4. ● Smooth relationship with
default rates
● Small inventory for
riskier loans
gradesarestrong
predictorsoFdefault
5. ● Problematic for two reasons :
○ Low information value for
outliers
○ small sample on outliers
● Binning + Categorical
highlyskeweddistributions
PublicRecords
DELINquencylast 2years
Inquirylast6months
8. ● Random Forest performs
marginally better than
Logistic Regression
● The extra complexity is
not justified by the small
performance improvement.
● Bottomline : Logistic
Regression offers a good
compromise between
accuracy and complexity
LogisticRegression
vs.randomforest
ROCCurve
11. ● Default Rate is not the
relevant metric for an
investor
● Avoiding risk is not a
profitable trading
strategy
● Each trading strategy is
a trade-off between
expected return,
scalability and risk
tolerance
Astatisticaltradingmodel
Expected
Return
default
rate
Liquidity
15. nextsteps:lendingbot
● Automate the execution : LendingBot
● Include recent data, target delinquent loans as well as
“charged-off” .
● Use a scheduler to place orders automatically 4 times a
day.
● Place orders in a segregated portfolio and scale after a
monitoring period of 3 months.
Q&A/talkingpoint
● Is it ethical for a lender to use Zip Code data in its
underwriting policy ?