ML Ensemble Short-Term Crypto Direction (BTC 5-Minute)

How It Works The bot trains three complementary models — LightGBM and XGBoost (gradient boosted trees capturing non-linear feature interactions) and LSTM (capturing sequential time-series patterns) — on BTC price and possibly order book/volume features. The ensemble aggregates predictions, and when consensus probability exceeds a threshold, the bot enters the corresponding YES or NO position on Polymarket's BTC directional market.

Entry / Exit Conditions Entry: ensemble model confidence exceeds a threshold in either the UP or DOWN direction for the next 5-minute candle. Exit: the 5-minute market resolves naturally, or an early exit is triggered if the ensemble flips direction with high confidence before resolution.

Verified Results The repository uses GitHub Actions for automated retraining, suggesting a live or near-live operational pipeline. A GitHub Pages dashboard implies ongoing performance tracking. No specific Sharpe ratio, win rate, or P&L figures are provided in the description.

Implementation Notes Automated retraining via CI/CD is a meaningful operational feature that keeps the model current with regime changes. The very large repo size (567 MB) likely reflects stored model artifacts and historical training data. LSTM requires GPU or extended CPU training time, which GitHub Actions free tier may not handle efficiently.

Risks Short-term crypto prediction is a notoriously difficult problem with near-efficient markets at the 5-minute horizon. Ensemble model outputs may be overfit to historical BTC regimes. Polymarket's BTC 5-minute market may have limited liquidity and wide spreads that consume the edge. Model retraining lag can leave the bot using stale models during regime changes.