Crystall Ball

Home Page
Analysis Section
Prediction Market Implementation
Meet the team
Learn more

Inspiration

We were inspired by the recent investments made into prediction markets and the ever-increasing prevalence of fake news, which is only multiplied by the use of AI to create fake media and quickly churn out articles.

What it does

Uses our neural network model to output a fake probability score for news headlines or snippets of the article. Fetches data from prediction markets and calculates a fair probability of the event happening based on market activity.

How we built it

Machine learning was done through PyTorch and a publicly available Kaggle dataset. We made use of the polymarket API for the prediction market section and searched for relevant events, taking away the spread. Front-end was done on Streamlit for easy hosting during the 24-hour time frame.

Challenges we ran into

The polymarket API was only pulling outdated markets, and it took us a while to set the right parameters and make the decision that we needed to search through both markets and events. Another issue had to do with modeling - when training with subject, text, and title, we were able to reach a higher accuracy within our train and test, but were getting a lot of false positives when just titles were inputted on our website. We went back and trained a second model that only uses titles, and while less accurate, it was giving us better responses on the real titles we tried from reputable news sources.

Accomplishments that we're proud of

Being able to implement our most important ideas, and it all functioning properly together within the tight time frame. I'm proud of both the model detection and getting polymarket integration, as I am a firm believer in the wisdom of the crowd and sharp money.

What we learned

We learned a lot about modeling, working together on a single database, and integrating publicly available information from prediction markets.

What's next for Crystall Ball

I wanted to add insights into why the model leaned towards a certain fake probability, but the model was pretty complex, and simply highlighting suspect words didn't encapsulate the decision very well. Additionally, I want to try to improve the model with different parameter weights and outlier detection. The prediction markets also have a ton of potential, had the idea of discussing liquidity, market patterns, and analyzing for potential insider traders when markets see huge spikes towards a certain outcome. There's an ongoing debate about whether insiders should be allowed to trade on these markets, and we believe that this information should be harnessed, not cast away.