Inspiration

We know that people like the "what if". Making predictions and bets are very common in today's world, and it is not surprising that 4.2 billion people worldwide place bets every year! You can bet on literally everything. But we wanted to find out if we could find a pattern in what we bet on. What brings us excitement, and what predictions, we find quite frankly boring. We want to predict the behaviour of the predictors

What it does

Predicting Popular Predictions is designed to calculate how likely a new prediction market is to be popular. Allowing users to see what other predictions catch peoples attention (and their wallet!)

How we built it

We used Kalshi API to obtain details of events and using parameters such as title, duration, category and frequency of event, we were able to train an LightGM regression model using sk-learn to predict the total volume (money put in from either side of a bet) for a given market. The input from the user was a title. We used google gemini API to obtain both the category and frequency from this.

Challenges we ran into

When collecting data and training our model, we ran into a variety of issues. Firstly, many markets exist that are exceedingly likely/unlikely to happen and as a result had very little volume. For example, the market for 'What will Donald Trump mention at the upcoming meeting with the Japanese ambassador?' had multiple markets, with each word having a 'yes' and 'no' outcome, leading to certain words which would be exceedingly unlikely to be mentioned and thus have little to no trading volume, despite having the same title and description. Moreover, tail events such as "Will Real Madrid win the champions league final?" which attract millions of dollars worth of volume significantly deviate from the median and occur very sparsely, meaning our initial model worked well in the average case (e.g "Will the temperature exceed 57f in New York City Tomorrow") but struggled with the tail events. To overcome this, we implented further features in an attempt to classify these events (Screening for common words such as 'final' and 'election'), although it still tends to undervalue these tail events.

What we learned

During the project, we learned how to use the Kalshi API, learned more about applying a regression model to a variety of different feature types, particularly when encoding the title of the market was a unique challenge we hadn't encountered before. Learned how to use the gemini API to use an LLM from within the program to estimate category and frequency based on the title supplied by the user.

What's next for Predicting Popular Predictions

We plan to significantly improve our model, mainly by trying to classify when an event is an outlier.

Built With

Share this project:

Updates

Submission history