RESEARCH · GLOBAL
XGBoost, LSTM, and iTransformer show positive gross returns on Bitcoin but fail after transaction costs above 10bps
Walk-forward study (2018–2026, 70K hourly BTC-USDT observations) evaluates three ML models (XGBoost, LSTM, iTransformer). All generate positive trading PnL before costs; once transaction costs exceed 10 basis points, naive strategies fail. A cost-aware filter helps.
WHY IT MATTERS
Crypto trading teams relying on ML should model realistic execution friction. The paper's cost-aware filter (skipping low-conviction trades) offers a practical guardrail against overtrading and whipsaw losses.
Source: arXiv · 2026-06-02