ML Ensemble Short-Term Crypto Direction Trading

How It Works The bot trains an ensemble of gradient boosting models (LGBM and XGB) and a recurrent neural network (LSTM) on BTC price and market data to forecast whether BTC will be up or down over a 5-minute resolution window. Predictions from the three models are combined (likely via averaging or stacking) to generate a final directional signal, which is then used to trade the corresponding Polymarket binary market.

Entry Conditions A trade is entered when the ensemble's predicted probability for one outcome (Up or Down) exceeds the current Polymarket market price by a sufficient margin to provide positive expected value after fees. The auto-training pipeline via GitHub Actions suggests the model is continuously retrained on new data.

Exit / Resolution Given the 5-minute resolution window, positions are held to expiry in most cases. The short duration limits the need for active exit management.

Verified Results No live P&L or backtested Sharpe ratio is disclosed. The continuous retraining via GitHub Actions is a notable implementation detail that addresses model staleness. A live dashboard exists for monitoring performance.

Implementation Notes & Risks Predicting 5-minute BTC direction is an extremely competitive and well-studied problem. The main risks are: (1) the model captures noise rather than signal, especially after transaction costs; (2) 5-minute Polymarket markets may have insufficient liquidity or wide spreads; (3) the Polymarket market price may already reflect the consensus directional view, leaving no exploitable edge; (4) overfitting during training is a significant concern for short-horizon models.