XGBoost on 70,000 Hours of Bitcoin Data Yields 65% Return, but Edge Lies in Execution

A 2026 study (arXiv:2606.00060) applied XGBoost to roughly 70,000 hourly BTC/USDT observations across 27 walk-forward folds spanning 2018–2025, reporting an annualised return of about 65% at 10 basis points in costs with a Sharpe ratio above 1. The research found that naive sign-based trading strategies generate positive gross returns but collapse once transaction costs are applied, because they trade too frequently. The key fix was a cost-aware execution filter that only triggers a position when the forecast magnitude exceeds a set threshold, reducing turnover by more than tenfold and restoring net profitability. Feature importance analysis highlighted the 4-hour Bollinger Band position and RSI as the most discriminative signals, reinforcing the role of intermediate-term volatility and momentum. Crucially, the study notes that the model's edge does not survive strict statistical significance tests against a simple buy-and-hold strategy, underscoring that execution discipline — not model sophistication — drives the results.