Predicting Lending Club Loan Defaults and Using Them to Maximize Returns | Hacker News Reader

Predicting Lending Club Loan Defaults and Using Them to Maximize Returns

isaac-thedataincubator-project.herokuapp.com

37 pointsTheDataInc8 comments

Minimizing defaults during a period of low defaults is a different exercise than protecting a portfolio against defaults period. Doing the latter is difficult with alternative lenders because the data doesn't go back far enough to be meaningful (most of the loan volume knows only the good part of the economic cycle).

I was in a meeting at a bank recently and the topic of alternative lenders came up. The bank's chief credit officer limited his comments to "I'd like to see how their portfolios and models look after the next economic downturn." Wise words.

How did your banks portfolios and models look after the previous one? It's easy to speak 'wisely' like this when the previous economic down turn was just a few years ago, but if recent history is any predictor the banks will have exactly the same naieve portfolios and models 10 years from now.

Anyway, this page seems to have data from 2007 and on, so I wouldn't really call that a 'low default' period..

It's a common mistake to refer to "the banks" as if they are all the same when they are not. The bank in question is a community bank with a conservative loan portfolio. It did not experience the same losses banks with more aggressive portfolios did.

As for the data you reference, in all of 2007, Lending Club did about $6 million in loans[1]. For comparison, it did over a billion dollars in loans in Q3 2014. Again, alternative online lenders have built the vast majority of their portfolios (or originated most of their loans in the case of services like Lending Club) during an unprecedented period of extraordinarily low interest rates and historically low defaults.

I'm not suggesting that all of these loans are going to go bad when the economy turns; I was simply pointing out that exercises like the one linked to here are quite limited in their real-world utility.

[1] https://www.lendingclub.com/info/statistics.action

Has anyone considered the grameen model of peer pressure lending in P2p marketplaces?

You signup with your facebook account, it records all the details of your facebook friends, and if you go into default on your loan for more than 3 payments then it starts contacting your facebook friends automatically and stating that unfortunately you were unable to keep up with your debts and could your friends chip in some money to help cover things??

the threat of losing face would probably decrease default rates significantly, therefore allowing a decreased interest rate to be charged.

light on details or analysis

There is a github behind it. It has an ipython notebook, but I haven't opened it up to check out if there is enough there for reproducible research. https://github.com/tianhuil/isaac-thedataincubator-project

The data files (data/LoanStats3a.csv) are missing.

Downloadable from Lending Club though: https://www.lendingclub.com/info/download-data.action

yhat had a nice article on this exact subject and data set with a good walk through of the R code.

http://blog.yhathq.com/posts/machine-learning-for-predicting...

Reply on news.ycombinator.com