential profitability for investors?
- Yes I am concerned with the lowering of the expected return each month. I only focus on my own returns, and based on changes lendingclub is implementing I hope that the return rate will stabilize soon and lead to more higher quality loans.
- Is it a problem that institutional investors pick "cherry loans" within seconds - leaving only shitty loans?
- No. I don't concern myself with that institutional investors are investing. I don't consider any loans as "shitty". Every loan has a purpose and place depending on an investors level of risk.
- Do you see LC as an attractive place to invest and why or why not?
- Yes. My return is still far above what I could receive in a savings account. I use lendingclub as a holding place for cash that I am going to at eventually deploy into a different investment. I don't expect super high returns from placing money in lendingclub.

What confuses me was the sharp drop of the loan volume in Q2 2016. Up to Q1 2016 the loan volume has grown nearely expotentially. Then there was this sharp drop:

http://ir.lendingclub.com/Cache/1001216391.PDF?Y=&O=PDF&D=&fid=1001216391&T=&iid=4213397
I thought maybe, that many investors were disappoited by the performance and therefore left.
(I don't think that the resignment of the CEO should have an effect on private investors, snce they invest to earn money not because the CEO is a great guy)

Again, many thanks for the hughe effort - it's so great to have this kind of information source!
Idea sharing brings the world a step further!

#### Fred93

• Hero Member
• Posts: 1966
##### Re: Develop Loan-Picking Algorithm
« Reply #8 on: October 12, 2017, 01:41:34 PM »
What confuses me was the sharp drop of the loan volume in Q2 2016. Up to Q1 2016 the loan volume has grown nearely expotentially. Then there was this sharp drop:
(I don't think that the resignment of the CEO should have an effect on private investors, snce they invest to earn money not because the CEO is a great guy)

It was a result of the scandal.  The banks and other institutional investors pulled back at first whiff of scandal.  Retail investors slowed down, but hung in better than the institutions.

Imagine you're a mid-level person at a bank or other institutional investor.  Suddenly there's a scandal and a lot of negative news about this company you've been investing with.  Big headline stuff.  CEO thrown out.  Words like "accounting irregularities" in all the articles.  This is the stuff you live to avoid.  Chances are, you're about to get a call from somebody two or three levels above your pay grade to explain.  What do you do?  You immediately stop until you can show that you've done a bunch of review, which includes stopping until LC had completed a bunch of auditing and whatnot, and some committees have reviewed the new documentation.  Gotta show that you reacted responsibly.  Continue investing when your ass is well covered with paperwork, and perhaps some other banks have turned back on to give you some cover.  Bureaucrats' main career goal is avoiding embarrassment.  This makes a big hiccup in loan volume for a few months.  Many institutions came back after some checklists of auditing things etc are complete.  Some didn't.  etc.

Renaud was important.  He was the face of the company and an inspirational leader for the whole industry.  Great communicator.  So when he was thrown out, a lot of people began questioning the whole story.  If you begin questioning the whole proposition, then you think differently about how you estimate your return.  Took quite a while for this to settle out.

#### Jogi

• Newbie
• Posts: 18
##### Re: Develop Loan-Picking Algorithm
« Reply #9 on: October 12, 2017, 06:01:29 PM »
@Fred93: I understand your point...Maybe I do view it not so sererely since I'm not personal invested into LC (I would really like but I'm located in Germany).
These herding effects are quite sad since they're very often irrational - like a bank run. The cool thing about a bank run is, that the government secures your savings up to a certain limit.
And here is the drawback of LC not beeing a bank...Only the future knows whethere it pays of to be a non bank...

This rush off was though very sad for LC...and very sad for the remaining investors. Would the volume have staid the same, LC could have generated more revenues and could thus further improve the credit scoring systems....

#### kib

• Newbie
• Posts: 8
##### Re: Develop Loan-Picking Algorithm
« Reply #10 on: October 16, 2017, 11:53:16 AM »
I'd be interested in knowing why different loans are handled differently in terms of recovery attempts.  I'll have one loan where an attempt is made to contact the borrower just about every day, and then another one that looks similar to my eye where apparently no attempt whatsoever is made for weeks.  Possibly, knowing how a loan would be treated if it went delinquent would contribute to modeling (along with information about which recovery method was most reliable.)  E.g. I would be less likely to invest in loans allowed to go three weeks delinquent without any outreach; I'd like to know which ones those are upfront.