By Sahil J, Sarthak K, Brandon G, Andy T

Problem: Loan Inequality

- Despite legislation banning racial discrimination in lending, minorities still struggle to receive loans, even after controlling for other factors.

- Though women make up 30% of loan requests for small businesses, they receive only 4.4% of the total loans.

- Loans are a way to move up the economic ladder, but it’s an unfair game. This exacerbates income inequality

​

​

How can we design an end-to-end system that fosters economic equity in the loan system, matching donors and requests seamlessly?

Presenting...

An unbiased lending system

• Users can either donate or request loans through the website
• Loans are subjected to deep learning classification to determine the probability that the loan will be paid back.
• This deep learning classification runs independent of factors that often discriminate against specific demographics -- and gives users access to zero or low interest loans depending on their profile.
• The money paid back in loans goes back into the system to cyclically fund more people

​

Model

• Fully Connected Neural Network, with a Binary Classifier that predicts probability of payment and charge-off
• Consists of several dense layers followed my dropout layers for regularization.
• Softmax function to transform to [0, 1] space.

75%

Loan status classification accuracy

Features

- Pairing system between low-interest lenders and requests based on interest

- Debiased, deep-learning algorithm to “mask” features of the borrower with no direct correlation to loan return or defaulting

- End-to-end encryption and prediction from borrower to lender

Workflow

AI Backend

Loans donated/requested on website

Flask Server Gateway

Dataset

(1) LendingClub Loan Data (.csv)

2.26 million loans issued from 2007-2015 by LendingClub, with 75 features per loan

​

​

Subjected To

• Removal of features correlated with discrimination
• Removal of loans with substantial missing values

Challenges Faced

• We had to clean the data to deal with NaN values
• Feature engineering and identification relevant to our problem
• We had to adapt our models/architecture based on our results
• We had to do constant hyperparameter (architecture, learning rate, depth, batch sizes etc.) tuning in order for the model to achieve the level of accuracy that we reported.

​

Screenshots

​

Future Work

• Continued hyperparameter experimentation will lead to higher classification accuracy of loan payback.�
• Estimating not just probability of paying back, but timeframe of paying back.�
• Integrating join-filing for loans into the model.�
• Complementing deep learning method with vital domain knowledge to better engineer features.

​