BTC Multi-Timeframe Stacked Ensemble Arb

The strategy uses two leading gradient boosting frameworks—XGBoost and LightGBM—in a stacked configuration to predict price direction for Bitcoin across three different time horizons (5 minutes, 15 minutes, and 1 hour). The model is trained on historical order book data and price action to generate probability estimates for Up or Down outcomes.

The 'edge-based arbitrage' concept exploits discrepancies between the model's predicted probability and the market's implied probability. When the model identifies a significant edge (predicted probability significantly higher than price implies), it takes a position on the corresponding Yes/No contract.

Created in mid-August 2026 and updated through early September, the bot appears to be actively developed. The inclusion of LLM marks in the topics suggests possible incorporation of language models for sentiment or news analysis.

Entry is triggered when the model's confidence exceeds a threshold relative to market price; exit occurs at market resolution or when the edge disappears. The strategy targets short holding periods, reducing overnight risk.

Risks include model overfitting, especially with high-frequency intraday data, and execution lag between model signal and order placement in fast-moving BTC markets. The small repo size (approx. 1 MB) suggests no historical data is stored locally, meaning the model depends on real-time data feeds.