Random Forest Probability Calibration Model for Systematic Polymarket Trading

How It Works A Random Forest classifier is trained on historical Polymarket market data and relevant features (e.g., time to expiry, volume, historical accuracy of similar markets, external data signals) to produce calibrated probability estimates for market outcomes. When the model's probability estimate differs from the current market price by more than a threshold, a position is taken in the direction of the model's edge.

Entry/Exit Conditions Entry is triggered when the absolute difference between model probability and market price exceeds a minimum edge threshold (accounting for fees and slippage). Position size is determined by Kelly criterion or a fractional variant based on the estimated edge. Exit occurs at contract resolution or earlier if the market price converges to the model estimate and a satisfactory profit can be locked in.

Verified Results The repository includes backtesting infrastructure and automated paper trading, suggesting the model has been validated on historical data before live deployment. However, no specific accuracy metrics, Sharpe ratios, or live trading results are disclosed publicly. The existence of a full pipeline (training → backtesting → paper trading → live) indicates methodological rigor.

Implementation Notes The pipeline is built in Python and includes: feature engineering from Polymarket historical data, Random Forest model training with probability calibration (likely Platt scaling or isotonic regression to convert raw scores to true probabilities), backtesting against historical markets, and an automated paper trading bot for live validation. The small repo size (804KB) suggests a focused, clean implementation rather than a data-heavy system.

Risks Random Forest models can overfit to historical market regimes that no longer hold. Feature selection is critical — using features that are not available at trade time (look-ahead bias) will produce inflated backtest results. Polymarket markets are heterogeneous (political, crypto, sports, etc.), and a single model may not generalize well across categories. Market microstructure changes (new liquidity providers, changing participant mix) can cause model drift.