Inspiration
Prediction markets, due to their nature, currently experience tons of insider trading activity. This makes many prediction markets fundamentally unfair to laypeople. We wanted to build an application that levels the playing field and identifies these insider traders so that people can capitalize on their activity.
What it does
PolyPredict collects insider-trader market data and trains a custom machine learning model on 20+ indicators of insider trading activity. It then displays ranked trader-level predictions, grouped market consensus views, and model performance stats in a web dashboard.
How we built it
We built a full pipeline to scrape insider trading activity, fetch/parse outcomes, clean and normalize mixed-format CSV features, train a PyTorch binary classifier, and generate predictions for future markets. The frontend is a Django app with pages for per-trader predictions, grouped market consensus scoring, and model metrics/threshold accuracy summaries.
Challenges we ran into
The hardest part was data quality: inconsistent field formats, missing outcomes. We also had to handle external API variability.
Accomplishments that we're proud of
Our machine learning model performs extremely well, predicting 100% of insider trades accurately at a 99% model confidence level, and 98.2% of insider trades accurately at a 95% model confidence level.
Log in or sign up for Devpost to join the conversation.