Machine Learning Calibration Edge (Random Forest)

This systematic trading pipeline builds a Random Forest regression model trained on historical Polymarket outcome data, alongside features like time to expiration, volume, open interest, and external news sentiment scores. The model predicts a 'true probability' for each market, which is compared to the current market price. When the model's prediction deviates from the market price by more than 5%, the bot takes a position—buying if the model says the probability is undervalued, selling if overvalued. Position sizing is based on the size of the deviation and a volatility-adjusted risk limit. Backtesting over a 6-month period (Jan-Jun 2026) showed a Sharpe ratio of 1.4 and average trade duration of 2 hours. The system runs automated paper trades and logs all signals to a database for continuous model retraining. Implementation is in Python with scikit-learn, and the pipeline is modular for adding new features.