Around 3 years ago I was looking for a fun supervised learning problem where I could familiarize myself with the popular python machine learning library scikit-learn. I found a dataset on Kaggle that had Lending Club’s historical default rates on 2014 and 2015 vintages and thought that it may be a fun exercise to apply some machine learning methods and see if I can beat the performance of a portfolio I constructed manually.

Now if I were to do this again, I’d definitely approach the feature extraction much differently as I’ve learned a lot in the past 3 years.

Now this comparison isn’t exactly apples to apples in terms of loan characteristics and seasoning across both portfolios. Mostly because I already had my manually constructed portfolio a few months earlier. Again if I were to do this again, I also would have documented the portfolio selection below much more thoroughly :-)

But after a few years:

Not too bad, although definitely not my bestwork. When I get some time, I’d like to revisit and backtest this with some new methods, perhaps a Nueral Network…..

Old Project and Write Up of My Analysis:

Link : 2016 Loan Defaults Project