Big DUT-A

Inspiration

Our group was inspired by standard techniques used to analyze and predict stock trends, like sentiment analysis with LLMs and using LSTMs to forecast data. Our idea was to apply these techniques to financial statements, like the 10Q reports published quarterly by companies.

What it does

Our project attempts to forecast future data and provide sentiment analysis of news regarding financial statements and their metrics, like Earnings per Share, Total Revenue, and Net Income.

How we built it

The frontend was constructed in JavaScript using React. Numerical data in financial statements is analyzed and forecasted in the future using an LSTM with Monte Carlo dropout to introduce uncertainty into the model, and news sentiment is analyzed with the Mistral-7b LLM. Both models use Python, and Flask sends the information to the frontend.

Challenges we ran into

The largest challenge was trying to find proper data. 10Qs tend to vary between companies, so finding standardized data for multiple companies was nearly impossible. 10Qs can be found on SEC's EDGAR platform, but the data is not very usable without heavy pre-processing. In the end we scaled the project back down to just analyzing Goldman Sachs.

Accomplishments that we're proud of

We were able to create a functional frontend, and the Mistral-7b LLM was able to be fine-tuned on news sentiment data on companies. Despite an initial training time of 13 hours, we were able to scale it down to 20 minutes.

What we learned

We learned that one of the hardest parts of data science is finding a viable dataset and cleaning it, especially when the data comes from multiple sources. We also learned how to work together as a team, using version control to collaborate on the same project.