Author Topic: New note selection model & website. Would love feedback  (Read 38643 times)

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
New note selection model & website. Would love feedback
« on: October 24, 2014, 01:56:24 PM »
Over the last 5-6 months I've been working with my brother on a note selection tool to use for our own investing on Lending Club.  We decided to make the site public so others can use it too and we launched it yesterday.

PeerToPeerQuant.com

I would love to hear feedback.

Its free to make an account, free to see notes we recommend, and first 25 clicks on the notes are free ($0.15 after that).

We use a genetic algorithm to come up with our selection criteria and so far it has been working really well (15.98% ANAR and 11.4% XIRR).  We only have 3 months of data but we outperform most other accounts at that age and risk level.  Obviously we will need more time to know how it does in the long run.

Info on Methodology and Genetic Algorithm
Screenshots of results

Fred93

  • Hero Member
  • *****
  • Posts: 2235
    • View Profile
Re: New note selection model & website. Would love feedback
« Reply #1 on: October 24, 2014, 05:18:32 PM »
Happy to see new players appear.  Wish you luck.

Average age of portfolio is very short for performance measurement purposes. 

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #2 on: October 24, 2014, 05:42:11 PM »
Happy to see new players appear.  Wish you luck.

Average age of portfolio is very short for performance measurement purposes.

Thanks.

Yeah, I agree we need more time to know how the performance is.  On the LC benchmark chart we are in a good place, but there is much more variability at 3 months.

lascott

  • Hero Member
  • *****
  • Posts: 1435
    • View Profile
    • Appreciate my post and want to try LendingRobot? URL below
Re: New note selection model & website. Would love feedback
« Reply #3 on: October 24, 2014, 07:17:03 PM »
I would love to hear feedback.

Its free to make an account, free to see notes we recommend, and first 25 clicks on the notes are free ($0.15 after that).

We use a genetic algorithm to come up with our selection criteria and so far it has been working really well (15.98% ANAR and 11.4% XIRR).  We only have 3 months of data but we outperform most other accounts at that age and risk level.  Obviously we will need more time to know how it does in the long run.

The concept Genetic Algorithms is pretty interesting.

a) I may be willing to try it but I would like to filter by 36 month notes.  Being able to sort the columns would solve that as well. Filtering would be nice and then being able to sort on interest rate.  What is the current sort order?!?!



b) I'm not sure I like tying it to my main google email tho.  I would probably create a gmail account just for this usage.

c) I'm not sure I like the payment for note that don't get funded or completed (ie. Order complete: 1 Note(s) not issued email).  I realize it is hard to track (lendingrobot does) and it is "only" $0.15.  All in all it seems like the P2P-Picks concept was if memory serves me (but they never ended up billing).
« Last Edit: October 27, 2014, 02:48:41 PM by lascott »
Tools I use: (main) BlueVestment: https://www.bluevestment.com/app/pricing + https://www.interestradar.com/ , (others) Lending Robot referral link: https://www.lendingrobot.com/ref/scott473/  & Peercube referral code: DFVA9Y

rawraw

  • Hero Member
  • *****
  • Posts: 2793
    • View Profile
Re: New note selection model & website. Would love feedback
« Reply #4 on: October 25, 2014, 06:34:05 AM »
Are you actually charging retail investors?  If so, there are some legal hoops you have to jump through otherwise it is illegal.  Be sure to take care of that, if it isn't taken care of already.

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #5 on: October 25, 2014, 08:29:04 AM »
Are you actually charging retail investors?  If so, there are some legal hoops you have to jump through otherwise it is illegal.  Be sure to take care of that, if it isn't taken care of already.

Definitely wise advice.  We have spoken with legal counsel and looked into all the laws about the service.  We feel we are legal because we meet several exceptions for not needing to register as financial advisors.

http://www.sec.gov/divisions/investment/iaregulation/memoia.htm

The basics:

* we fit Supreme court rules for classification as a newsletter or publication
* we do not claim to be investment advisors
* we are not big enough to register with SEC and we are based in WY which has no financial advisor laws

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #6 on: October 25, 2014, 08:35:04 AM »
a) I may be willing to try it but I would like to filter by 36 month notes.  Being able to sort the columns would solve that as well. Filtering would be nice and then being able to sort on interest rate.  What is the current sort order?!?!

b) I'm not sure I like tying it to my main google email tho.  I would probably create a gmail account just for this usage.

c) I'm not sure I like the payment for note that don't get funded or completed (ie. Order complete: 1 Note(s) not issued email).  I realize it is hard to track (lendingrobot does) and it is "only" $0.15.  All in all it seems like the P2P-Picks concept was if memory serves me (but they never ended up billing).

Thanks for the feedback lascott. 

a) Most people have suggested this feature too.  We had initially left this out because it would provide lower expected returns than if choose random notes.  But most users seem to want it, so I think we will look into adding that.

b) What would you prefer for login?

c) I definitely understand not wanting to be billed for notes that don't complete.  We decided to set the fee lower to compensate, roughly 0.6% fee if you buy $25 notes.  Much lower if you buyer larger notes.  We could probably check LC site to see if each loan funded or not and refund fees for those that were removed for some reason.  I will look into that.

lascott

  • Hero Member
  • *****
  • Posts: 1435
    • View Profile
    • Appreciate my post and want to try LendingRobot? URL below
Re: New note selection model & website. Would love feedback
« Reply #7 on: October 25, 2014, 12:33:50 PM »
a) I may be willing to try it but I would like to filter by 36 month notes.  Being able to sort the columns would solve that as well. Filtering would be nice and then being able to sort on interest rate.  What is the current sort order?!?!

b) I'm not sure I like tying it to my main google email tho.  I would probably create a gmail account just for this usage.

c) I'm not sure I like the payment for note that don't get funded or completed (ie. Order complete: 1 Note(s) not issued email).  I realize it is hard to track (lendingrobot does) and it is "only" $0.15.  All in all it seems like the P2P-Picks concept was if memory serves me (but they never ended up billing).

Thanks for the feedback lascott. 

a) Most people have suggested this feature too.  We had initially left this out because it would provide lower expected returns than if choose random notes.  But most users seem to want it, so I think we will look into adding that.

b) What would you prefer for login?

c) I definitely understand not wanting to be billed for notes that don't complete.  We decided to set the fee lower to compensate, roughly 0.6% fee if you buy $25 notes.  Much lower if you buyer larger notes.  We could probably check LC site to see if each loan funded or not and refund fees for those that were removed for some reason.  I will look into that.
a) Re: "this feature" -- So did you mean the filter or the sort?  Ideally I'd like both.  It was unclear (unless I overlooked something) how the investing process was meant to work?  Was you expectation just that people would just manually pick starting at the top until they ran out of money.  I had also asked the question: What is the current sort order?!?!

b) I would prefer just any email address.  Not sure what advantage it is for you guys unless it simply means you don't have to build in a confirmation email being sent to that email with a verification url link which is by far the most common thing I see.  I could live with what you are doing but I would not use my main google email (account). I just try to do reasonable things to limit my exposure.

c) I think you users would appreciate that refund even tho we are talking about $0.15.  30%+ notes don't go through. (Morning emails: Order Complete: 1 Note(s) not issued).  I'm not sure what the percentage is based on the statistical means of picking notes (ie. P2P-Picks % that don't go through, or my particular LendingRobot Criteria % that don't go through).  I have a gut feeling that really "good looking" notes may be from folks that are "smart" and may decide to get funds elsewhere (... or payoff the loan within a few months.)
Tools I use: (main) BlueVestment: https://www.bluevestment.com/app/pricing + https://www.interestradar.com/ , (others) Lending Robot referral link: https://www.lendingrobot.com/ref/scott473/  & Peercube referral code: DFVA9Y

lascott

  • Hero Member
  • *****
  • Posts: 1435
    • View Profile
    • Appreciate my post and want to try LendingRobot? URL below
Re: New note selection model & website. Would love feedback
« Reply #8 on: October 25, 2014, 12:44:59 PM »
I realized there is another issue if folks use other methods to pick notes already (IR, LR, LC, BV, BV+P2PPicks, NSR, etc).  What if they want to make sure they don't spend $0.15 to get a duplicate note that they already purchased?

In one theory, you could say that if I picked a note for $25 via BlueVestment+P2PPicksModel and if P2PQ picks in then it must be a really "good" note and may want to double down ($50 as $25 for each systems pick).
« Last Edit: October 25, 2014, 01:21:12 PM by lascott »
Tools I use: (main) BlueVestment: https://www.bluevestment.com/app/pricing + https://www.interestradar.com/ , (others) Lending Robot referral link: https://www.lendingrobot.com/ref/scott473/  & Peercube referral code: DFVA9Y

BruiserB

  • Sr. Member
  • ****
  • Posts: 420
    • View Profile
New note selection model & website. Would love feedback
« Reply #9 on: October 25, 2014, 07:00:28 PM »
At least a few of these allow you to specify you don't want to buy already owned notes. (I'm pretty sure BV allows this stipulation).  But I don't know if it always works in the feeding frenzy of new notes being issued. I suppose more than one service could put the same note in your shopping cart before actually buying it...and at that time both services wouldn't think you already own the note. But like you say, I wouldn't be too upset if I ended up with a few doubles because two of my rules/services picked it.


Sent from my iPhone using Tapatalk

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #10 on: October 27, 2014, 11:18:42 AM »
a) Re: "this feature" -- So did you mean the filter or the sort?  Ideally I'd like both.  It was unclear (unless I overlooked something) how the investing process was meant to work?  Was you expectation just that people would just manually pick starting at the top until they ran out of money.  I had also asked the question: What is the current sort order?!?!

We will probably look into sort and filter as options.  I can't guarantee we can do both or how quickly because I don't know the backend as well as my partner, but I think we will give users the option for both.

Right now the notes are in random order for each user.  Random notes from our recommendations gives the best expected return and minimizes the race to buy the "top" loan if everyone saw the same order.

On our FAQ section on the note page we state the below statement but should make this more clear because most people are asking about it:

"Our data shows that you get the best results by picking random notes on our list. To give you the best return, we shuffle the notes differently for each user. The best strategy is to buy notes working down from the top of the list.

If you pick a smaller subset, like D grade or 36 month notes, you'll have a lower expected return than our algorithm predicts."

b) I would prefer just any email address.  Not sure what advantage it is for you guys unless it simply means you don't have to build in a confirmation email being sent to that email with a verification url link which is by far the most common thing I see.  I could live with what you are doing but I would not use my main google email (account). I just try to do reasonable things to limit my exposure.

Using Google accounts cut dev time a great deal because we are hosting on Google's App Engine (cloud service from google) so using their accounts is a built in feature.  Thanks for letting me know your preference.  We probably won't have this as our top priority right now but will have it on our list to address.

c) I think you users would appreciate that refund even tho we are talking about $0.15.  30%+ notes don't go through. (Morning emails: Order Complete: 1 Note(s) not issued).  I'm not sure what the percentage is based on the statistical means of picking notes (ie. P2P-Picks % that don't go through, or my particular LendingRobot Criteria % that don't go through).  I have a gut feeling that really "good looking" notes may be from folks that are "smart" and may decide to get funds elsewhere (... or payoff the loan within a few months.)

Since I've started picking notes I've kept track of every single one that has been recommended by the algorithm and which funded and which did not.  Our removal rate is 22%.


At least a few of these allow you to specify you don't want to buy already owned notes. (I'm pretty sure BV allows this stipulation).  But I don't know if it always works in the feeding frenzy of new notes being issued. I suppose more than one service could put the same note in your shopping cart before actually buying it...and at that time both services wouldn't think you already own the note. But like you say, I wouldn't be too upset if I ended up with a few doubles because two of my rules/services picked it.

You might double down on a note if multiple services pick it.  Our service doesn't automatically buy the note, it just adds it to your cart so you could double check it wasn't already in your account before making purchase.

For regulatory reasons, we can't change our recommendations per user (or else we would be considered financial advisors) so even if we could see what you had already purchased we wouldn't be able to change our recommendations.

rawraw

  • Hero Member
  • *****
  • Posts: 2793
    • View Profile
Re: New note selection model & website. Would love feedback
« Reply #11 on: October 27, 2014, 12:08:51 PM »
What time period was used for analysis?  Was there out of sample testing?  How often are variables checked for changing predictive power?  Is the LC grade a variable?

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #12 on: October 27, 2014, 12:40:35 PM »
What time period was used for analysis?  Was there out of sample testing?  How often are variables checked for changing predictive power?  Is the LC grade a variable?

All loans in the download file were used in analysis.

There was some out of sample testing done.  We also used some statistical methods to avoid overfitting with the model including limiting/penalizing complexity, and limits on the number of generations the algorithm could use.

We have not checked the variables for changes in predictive power since it is a fairly new model.  We do plan to review the model on a regular basis to adjust if variables are deemed to no longer be predictive.

I can't be specific about what variables are used in the final algorithm, but LC score and FICO were considered as variables as part of the process.  LC and FICO are different models that predict borrower fitness.  We feel it can be very useful to combine different models to come up with a better model.  Similar to how all the top teams improved in the Netflix competition.

rawraw

  • Hero Member
  • *****
  • Posts: 2793
    • View Profile
Re: New note selection model & website. Would love feedback
« Reply #13 on: October 27, 2014, 01:21:09 PM »
So you're using LC grade, although they continually and unannounced change it. And when you say all notes, how can there be out of sample testing?

turing

  • Full Member
  • ***
  • Posts: 222
    • View Profile
    • PeerToPeerQuant.com
    • Email
Re: New note selection model & website. Would love feedback
« Reply #14 on: October 27, 2014, 02:35:23 PM »
So you're using LC grade, although they continually and unannounced change it.

I didn't say we use LC in our final algorithm.  I won't say what variables we use in the final algorithm because it is propriety.  But we did feed LC grades into the genetic algorithm along with most other variables in the data file.

Yes, LC's changes to that variable would complicate its usage compared to other variables.  I agree with you on that.

And when you say all notes, how can there be out of sample testing?

I meant that we downloaded all notes and didn't exclude certain types.  The out of sample part was not in the main run of the algorithm.  I thought that would be implied when I said there was out of sample testing, but I should have been explicit.  Sorry