Random Forest Probability Calibration Model
Edge sourced from systematic miscalibration of Polymarket prices, identified via a Random Forest model trained and backtested out-of-sample.
This strategy uses a Random Forest classifier to estimate true event probabilities and compares them against live Polymarket prices. When the model's calibrated probability diverges sufficiently from the market price, a trade is initiated in the direction of the model's estimate. The pipeline includes feature engineering from historical market data, model training with cross-validation, and probability calibration (likely via Platt scaling or isotonic regression) to ensure output probabilities are well-formed.
Entry conditions are triggered when the model's predicted probability exceeds a threshold spread from the current market price, providing a positive expected value cushion after fees. Exit conditions are not explicitly described but likely involve either resolution, a time-based stop, or a price reversion to the model's fair value estimate.
The repository includes a backtesting framework, which suggests some degree of out-of-sample verification was performed. However, with 0 stars and no public performance metrics disclosed, the verified results are unknown. The automated paper trading bot component implies the strategy has moved to a simulated live execution phase.
Implementation uses Python, typical of ML pipelines with scikit-learn. Key risks include model overfitting to historical Polymarket data, concept drift as market participants become more sophisticated, and the relatively thin liquidity in many Polymarket markets making large position execution difficult without moving prices.