ML Calibration and Predictive Edge Trading
Uses a Random Forest model to estimate true probabilities of event outcomes and trades when Polymarket prices deviate from model predictions.
This strategy, implemented in BallesJr/polymarket-edge-model, trains a Random Forest classifier on historical market data (prices, volume, order book imbalances, and external features like news sentiment). The model outputs a calibrated probability for each outcome. When the market price implies a probability significantly lower or higher than the model's estimate, the bot takes a position (buy if undervalued, sell if overvalued). Entry is based on a confidence threshold and a minimum expected value. Exit is dynamic: either price convergence to model prediction, a stop-loss if the model is disproven, or market resolution. Backtesting results indicate a Sharpe ratio above 1.5 on a diverse set of political and financial markets. Implementation includes a pipeline for feature engineering, retraining schedules, and a paper trading bot that executes orders via the Polymarket API.