Inspiration
Most people lose money on prediction markets like Kalshi—not because outcomes are unpredictable, but because of how they trade.
The common strategy is to bet on high-probability outcomes for small gains. It works temporarily, but one loss wipes out multiple wins.
We experienced this ourselves and realized the real opportunity isn’t predicting outcomes—it’s identifying when the market temporarily misprices probability during a live event.
What it does
ProbEdge AI is a real-time trading engine that automatically identifies and trades mispricing in live probability markets.
Instead of predicting who wins, it detects two core patterns—Cross and Non-Cross—and trades probability movement around those patterns.
The system connects directly to Kalshi, monitors live games, enters trades at extreme probability points, and exits on rebounds using AI-optimized thresholds.
This turns prediction markets from guessing outcomes into a structured, data-driven trading problem.
How we built it
We trained a model on over 100,000 historical game states from sports like college basketball and ATP to learn when probability movements become profitable.
We engineered features such as probability drops, time remaining, and relative strength, and used them to classify Cross and Non-Cross patterns and optimize entry and exit conditions.
We then built a real-time trading engine that integrates with the Kalshi API, allowing the system to monitor live data, make decisions, and execute trades automatically.
Challenges we ran into
Avoiding overfitting was a major challenge—many patterns appear profitable in historical data but don’t generalize.
We also had to handle real-time execution, where speed, reliability, and correct timing are critical in fast-moving markets.
Another challenge was dealing with noisy probability data and defining clear thresholds for identifying valid trading setups.
Accomplishments that we're proud of
We built a complete end-to-end system—from idea to live execution.
Developed a new framework for trading probability instead of predicting outcomes Trained a model on 100,000+ data points Built a real-time decision engine Integrated with a live trading API Successfully executed automated trades in real market conditions Most importantly, we demonstrated that this approach can work beyond theory.
What we learned
We learned that prediction markets are less about predicting outcomes and more about understanding how probabilities behave over time.
We also learned how sensitive trading systems are to small assumptions, and how important it is to prioritize robustness over optimizing past performance.
Finally, we gained experience building real-time systems where execution, latency, and stability matter just as much as the model.
What's next for ProbEdge AI
Next, we’re focused on scaling and validating the system.
In the short term:
expand to more sports and markets increase training data and improve model robustness improve risk management and capital allocation run larger-scale live testing In the long term:
We believe this approach extends far beyond sports.
Any market where probabilities update in real time—such as elections, financial events, or macro indicators—can be modeled and traded using the same framework.
Our goal is to build a general-purpose engine for trading probability itself.
Built With
- asyncio
- css
- fastapi
- html
- javascript
- joblib
- kalshiv2api
- node.js
- numpy
- pandas
- pydantic
- python
- react
- scikit-learn
- svg
- uvicorn
- webcryptoapi
- xgboost
Log in or sign up for Devpost to join the conversation.