Inspiration - Our team is very interested in machine-learning, artificial intelligence, and making things for the people using technology. From this idea came the thought that we were going to do a project that involved machine-learning. The next step was to find a problem where this would be very efficient and useful. What better field than one with massive datasets that almost no-one wants to comb through? Finance. Using our neural network, we created a machine-learning loan predictor which will be very useful to banks.
What it does - It uses a custom neural network with machine learning to go through the Fannie Mae datasets, covert it to understandable CSV files, and then take those files, train the neural network to see which loans foreclosed under which parameters and which ones succeeded under other parameters. From this, the program can predict if future loans with similar parameters will be paid or foreclosed.
How we built it - Using a neural network, we were able to create a self-evolving program that took the the information and test results of the base trial , and then produce better equipped “children” through multiple trails with neurons. This led to the program machine-learning over time and allowed us to get closer and closer to the thought-process of a human run bank when they are making decisions about loans, and whether or not the client will pay off the loan. We used Fannie Mae data as our test cases and learned from that data, using fields we felt were most important to come to a percentage decision.
Challenges we ran into - We wanted to use the BigParser API as well, to qualify for another challenge, and to use the power of the Grid. We had actually created 2 Grids, but in the end, we were short on time and manpower (2 of the 4 members got sick and went home), and decided to just to the raw text files. We also had to the change the type of neural network from a more general type with a common structure to a the specific “evolving” genetic type.
Accomplishments that we're proud of - We were able to implement and create a self-learning program that thought like a human, and were able to accomplish this through the use of complex and self-evolving algorithms with datasets that evolved. This constant improvement in almost every facet of the program led to an exponential growth in the accuracy and reliability of the results at the end of the program.
What we learned - We learned how to implement neural networks, work with many files and large datasets, and split the work over one long stretch. These skills are very useful in the real world, as almost all real-world projects use large datasets, multiple files, and the possibilities of neural networks and machine-learning has no ceiling in this technology-driven world.
What's next for Loan Prediction Neural Network - Next, we hope to be able to use the BigParser API to make our program much more efficient. We also hope to create a clean Website that is interactive, where the user can enter in information about a loan, and in real-time, the website will fetch data from the Fannie Mae servers, convert it into CSV files, push to BigParser, and the neural network can calculate the risk rate and the chance the loan will either be paid or foreclosed.