Capital Two

VT Hacks8 Hackathon

Hackers: Zaid Al Nouman, Miranda Manka, Antonio Chelala, Michael Trail

Abstract

Capital Two is a fictional banking system that uses custom generated data, the Nessie API, and displays this information on a flask hosted website. This study aims to investigate any user’s bank account and draw conclusions from their account and display that the Capital Two system can be readily integrated with the Nessie API. This was accomplished by generating a set of data containing 10,000 customers each with 45 data fields and outputting it as a .csv file. This file was then taken and all necessary data points were uploaded to the Nessie API and relevant statistical analysis was attempted. Finally, the results were published on a flask hosted website. We conclude that the Capital Two system is a sufficient system for the goals it sought to achieve.

Purpose

The purpose of Capital Two is to create a dataset that would attempt to model “real-world” banking data. We wanted to create a system that could simulate extracting and analyzing data using an API provided by Capital One.

Technical Approach

Part 1: Data Set Generation

We started out thinking about what we wanted our dataset to be about and include. We looked at datasets available on sites like kaggle, but didn’t find what we were looking for. One of the main problems we had was finding the right data. We tried using the data from the nessie API but ran into problems with the data, including many obviously made up fake values. We ended up making our own dataset and created all of the variables we needed to interact with the nessie API. We looked online and researched what we could about different financial data to model our own data. We wanted our data to be as close to real world data as possible. To do this, we researched online to find what the average american might have in their savings, or how much they pay in loans each month. One big problem was that most articles or websites we came across only listed the mean and median. It was difficult to create randomized data without either a range of values or a standard deviation to work with. While we worked with the best information we could find, there were still estimations and educated guesses to create ranges and distributions from the data. However, this creates problems with any analysis we had hoped to do. We used census data to create an income variable to estimate income brackets and most of the rest of the data from. We also used other sources to create other variables. This allowed us to create Capital Two with what we think is a good set of estimates for what random data might look like. While this was good for generating data, it didn’t work well for analysis. Since a lot of the data was estimated from an income variable, most of the data was related to each other, and fairly collinear. While realistically it made sense that someone earning more would be able to deposit more and spend more per month, this didn’t work well as there were some statistical assumptions that were violated and would not make for good analysis. With real data, there would likely be fewer problems with this which would allow for analysis such as time series analysis or regression after further data cleaning.

To generate random values based on prior research data many helper functions were utilized along with 12 main data generating functions. A table below shows each major function and a short description about the functions.

Part 2: Nessie API Integration

We used the Nessie API to ensure that our hack could be applicable to real world situations such as live Capital One databases. First, we inquired about whether or not the Nessie API contained pre-loaded data that could be used for analysis, and set out to create our own dummy data that could be used. However, half way through implementing our algorithm for the dummy dataset, we were informed that the Enterprise endpoint already contained data, so we pivoted to analyze the pre-existing data. We quickly noticed that pre-existing data contained a large amount of arbitrarily large or small values, and set out to generate a dataset that more closely modeled the trends we saw while researching. The dataset was used to fill the Nessie API with a simulated “bank”, and then data was then extracted from the bank using the API and converted into a CSV. By utilizing the Nessie API provided by Capital One, we hope that our solution can be utilized with similar systems for analysis.

One interesting moment we had while working with the API was the amount of requests we would’ve liked to make. Initially, we were thinking about generating a dataset of 10,000 and entering it within the Nessie API. However, we quickly realized that with the number of samples we had, along with the number of factors for each column, the number of requests we would need exceeded 500,000. Despite being told the Nessie API had no limit, we decided to err on the side of caution and not push Nessie to Stressie. Thus, we ended up testing with a much smaller dataset until we were sure our solution would work for any size dataset.

Part 3: Static Website Integration

Initially, we were thinking of using a dynamic website for further statistical analysis of the data, but encountered several roadblocks. Pressed for time and resources, we switched gears and no longer needed a dynamic website to host our data, and opted to keep things simple by hosting our findings with a static website.

Results

Our end product is a collection of separate programs that when in conjunction, can take a large generated data set from prior research and utilize the Nessie API provided by capital to simulate a bank environment.

Conclusions

One thing we could improve would be the creation of the variables. This was our first attempt at something like this so we didn’t know how it would turn out. In the future, we would try to change some of the variable creation methods to allow for more “real-world” type data. Hopefully this would allow for further analysis that we had hoped to do. This is also applicable to similar uses of API for financial data and could be helpful in simulating data that banks have to analyze and potentially predict future outcomes. One thing many members learned was flask for web development. Unfortunately the end result was a static website due it being significantly easier, however, the skillset still remained. We unexpectedly ended up learning significantly more about distribution statistics than originally anticipated.

Reference Sites

https://data.census.gov/cedsci/table?q=United%20States&t=Income%20%28Households,%20Families,%20Individuals%29%3AIncome%20and%20Earnings%3AIncome%20and%20Poverty&g=0100000US&tid=ACSST1Y2019.S1901&hidePreview=false
https://www.valuepenguin.com/banking/average-savings-account-balance
https://www.thebalance.com/breakdown-of-average-monthly-household-expenses-4687519
https://www.fool.com/retirement/2019/10/27/this-is-how-much-the-average-american-saves-each-y.aspx
https://www.daveramsey.com/blog/americans-have-debt#:~:text=A%20recent%20study%20showed%20that,first%2010%20people%20you%20see

Credits To:

Zaid Al Nouman
Miranda Manka
Antonio Chelala
Michael Trail