LoanBAT

Inspiration

I researched biases in the financial world and a lot of information came up on loans. Specifically with people arguing whether AI or the regression model was more likely to give you a loan. We decide to use their arguments and come up with a program that models the two for risk analysts to make better judgements and mitigate bias.

What it does

We submit a form for an applicant that runs their data through a legacy linear regression model and a neural net, it displays both those risk scores and uses counterfactual testing to change one variable at a time and discover hidden bias. If both agree on something then the bias is likely small, however larger discrepancies signal a stronger bias.

How we built it

Using React, Next.js, and Flask for the backend. Pytorch, Google Colab and Kaggle to find and process data for the machine learning model.

Challenges we ran into

We had a big problem connecting everything, especiallhy what was in Google Colab and was training our neural net. We ended up unpacking the notebook into the React program and using Flask to run it.

Accomplishments that we're proud of

We are proud that the Machine Learning aspects are thorough and well researched, providing solutions to a real-world problem with modern applications.

What we learned

We learned that bias in the models can depend greatly on the applicant and their data, however with the right inputs, bias can clearly be seen with both models of loan applications.

What's next for LoanBAT

We plan to deploy it using AWS and Docker in the future, this will let risk analysts run the program on-premises or in the cloud.

Built With

• colab
• flask
• javascript
• kaggle
• next.js
• python
• pytorch
• react
• typescript