ML Ensemble BTC Short-Term Direction (LGBM + XGB + LSTM)

How It Works This bot targets Polymarket's 5-minute BTC Up/Down binary markets, which resolve based on whether Bitcoin's price is higher or lower at the end of a 5-minute window. It combines three model types: LightGBM (LGBM) and XGBoost (XGB) for tabular feature-based prediction, and an LSTM (Long Short-Term Memory) recurrent neural network for sequence-based pattern recognition. The ensemble vote or blend of these three models produces a probability estimate that is compared against the market price.

Entry/Exit Conditions A position is taken when the ensemble's predicted probability diverges from the market's implied probability by more than a configurable edge threshold. The bot presumably bets YES when the model probability exceeds market price and NO when it falls below. Positions naturally close at the 5-minute market resolution.

Verified Results No live trading results are confirmed — the repo includes a live dashboard hosted on GitHub Pages, which suggests ongoing performance tracking. Auto-retraining via GitHub Actions implies the model is kept current with recent market microstructure.

Implementation Notes The extremely large repo size (~577GB) is suspicious and may indicate bundled data or model artifacts. Auto-training pipelines suggest a MLOps-aware implementation. The dashboard provides real-time monitoring of bot decisions and P&L.

Risks Five-minute crypto markets are highly efficient and dominated by latency-sensitive participants. Ensemble models risk overfitting to historical BTC regimes. GitHub Actions latency may be too slow for competitive 5-minute execution windows. Model retraining frequency and data leakage controls are unverified.