Developing a loan filter strategy?

[J2E]

How do you find the filters that work best for you?  ie, reduce default risk while maximizing interest earned?  When I look at the historical data, there are so many possible combinations of factors (income, FICO, DTI, delinquencies, that could influence a loan's 'standing'.  What factors have you come up with that you think have lowered your default rate?  Right now I'm following closely with Peter's High and Medium income filters he posted, thanks!  I really don't like just shotgunning trying out various filter combinations and see what keeps the best rate.

  Doing a brute force on each possible combination is possible, taking hundreds of computer hours I'm sure, but possible, but I'm not sure how beneficial.  The other issue is that you'd have so few matching results, you'd need to compile those results into ranges to get some reasonable pool of loans to have any meaningful statistics.  A sample of 1-2 loans isn't enough.  I'd love to develop a loan strategy for each credit grade, so to have some conservative and some more high risk across the board.  The factors that influence an A rated loan may not be the same as an F or G.

Also, I think something like this has already been done, and no real reason to re-invent the wheel.  See http://blog.dmpatierno.com/post/3161338411/lending-club-genetic-algorithm where he applies an genetic algorithm to get his results (Way above my skill level).  Amazingly he's shared his code so I'm going to look things over and see what I can come up with.

[AnilG]

Backtesting using the historical loan data is the way to go. Statistics, data mining, and pattern recognition can help in identifying important and combinations of factors.

In my opinion, there are two type of analysis. One where someone throws different algorithms at set of data and sees what comes out. The blogger who used genetic algorithm on LC historical data will be in this camp. Second where someone explores the data first to understand what different factors are,  how they are related and impact the results. I am in the second camp as I prefer to understand the factors and explore the relationships before developing quantitative models. If you are interested in learning Statistics more, I will recommend to start with textbook "Practical Business Statistics" by my stats professor Andrew Siegel.

Some of the important factors for lowering defaults based on my analysis are Loan Grade, Interest Rate, Borrower's Location, Loan Purpose (impact declining with time as I believe borrowers are manipulating it), Revolving Credit Utilization, Monthly payment to monthly income ratio, and Months since last delinquency and public record.

You may also want to check out my blog Random Thoughts at http://andirog.blogspot.com  where I have been exploring different factors and its impact on defaults. As a conservative contrarian investor, I am biased toward reducing risk first and then maximizing return second. Based on prior research in consumer lending, recently I introduced BLE Risk Index (BLE = Bad Loan Experience) on my blog and at PeerCube. Also, if you are interested in reviewing loans available through filters created by others, you may want to check out Peer Filters on PeerCube at http://www.peercube.com/lc/peerfilter. I continue to add new filters as I find them online and also encourage users to share their filters. Also, if you share your filter and send me a note to request analysis, I will be happy to analyze and publish findings on my blog and at PeerCube.

[SeanMcD]

Out of curiosity, Anil, what statistical modeling software are you using?

[J2E]

Thanks AnilG, I'm actually a PeerCube member and have been looking around more now that SmartPeerLending has pretty much ceased.  I like that you've made the filters available for viewing, not hidden and only showing the loans that match, and that it's open to member content.

[Peter]

Statistical modeling is how most of the sophisticated investors do it. Now, everyone will approach it slightly differently but I like Anil's approach the best. That is to understand the data before and what data points are important and what can be ignored.

For me, I have been having success doing some basic analysis on Lendstats and Nickelsteamroller for Lending Club data and the new Prosper Stats site for Prosper data: http://prosper-stats.appspot.com/SearchForm.jsp

I see this area being a big growth area for p2p investors in the next 12 months.

[AnilG]

@Sean,

I use quite a few tools. Two primary tools are Excel and Tableau. Though pricey (annual license $1K), Tableau is very good for data exploration. Excel with its Analysis tool pack and Solver add-in and a VBA add-on from my stats prof has been sufficient for most analysis. Time to time I also try to use R and Weka.

@J2E,

Thanks for being a PeerCube user and for your comments. I am continuing to evolve the platform. So, anytime you come across "I wish PeerCube did this ..." moment, let me know. I will try my best to make those thing happen.

@Peter,

I agree with your assessment. Most lenders don't need advanced technique, some basic analysis and data exploration is enough to build up intuition in p2p lending.

[rawraw]

I spent an unimaginable amount of hours in Nickelsteamroller (much more pleasing to the eye than Lendstats).  I then went to Lendstats and other sites to make sure they agreed with my screen.