1 of 19

Improving Underwriting of Branch Users

Mitchell Reynolds

May 3, 2018

2 of 19

Project objective:

Provide Insights that guide the underwriting process

3 of 19

Summary of Insights

Time Series

A key determining factor for repayment is when the loan is disbursed.

​

High Level Features

Of all easily derived variables, it seems that how “popular” a person is predicts if they will repay or default.

Natural Language Processing

Two different approaches

Several words seem to shed light onto defaults and repayments.

4 of 19

Time Series

All loans defaulted if they were disbursed before May 25th, then defaults greatly reduce.

​

Possible Explanation: Repayment time is within a week for Branch and these users get paid around the first of the month.

5 of 19

High Level Features - Logit

6 of 19

Most Important (Pearson R)

After date of disbursement, the highest Pearson R score is the unique count of each variable below for each user:

  1. Thread_id
  2. SMS_Address

SMS Derived Variables:

  1. Total Word Count
  2. Total Character Count

​

​

MULTICOLLINEARITY

7 of 19

Natural Language Processing (NLP) Crash Course

8 of 19

NLP Approach 1 - Combined Messages

9 of 19

NLP Approach 1 - Combined Messages

Word Counts

TFIDF

10 of 19

Combined Messages

Milk, number, want, and tissue seem to move with defaults.

​

No English words made it into the top 10 for repayment

11 of 19

NLP Approach 2 - Each Sentence

12 of 19

NLP Approach 2 - Each Sentence

Word Counts

TFIDF

13 of 19

Each Sentence

This doesn’t make a lot of sense to me…

​

I think the models might be overfitting to certain words.

14 of 19

Model Comparison

High Level Logit

  • 34 Variables
  • F1 = 0.77

NLP Logit

  • Combined: F1 = 0.595
  • Sentence: F1 = 0.672

Stacked Logit

  • Next step would to combine both of the above once I have more data

15 of 19

Thank You!

Questions?

16 of 19

Appendix

17 of 19

18 of 19

19 of 19

Time Series

Most phone calls happen in May then decay thereafter.